UNIVERSITY OF CALIFORNIA
Santa Barbara
An Assessment of Hyperspectral and Lidar Remote Sensing for the Monitoring of
Tropical Rain Forest Trees
A Dissertation submitted in partial satisfaction of the
requirements for the degree Doctor of Philosophy
in Geography
by
Matthew Loren Clark
Committee in charge:
Professor Dar A. Roberts, Chair
Professor Oliver A. Chadwick
Professor David B. Clark
Professor Phaedon C. Kyriakidis
September 2005
The dissertation of Matthew Clark is approved.
____________________________________________
Oliver A. Chadwick
____________________________________________
Phaedon C. Kyriakidis
____________________________________________
David B. Clark
____________________________________________
Dar A. Roberts, Committee Chair
September 2005
An Assessment of Hyperspectral and Lidar Remote Sensing for the Monitoring of
Tropical Rain Forest Trees
Copyright © 2005
by
Matthew Loren Clark
iii
ACKNOWLEDGEMENTS
This research would not have been possible without financial support from
NASA Headquarters, through an Earth System Science Fellowship Grant NGT530436, and a Teaching Assistantship from the UCSB Geography Department. I am
also deeply thankful to the AVIRIS team at NASA’s Jet Propulsion Laboratory, who
let me use their portable spectrometer and who paid the costs to ship the instrument
to Costa Rica for my field campaign. In particular, I would like to thank Rob Green
and Jessica Faust for training and logistical support during my graduate studies.
Another critical part of this research was access to unique tropical rain forest
datasets. The Spectral Information Technology Application Center (SITAC)
collected the HYDICE imagery and the U.S. Army Corps of Engineers Topographic
Engineering Center collect the FLI-MAP lidar data used in this research. Both
datasets were graciously donated to the Organization for Tropical Studies (OTS),
who manages the La Selva Biological Station in Costa Rica. OTS provided
logistical support for these campaigns, as well as for my field work in Costa Rica.
Topographic field data in my research was based upon work supported by the
National Science Foundation under Grant DEB-0129038. Old-growth tree data were
collected by David and Deborah Clark, with support from the National Science
Foundation (Grant DEB-0129038) and the Andrew W. Mellon Foundation. Plotscale tree height measurements were provided by Jack Ewel, whose work is
supported by the National Science Foundation (Grant DEB-9975235) and by the
Andrew W. Mellon Foundation.
iv
I would like to thank my main academic mentors in my doctoral research,
Dar Roberts and David Clark. Dar taught me to appreciate the physics behind
remote sensing and he propelled me forward with his keen intellect, enthusiasm and
energetic teaching abilities. David instilled in me a profound appreciation for
tropical scientific research and the wonders of the La Selva forest. His enthusiasm
for bridging the gap between remote sensing and forest ecology inspired me to return
to graduate school and to conduct my research at La Selva. In addition, I would like
to thank my committee members, Phaedon Kyriakidis and Oliver Chadwick, for their
academic guidance in helping me complete my dissertation.
Thank you to my parents, Adrianne and Don Clark, as well as my sister
Phoebe Kleiger. My mother has been a solid pillar of emotional support and a
wealth of wisdom during my entire academic career. I would not have been able to
complete my doctoral program without my loving wife, Ana Horta. I am deeply
grateful to have Ana in my life, and I thank her for her love, companionship and
willingness to leave her family and country and to put her career on hold so that I
could complete my doctoral degree.
I would like to thank past and present OTS staff, especially Robert Matlock,
Jorge Jimenez, Cristián Coronas, Orlando Vargas, and Antonio Trabucco for their
assistance in helping bring my research to fruition. I am greatly indebted Leonel
Campos and William Miranda at La Selva, who provided me with warm
companionship and invaluable expertise while conducting my field work.
I also thank my colleagues in the VIPER lab (and honorary members) for their
friendship, computer code, morale and lively discussions: Phil Dennision, Kerry
v
Halligan, Carl Legleiter, Izaya Numata, Seth Peterson, Becky Powell, Dylan
Prentiss, and Carlos Souza. I also thank the Geography Department staff, especially
Michelle Kueper, Connie Padilla, and Beilei Zhang. Finally, I would like to thank
Deborah Clark, Stephanie Bohlman and Jim Kellner for their friendship and
enthusiastic support of my research.
vi
VITA OF MATTHEW LOREN CLARK
September 2005
EDUCATION
Bachelor of Arts in Integrative Biology and Environmental Science, University of
California, Berkeley, December 1993
Master of Science in Ecosystem Analysis and Conservation, University of
Washington, Seattle, December 1998
Doctor of Philosophy in Geography, University of California, Santa Barbara,
September 2005 (expected)
PROFESSIONAL EMPLOYMENT
2003-2005: Graduate Student Researcher, Department of Geography, University of
California, Santa Barbara
2001-2002: Teaching Assistant, Department of Geography, University of California,
Santa Barbara
1998-2001: GIS Laboratory Manager, La Selva Biological Station, Organization for
Tropical Studies, Costa Rica
1996-1998: Graduate Research Assistant, Long Term Ecological Research Network
(LTER), University of Washington, Seattle
1994-1996: Senior Biochemical Technician, Genentech, Inc., South San Francisco
AWARDS
National Aeronautics and Space Administration Earth Systems Science Fellowship,
2002-2005
California Space Grant Fellowship, 2004
PUBLICATIONS
Clark, M.L., Roberts, D.A., & Clark, D.B. (2005). Hyperspectral discrimination of
tropical rain forest tree species at leaf to crown scales. Remote Sensing of
Environment, 96(3-4), 375-398.
Clark, M.L., Clark, D.B., & Roberts, D.A. (2004). Small-footprint lidar estimation of
sub-canopy elevation and tree height in a tropical rain forest landscape. Remote
Sensing of Environment, 91(1), 68-89.
Clark, D.B., Read, J.M., Clark, M.L., Murillo Cruz, A., Fallas Dotti, M., & Clark,
D.A. (2004). Application of 1-m and 4-m resolution satellite data to studies of
tree demography, stand structure and land-use classification in tropical rain
forest landscapes. Ecological Applications, 14(1), 61–74.
Clark, M.L., Roberts, D.A., Gardner, M. & Weise, D.R. (2004). Estimation of
Hawaiian Islands fire fuel parameters from AVIRIS imagery. Proc. 13th Annual
JPL Airborne Earth Science Workshop, Jet Propulsion Laboratory, Pasadena,
CA.
Powell, R.L., Matzke, N., de Souza, Jr., C, Clark, M.L., Numata, I., Hess, L.L., &
Roberts, D.A. (2004). Sources of error in accuracy assessment of thematic landvii
cover maps in the Brazilian Amazon. Remote Sensing of Environment, 90, 221234.
Clark, M.L. (1998). An Analysis of Western Olympic Peninsula Forest Structure
Using Combined Synthetic Aperture Radar and Landsat Thematic Mapper
Images. Master of Science Thesis, University of Washington, Seattle, WA. 214
pp.
PRESENTATIONS
Clark, M.L., Clark, D.B, & Roberts, D.A. “Remote sensing of tropical rain forest
structure with small-footprint lidar”, Ecological Society of America &
International Congress of Ecology, 90th ESA Annual Meeting, August 7-12,
2005, Montreal, Canada.
Clark, M.L., Clark, D.B, & Roberts, D.A. “Lidar estimation of sub-canopy elevation
and tree height in a tropical rain forest landscape”, Guest lecture presentation,
University of California, Geography Dept., May 27, 2005, Santa Barbara, CA
Clark, M.L., Roberts, D.A., & Clark, D.B. “Hyperspectral discrimination of tropical
rain forest tree species at leaf to crown scales”, American Society of
Photogrammetry and Remote Sensing Annual Conference, March 7-11, 2005,
Baltimore, MD
Clark, M.L., Roberts, D.A., Gardner, M. & Weise, D.R. “Estimation of Hawaiian
Islands fire fuel parameters from AVIRIS imagery”, 13th Annual JPL Airborne
Earth Science Workshop, March 31-April 2, 2004, Pasadena, CA.
Clark, M.L., Clark, D.B., & Roberts, D.A. “Lidar estimation of sub-canopy elevation
and tree heights at La Selva”, Visiting researcher colloquium, April, 14, 2004, La
Selva Biological Station, Costa Rica.
Clark, M.L., Powell, R., Matzke, N., de Souza, Jr., C., Numata, I., Hess, L.L.,
Roberts, D.A. “Accuracy assessment of remote sensing products using airborne
videography: A case study from Rondônia, Brazil”, Invasive Exotic Plants:
Approaches for the Florida Landscape, Conference and Workshop, February 1214, 2003, Miami, FL.
Clark, M.L. Evaluación de sucesión usando sensores remotos comerciales. Guest
lecture presentation, Universidad de Tucumán, April 22, 2003, San Miguel de
Tucumán, Argentina.
Clark, M.L. (1999). “Teledetección y ArcView GIS”. World Bank Central
American Ecosystem mapping project workshop, June, 1999, Panama City,
Panama.
viii
ABSTRACT
An Assessment of Hyperspectral and Lidar Remote Sensing for the Monitoring of
Tropical Rain Forest Trees
by
Matthew L. Clark
The main objective this research was to assess two types of emerging remote
sensing technology, hyperspectral and lidar sensors, for the automated
discrimination of tropical rain forest tree (TRF) species. The hyperspectral data
contain information on the biochemical and structural properties of crowns, while
the lidar data contain structural information. I hypothesized that these two datasets
combined would permit greater species classification accuracy than either dataset
alone.
Working in an old-growth TRF in Costa Rica, canopy-emergent individual
tree crowns (ITCs) for seven target species were manually digitized with reference
to high spatial resolution hyperspectral and lidar datasets that were acquired from
airborne sensors. Multispectral and hyperspectral classification was performed
using pixel- and crown-scale spectra and spectral angle mapper (SAM), maximum
likelihood (ML), and linear discriminant analysis (LDA) classifiers. Pixel-majority
and crown-scale ITC classifications were significantly more accurate with
hyperspectral data relative to multispectral data, revealing the importance of the
ix
spectral detail offered by hyperspectral imagery. Additional techniques were
explored to best harness this spectral information. These included incorporating
hyperspectral metrics into decision trees (DTs) and multiple endmember spectral
mixture analysis (MESMA). The best spectral-based classification accuracy was
with crown-scale spectra and a relatively simple LDA procedure. These results
suggested that hyperspectral imagery need not be acquired at a very high spatial
resolution or analyzed with sophisticated techniques to provide adequate
discrimination of species. Leaf phenology was important in mapping TRF tree
species. Leaf-off trees had distinct volume-scattering and spectral mixing
properties that influenced classifier variable selection as well as final classification
accuracy.
Crown-scale hyperspectral data were combined with structural data from the
lidar sensor in LDA and DT classifiers. There were significant differences in the
majority of lidar-derived structural metrics among the study tree species; however,
the addition of this information to the classifiers did not improve classification
accuracies. Although lidar data was not useful for species discrimination, it did
provide an unprecedented view of canopy topography and sub-canopy elevation that
is difficult to measure using traditional techniques.
x
TABLE OF CONTENTS
CHAPTER 1: Introduction............................................................................................ 1
CHAPTER 2: Lidar estimation of sub-canopy elevation and tree heights ................. 17
CHAPTER 3: Discrimination of tree species at multiple scales................................. 71
CHAPTER 4: Classification of tree species with absorption features and binary
decision trees............................................................................................................. 132
CHAPTER 5: Classification of tree species with multiple endmember spectral
mixture analysis ........................................................................................................ 186
CHAPTER 6: Comparison of lidar and hyperspectral data for tree classification ... 221
CHAPTER 7: Conclusions........................................................................................ 254
REFERENCES.......................................................................................................... 273
APPENDIX I: List of Acronyms .............................................................................. 306
APPENDIX II: Summary of spectral metrics (Chapter 4)........................................ 309
xi
CHAPTER 1: Introduction
1.1. Background
Mapping and monitoring of threats to tropical rain forest biotic diversity
Tropical rain forests (TRF) now cover only 6.4 percent of the Earth’s terrestrial
surface (9.5 x 108 km2) yet they maintain a large proportion of the world’s biotic
diversity (Thomas et al., 2004; Whitmore, 1990). There are many reasons to be
concerned about TRF biotic diversity. Monetary benefits include timber and nontimber (e.g., fruits, nuts, medicines) products (Fearnside, 1999). The vast genetic
diversity contained in TRF could provide future economic wealth and utility, such as
through undiscovered pharmaceuticals or food resources (Fearnside, 1999;
Laurance, 1999). Besides these utilitarian benefits of TRF, many authors have
written about the ethical reasons to protect the irreplaceable species and human
cultures maintained by these forests (Raven, 1988; Laurance, 1999).
Often
neglected in evaluating TRF value are the many ecosystem processes or “services”
that an intact forest provides, such as flood control, soil conservation, carbon storage
and regional and global climate regulation (Laurance, 1999). For example, it is
estimated that tropical forests store 59% and 27% of global carbon stocks in
vegetation biomass and soils, respectively (Dixon et al., 1994).
Deforestation and forest fragmentation are well-documented immediate threats to
tropical forest biodiversity (Achard et al., 2002; Fearnside, 1999; Phillips, 1997;
Skole & Tucker, 1997). Digital images from satellites have been crucial for
understanding the spatial extent and temporal dynamics of these threats because the
1
sensors provide continuous spatial and frequent temporal measurements of reflected
radiance from TRF canopies. Mapping of tropical forests has primarily relied on
medium spatial resolution satellite imagery from multispectral sensors, such as
Landsat Thematic Mapper (TM) with 30-m pixels and 6 optical bands. This imagery
is relatively inexpensive and has permitted the mapping of general forest cover
classes needed to calculate the rate and extent of deforestation and forest
fragmentation (Cochrane et al., 1999; Roberts et al., 2002; Skole & Tucker, 1993;
Steininger et al., 2001). However, variability in forest types due to high tree
diversity and natural and human disturbances results in complex radiance signals
that are difficult to discriminate using coarse spectral and spatial resolution sensors,
leading to significant errors in estimates of land cover area and temporal change
(Achard et al., 2002; Cochrane, 1999; Foody, 2003; Powell et al., 2004; Skole &
Tucker, 1993).
Tropical forest degradation is also a threat to biodiversity. Factors causing
degradation include: 1) the selective extraction of plants, such as the removal of
commercially-valuable trees (i.e., selective logging); 2) selective removal of animals
(i.e., defaunation); 3) invasive species; 4) climate change; 5) changing atmospheric
composition, such as elevated CO2 on plant growth; 6) changing tree mortality and
recruitment rates; 7) fire intrusion into forests; and 8) forest fragmentation
(Cochrane et al., 1999; Nepstad et al. 1996; Phillips, 1997; Uhl & Kauffman, 1990).
Many of these biodiversity threats have direct and indirect effects on tropical trees,
which are the major components of forest structure and largely determine the
2
microclimate, substrate and food resource environments that sustain the rich plant
and animal biota in the forest. One example of an indirect effect on biodiversity is
climate change. Warmer global temperatures and associated changes in precipitation
patterns linked to greenhouse gas emissions may alter tree growth, recruitment and
mortality rates, thereby creating new assemblages of trees (Clark, Piper et al., 2003;
Laurance et al., 2004; Phillips, 1997). If these altered tree communities fail to
sustain the complex interactions among trees, pollinators, seed dispersers,
herbivores, symbiotic fungi and other species that are common in tropical forests,
then overall biodiversity will likely decline (Laurance et al., 2004). One recent
global-scale study concluded that climate-change effects on tropical forests over the
next 50 years may pose as much risk to species survival as deforestation (Thomas et
al., 2004).
Our scientific understanding, monitoring, and management of these threats to
biodiversity are greatly hindered by a lack of spatially- and temporally-extensive
information on tree demography (Clark, Read, et al., 2004), such as species,
location, growth, and survivorship. Forest degradation affects one to several trees,
and it may go undetected in coarse spatial resolution imagery unless it causes clear
spectral change. For example, selective logging may increase fractions of exposed
soil or dead vegetation, permitting detection in coarse resolution pixels (Asner et al.,
2002; Souza & Barreto, 2000). Individual tree crowns (ITCs) are not resolved in
coarse-resolution imagery from a sensor such as Landsat. Most available data on
tree demography come from relatively small, sparsely-sample field plots with
infrequent re-sampling intervals. It is difficult to generalize such field data to the
3
landscape, regional and global scales needed for understanding the important
processes affecting biodiversity (Foody et al., 2003; Tuomisto, Poulsen et al., 2003;
Phillips et al., 2003). However, more intensive sampling over broader spatial scales
is generally not possible due to inaccessibility or cost constraints.
For more detailed maps of ITCs over broad spatial scales, scientists and land
managers in the tropics have relied on visual interpretation of aerial photographs
from film cameras (Clément & Guellec, 1974; Herwitz et al., 1998; Myers &
Benson, 1981; Trichon, 2001). Aerial photography has limited use in the tropics
because these environments are mainly located in developing countries, where
imagery is prohibitively expensive or otherwise difficult to obtain (Clark, Read, et
al., 2004). In addition, because photo-interpretation relies on human intervention, it
is prone to inconsistencies among interpreters (Congalton & Mead, 1983; Myers and
Benson, 1981).
Emerging remote sensing technology for individual tree crown analyses
A new generation of high spatial resolution (< 4 m), multispectral airborne and
spaceborne digital sensors now exists that can resolve ITCs as groups of digital
image pixels (McGraw et al., 1998; Gougeon & Leckie, 2003). In particular, the
launches of the commercial IKONOS and Quickbird satellites in 1999 and 2001
(Space Imaging, Thornton, CO; Digital Globe, Longmont, CO), respectively, have
greatly increased the availability of high spatial resolution imagery for applications
in tropical forests (reviewed in Hurtt et al., 2003). Preliminary TRF studies in Costa
4
Rica and Brazil have shown that IKONOS imagery can resolve ITCs and permit
crown size measurements (Asner et al., 2002; Clark, Read, et al., 2004; Read et al.,
2003) and the tracking of tree mortality (Clark, Castro, et al., 2004). These details of
crown structure, along with other canopy-scale information such as
presence/absence of tree species, logging roads and patios, may greatly improve the
discrimination of selective-logging impacts (Clark, Read, et al., 2004; Read et al.,
2003; Souza & Roberts, 2005). This capability has immense potential for
monitoring forest degradation and insuring regulatory and certification compliance
over broad scales (Read et al., 2003).
In high spatial resolution optical imagery, each ITC encompasses many pixels
with an n-dimensional spectrum that can be used for automated species
discrimination.
Current research shows that species discrimination is best
accomplished by aggregating pixels into their respective crowns for object-based (as
opposed to per-pixel) classification using each crown’s spectral and spatial
properties (Gougeon, 1995; Leckie, Gougeon, Hill et al., 2003; Meyer et al., 1996;
Preston et al., 1999).
Locating and delineating these image objects can be
accomplished with manual digitization (Asner et al., 2002; Clark, Read, et al., 2004;
Lamar et al., 2005; Read et al., 2003). For consistency and reduction of production
costs, ITC inventory over large spatial extents will require automated algorithms for
crown detection and delineation. Automated ITC delineation is a sub-discipline of
image segmentation (Warner et al., 1999).
Recently developed algorithms are
reviewed extensively in Hill and Leckie (1999), McGraw et al. (1998), and Key et
5
al. (2001).
Techniques generally involve collapsing spectral data into one
illumination/albedo band (Warner et al., 1999) from which radiometric maxima and
minima are analyzed to locate tree centers (maxima) and define crown edges
(minima), respectively (Gougeon, 1999; Key et al., 2001; Culvenor, 2002). Most
segmentation algorithms were developed for forestry applications in coniferdominated stands (e.g., Brandtberg & Walter, 1998; Gougeon, 1999; Lamar et al.,
2005). There are still many challenges to crown delineation in complex, old-growth
hardwood forests, where trees may intertwine and overlap, inter-tree shadows are
narrow, and canopy-gap shadows are prevalent (McGraw et al., 1998; Warner et al.,
1999), causing automated segmentation schemes to identify clusters of trees rather
than individual crowns (Culvenor, 2002).
High spatial resolution optical sensors typically measure a few spectral regions
in broad bands (i.e., multispectral). Vegetation spectra are controlled by similar
factors, such as chlorophyll and water absorption in leaf tissues, and as a
consequence their spectral differences are subtle and substantial spectral detail may
be needed to distinguish among species (Price, 1994; Cochrane, 2000). Airborne
and spaceborne hyperspectral optical sensors offer such detailed spectral information
by measuring the visible to shortwave infrared regions of the electromagnetic
spectrum (400-2500 nm) in over 100 channels. The Hyperion sensor on the EO-1
platform is the only spaceborne sensor offering hyperspectral data over the large
spatial extents needed for regional ITC mapping; however, the 30-m resolution of
the imagery has precluded ITC classification research. High spatial resolution
hyperspectral data (typically > 4 m) is now available from commercial (e.g., HyMap
6
[Integrated Spectronics Pty Ltd., Baulkham Hills NSW, Australia]) and experimental
(e.g., Airborne Visible/Infrared Imaging Spectrometer [AVIRIS; Jet Propulsion
Laboratory, NASA, Pasadena, CA USA]) airborne sensors. It is anticipated that this
high spatial and spectral resolution imagery will make the automated classification
of tropical species a reality (Cochrane, 2000); however, this hypothesis has remained
untested with an airborne or spaceborne hyperspectral sensor.
Another promising technology for ITC-level analysis is LIght Detection And
Ranging (lidar). Lidar systems are an active form of remote sensing that uses a
scanning laser to measure surface heights (Baltsavias, 1999a). These sensors have
emerged as the premier instruments for the generation of detailed digital terrain
models (DTMs: Kraus & Pfeifer, 1998; Petzold et al., 1999) and the estimation of
forest height (Næsset, 1997; Magnussen & Boudewyn, 1998; Persson et al., 2002)
and aboveground biomass (Drake, Dubayah, Clark et al., 2002). Small-footprint
lidar sensors record the three-dimensional height distribution of crown components,
providing information on crown structure that is useful for conifer species
discrimination (Brandtberg et al., 2003; Holmgren & Persson, 2004).
There have
been no attempts to discriminate species with lidar in a tropical forest environment.
Automated, computer-based classification of tropical rain forest trees
In visual interpretation of aerial photographs, the species of tree crowns are
distinguished using crown color and structural properties, such as branch
architecture, canopy position, contour shape, size, foliage cover and texture
7
(Fournier et al., 1995; Trichon, 2001). These spectral and structural properties
should be equally important for automated, computer-based ITC classification.
Hyperspectral imagery provides visible color information, but also includes the
near infrared and shortwave infrared regions of the electromagnetic spectrum beyond
the range of human vision. Tree spectral response is largely determined by the
principle tissues in their crown—leaves and bark (Asner, 1998; Roberts et al., 2004).
In general, the optical properties of these tissues are controlled by surface and
internal structure and biochemical concentrations, such as water, chlorophyll, lignin,
and cellulose. At leaf scales, many species spectra may be similar in shape because
there are only a few factors that control leaf reflectance (Price 1994, Poorter et al,
1995; Cochrane 2000). Furthermore, there may be significant spectral variation
between individuals of a single species or within one individual’s crown, which
could limit our ability to discriminate species through spectral techniques (Cochrane,
2000). This is especially true in the tropics, where epiphylls quickly colonize
species with long-lived leaves (Roberts, Nelson et al. 1998) thereby increasing
conspecific (within species) leaf spectral variation. Unique species spectra are more
likely found at branch or crown scales, where crown structure exerts an influence on
overall spectral response in the sensor’s ground instantaneous field of view
(GIFOV). At these scales, important structural properties such as branch
architecture, leaf arrangement and morphology, and species phenology (e.g., leaf
turnover, flowering) combine to determine the relative mixtures and illumination of
photosynthetic vegetation, non-photosynthetic materials, and shadows that form the
radiance signal measured by the sensor (Asner et al., 1998, Roberts et al., 2004).
8
For imagery with a spatial resolution finer than the scale of a crown, the spatial
arrangement of spectral vectors (i.e., pixels) may encode additional information on
crown structure. Few studies have used pixel spatial information for automated ITC
species discrimination, and there is inconclusive evidence that it greatly improves
forest composition classification (Franklin et al., 2000; Leckie, Gougeon,
Walsworth, & Paradine, 2003, Wang et al., 2004; although see Franklin et al., 2001;
Zhang, et al., 2004). A more straight-forward approach to remote measurement of
crown structure is through the use of small-footprint lidar sensors, since their
measurements respond directly to the physical arrangement of crown tissues
(Brandtberg et al., 2003; Holmgren & Persson, 2004).
1.2. Research overview and objectives
The main goal of this research is to investigate high spatial resolution
hyperspectral and lidar sensors for automated tropical rain forest species
classification. The study site and data sets are described below in Section 1.3. This
research involved several analytical decisions and constraints, including: 1) selection
of tree species of interest, 2) crown detection and delineation methods, 3)
classification schemes, 4) spatial scale of observation, and 5) calculation and
selection of pertinent spectral and structural variables.
I focused on emergent individuals of tree species which could be visually
identified both in the remote sensing data and in the field. Species were chosen
based on their importance in ecological research or conservation efforts and my
ability to find a sufficiently-large sample size. I opted to manually delineate crowns
9
so that I could focus on questions regarding species discrimination rather than tree
detection and delineation (i.e., object segmentation).
Several different classification schemes have been used for ITC species
classification, including neural networks (Gong et al., 1997), linear discriminant
analysis (Brandtberg et al., 2003; Fung et al., 1998; Gong et al., 1997; Holmgren &
Persson, 2004), spectral angle mapper (Xiao et al., 1999), spectral mixture analysis
(Xiao et al., 2004), decision trees (Preston et al., 1999), nearest-neighbor (Wang et
al., 2004); parallelpiped (Meyer et al., 1996), and the popular Gaussian maximum
likelihood classifier (Gougeon, 1995; Meyer et al., 1996; Leckie & Gougeon, 1999;
Key et al., 2001; Leckie, Gougeon, Hill et al., 2003). I devote three chapters to an
exploration of the relative trade-off among five classification techniques: maximum
likelihood (ML), spectral angle mapper (SAM), linear discriminant analysis (LDA),
decision trees (DT), and multiple endmember spectral mixture analysis (MESMA).
These analyses explored several methods for isolating and extracting information
from the hyperspectral and lidar data for optimal species discrimination. The issue
of spatial scale of spectral measurement, and its effect on species separability is also
analyzed and discussed.
The general objectives of this research were to:
1. Develop hyperspectral techniques for classifying tropical rain forest tree
species
2. Identify the optimal spectral regions and spatial scale for species
discrimination
10
3. Evaluate the importance of lidar-derived crown structure information for
species discrimination
4. Assess lidar technology for ecological analyses of tropical rain forests
1.3. Study site and remote-sensing datasets
Study site overview: The La Selva Biological Station
My research was conducted using data acquired at the La Selva Biological
Station (LSBS), located in the Atlantic lowlands of north-east Costa Rica in the
Sarapiquí canton (84°00'13.0" W, 10°25'52.5" N). LSBS is a 1614-ha reserve that
contains a mixture of old-growth terra firme, swamp, secondary and selectivelylogged forests, as well as plantations, developed areas with buildings and mowed
grass, and abandoned pastures with large grasses, shrubs and remnant trees (Fig.
1.1). Precipitation averages 4244 mm annually, with a comparatively dry season
from January to April and a second smaller dry season from August to October
(Frankie et al, 1974; Organization for Tropical Studies [OTS] meteorological data
1957-2003, http://www.ots.ac.cr). The old-growth forest (Fig. 1.1) is classified as a
Tropical Wet Forest in the Holdridge Life Zone System and is characterized by a
species-rich, multi-layered community of trees, palms, lianas, and other terrestrial
and epiphytic plants (Holdridge, 1971; Hartshorn & Hammel, 1994). There are at
least 400 species of hardwood trees in the reserve. Although some overstory trees
can be completely deciduous for a part of the year, mainly in the dry season, the
canopy is considered evergreen (Frankie et al., 1974; Hartshorn & Hammel, 1994).
The geomorphology of the landscape is structured by two main features: 1) highly
11
eroded lava flows that contain a system of alternating ridges and stream valleybottoms separated by steep slopes, and 2) flat to gently undulating alluvial terraces
(Sanford et al., 1994). The reserve is covered by a 50 x 100-m grid, oriented 32
degrees from North, with permanent monuments at each grid intersection. The local
grid coordinate system permits researchers to accurately geo-locate field data in a
local Geographic Information System (GIS) Cartesian coordinate system.
La Selva
1
Costa Rica
Km
HYDICE extent
FLI- MAP extent
Rivers
Land Use
Developed Areas
Selectively- logged
Old- growth Forest
Secondary Forest
Pasture
Plantation
Swamp
Figure 1.1. The La Selva Biological Station study site and extent of HYDICE
hyperspectral and FLI-MAP lidar datasets.
Hyperspectral data
For this research, I had access to a unique, high spatial resolution hyperspectral
dataset from the airborne HYperspectral Digital Imagery Collection Experiment
12
(HYDICE) sensor (Basedow et al., 1995).
The U.S. Spectral Information
Technology Application Center (SITAC) flew HYDICE over LSBS in March 30,
1998, which corresponds to the end of the drier season in the region. HYDICE is a
push-broom, indium-antimonide hyperspectral sensor that measures 210 bands
covering the 400-2500 nm region of the electromagnetic spectrum (Basedow et al.,
1995). The LSBS flights were flown at a 3.17 to 3.20-km altitude between 7:558:27 am local time (13:55 to 14:27 UTC). Six runs of 0.5-km wide, variable length,
1.6-m spatial resolution data (0.5 mrad instantaneous field of view; IFOV) were
acquired over old-growth forest, secondary forest, selectively-logged forest, tree
plantations, pastures and the nearby town of La Guaria. The section of the dataset
used in my research is shown in Figure 1.1. Pre-processing of the HYDICE imagery
is described in Chapter 3.
Lidar data
Another unique dataset used in this research was from a high spatial resolution
lidar sensor called FLI-MAP (John E. Chance & Associates, Lafayette, Louisiana).
These data were acquired from a helicopter on September 12 and 13, 1997. FLIMAP is a small-footprint, first-return 900-nm laser sensor that has a 8000-Hz pulse
rate, 30-degree scan angle, 2-mrad beam divergence, typically 9-point/m2 sampling
density, 30-cm footprint spacing, and a rated vertical accuracy of ~10-cm (Blair &
Hofton, 1999, Huising & Gomes Pereira, 1998). The 10-cm footprints (Hofton et
al., 2002) were converted to a Triangular Irregular Network (TIN) (John E. Chance
& Associates), and the final product delivered for analysis by the data distributor
13
(Fig. 2a) was a raster digital surface model (DSM) interpolated from the TIN with
0.33-m cell support (each containing a height). The DSM covered a 754-ha area of
the LSBS reserve (Fig. 1.1).
1.4. Description of Chapters
Chapter 2 is an assessment of small-footprint lidar technology for the estimation
of ground elevation and tree heights in tropical landscapes. This research provided
the foundation for Chapter 6, which sought to combine lidar and hyperspectral data
for ITC species discrimination. In Chapter 2, I developed a method to retrieve
canopy heights, as a canopy height model (DCM), from the original lidar height
surface (the DSM). The DCM was then used in Chapter 6 for ITC species
discrimination (discussed below). A key step in creating an accurate DCM was to
estimate sub-canopy terrain elevation from the DSM. With the DCM as a final goal,
I developed a ground-retrieval scheme to find terrain points. I then evaluated
geostatistical techniques for interpolating a DTM from the points. The accuracy of
the lidar-derived DTM was rigorously tested with a comparison to field-survey
points. I discuss how differences in DTM accuracy are related to terrain slope and
land-use factors. Next, the DSM and the derived DTM were used to calculate the
DCM, and the accuracy of lidar-derived estimates of stem heights at individual tree
and plot scales was assessed with reference to comparable field measurements. This
research is considered an extreme test of lidar technology because: 1) there are
several forest types with dense, structurally-complex canopies that severely restrict
14
ground-level light transmission and lidar returns (Clark et al., 1996; Drake,
Dubayah, Clark et al., 2002; Montgomery & Chazdon, 2001), 2) all vegetation
classes are in “leaf-on” conditions, exacerbating ground-retrieval difficulties, and 3)
the study site covers a range of terrain conditions. A version of this research is
published in Remote Sensing of Environment (Clark, Clark, & Roberts, 2004).
Chapter 3 begins my analyses of ITC species discrimination. In this chapter, I
examine the relative trade-offs between spectral regions, spatial scale of
measurement, and traditional classification schemes for species discrimination using
hyperspectral reflectance bands. Field spectrometer and airborne hyperspectral
reflectance spectra were acquired from seven species of emergent trees at LSBS,
permitting analyses at leaf, pixel and crown scales. My main objectives in Chapter 3
were to: 1) use statistical tests to determine if spectral variation among TRF tree
species (interspecific) is greater than spectral variation within species (conspecific),
thereby permitting spectral-based species discrimination; 2) identify the spatial scale
and spectral regions that provide optimal discrimination among TRF emergent tree
species; 3) develop an analytical framework for the species-level classification of
ITCs based on their pixel- or crown-scale reflectance spectra; 4) assess the relative
importance of narrowband hyperspectral versus broadband multispectral imagery for
species identification; and, 5) assess the efficacy of three traditional classifiers:
LDA, ML, and SAM. This research is published in Remote Sensing of Environment
(Clark et al., 2005).
In Chapter 4, I developed a purely hyperspectral-based method for ITC species
classification. I first calculated a suite of hyperspectral metrics that respond to
15
crown structure and photosynthetic pigments, water and other biochemical
absorption features. At tissue, pixel and crown scales, I tested for significant
differences in mean response of these spectral metrics among my study tree species.
I then assessed the utility of these metrics for ITC species discrimination using a DT
classification scheme. A manuscript of this chapter is ready for submission to
Remote Sensing of Environment.
Chapter 5 explores MESMA for the species-level classification of ITCs. Starting
with a large spectral library of image and laboratory spectra, I devised an automated
approach to select optimal endmembers for two- and three-endmember MESMA
models. In particular, the selection scheme sought to find within-species specialists
while excluding among-species generalists. The selected endmembers were then
used in MESMA to classify ITC species. This chapter will be submitted
simultaneously with Chapter 4 for publication in Remote Sensing of Environment.
The objective of Chapter 6 is to extend the spectral-based analyses from
Chapters 3 and 4 to include lidar-derived, crown structure metrics calculated from
the DCM (Chapter 2). I first assessed how crown structure, as quantified by lidar
metrics, varies among species. I then explored the benefits of lidar-derived structure
information for species classification.
In Chapter 7, I summarize the findings from Chapters 2-6 and outline general
conclusions about these new forms of remote sensing technology for tropical tree
species discrimination. I also make recommendations for future research.
16
CHAPTER 2: Lidar estimation of sub-canopy elevation and tree heights
2.1. Introduction
Digital terrain models (DTMs) and canopy height estimates are two important
remote sensing products for studies of TRF ecology and management. DTMs
describe the variation of elevation across a landscape and have been used in
applications including mapping drainage basin geomorphology (Yin & Wang, 1999),
flood modeling (Bates & De Roo, 2000), calculation of biophysical controls on
vegetation distribution (e.g., temperature, solar radiation) (Dymond & Johnson,
2002), and spatial analysis of soil properties (Gessler et al., 2000).
Canopy height estimates from remote sensing technology have a variety of
potential applications in TRF, such as calculating surface roughness for atmosphereland interaction models (Raupach, 1994), spatial analyses of forest dynamics, such
as canopy gap formation, distribution and turn-over (Birnbaum, 2001), identifying
plant species (Brandtberg et al., 2003; Holmgren & Persson, 2004), mapping of
wildlife habitat (Hinsley et al., 2002), modeling canopy rain interception (Herwitz &
Slye, 1995), and modeling light penetration (Clark et al., 1996; Montgomery &
Chazdon, 2001). Vegetation height is allometrically related to forest structure
parameters, such as estimated aboveground biomass (Brown et al., 1995). Because
roughly half of biomass is composed of carbon, improvements in our ability to map
biomass through remote sensing will translate into better estimates of carbon stocks
and flux at broad scales. Such advances are particularly important in tropical forests,
which contain a large proportion of terrestrial carbon, and consequently have the
17
greatest potential to increase atmospheric carbon dioxide from deforestation (Dixon
et al, 1994). Although passive optical and active synthetic aperture radar (SAR)
signals and associated metrics are sensitive to forest aboveground biomass variation
(Ranson et al., 1997; Sader et al., 1989), biomass estimates from these sensors tend
to saturate at the high biomass levels typically found in tropical forests (Imhoff,
1995; Luckman et al., 1998, Steininger 2000).
There have been relatively few published applications that have used smallfootprint lidar sensors in tropical rain forest environments (Blair & Hofton, 1999;
Hofton et al., 2002). However, studies using the large-footprint LVIS sensor in
Costa Rica have shown that there is immense potential of lidar technology for TRF
research and monitoring efforts (Blair & Hofton, 1999; Drake, Dubayah, Clark et al.,
2002, Drake, Dubayah, Knox et al., 2002; Hofton et al., 2002; Weishampel et al.,
2000). For example, Drake, Dubayah, Clark et al. (2002) showed that lidar metrics
applied to large-footprint waveforms can accurately predict aboveground biomass
over a wide range of tropical forest conditions without saturation.
2.1.1. Digital terrain models and lidar
Photogrammetry has been the traditional source of broad-scale DTMs throughout
the world. There has been an increase in the use of IfSAR (interferometric synthetic
aperture radar) and lidar active sensors as more economic alternatives for producing
DTMs (Hodgson et al., 2003; Petzold et al., 1999). In many developing countries,
where most tropical forests occur, the best publicly available topographic
information is from the Shuttle Radar Topography Mission (SRTM), which used
18
IfSAR technology to produce a global terrestrial DTM with a 90-m horizontal
resolution (Rabus et al., 2003). However, the relative vertical accuracy of this
product at the 50-100 km scale is roughly 6 m due to errors from several systematic
and random factors that can only be partially reduced by data post-processing.
Lidar-derived DTMs estimated in open areas or under areas with low vegetation
can have vertical accuracies ranging from 0.06 to 0.61-m root-mean-square error
(RMSE) (Cobby et al., 2001; Huising & Gomes Pereira, 1998). Trunks, branches
and leaves in dense vegetation tend to cause multiple-scattering reflections or
absorption of the emitted laser energy so that fewer backscattered returns are
reflected directly from the ground (Harding et al., 2001; Hofton et al., 2002). This
effect increases with more canopy closure, canopy depth (or volume) and structural
complexity (Harding et al., 2001; Hodgson et al., 2003; Hofton et al., 2002), and it is
expected to be more severe for first-return only lidar systems because recorded
returns generally come from the canopy (Magnussen & Boudewyn, 1998). The
result is that the RMSE between the lidar-derived DTM and reference elevation data
tends to increase in areas of dense vegetation because: 1) there are fewer groundreturn samples for DTM surface interpolation, and 2) those samples selected as
ground may actually be understory vegetation, logs or rocks (Cobby et al., 2001,
Hodgson et al. 2003; Raber et al., 2002). When non-ground samples are included in
the interpolation, mean-signed residual error will be positive (lidar-reference), i.e.,
the lidar DTM overestimates the reference elevation. When actual ground samples
are mistakenly filtered out of the data before DTM interpolation, as discussed below,
results are less predictable and depend on local topographic curvature. DTM peaks
19
may be clipped or valleys filled due to inadequate retrieval of ground samples,
resulting in the underestimation or overestimation of local elevation, respectively.
Additional sources of error in lidar-derived DTMs include vertical and horizontal
error in positioning the laser platform, laser scan angle, surface reflectivity, and
slope of the terrain, all of which combined can add 0.20-2.00 m of error to an
elevation estimate (Baltsavias, 1999b; Hofton et al., 2002; Huising & Gomes
Pereira, 1998; Kraus & Pfeifer, 1998).
The filtering of vegetation from sub-canopy ground returns for the interpolation
of DTMs, or “bald Earth”, has been an active area of research. However, most
algorithms for small-footprint lidar data are proprietary, and reported instrument and
DTM accuracies are often poorly documented and generally assumed to be measured
under optimal conditions—flat areas with no vegetation (Baltsavias, 1999a; Huising
& Gomes Pereira, 1998;). Some vegetation-filtering techniques include
morphological filters and statistical analyses of heights in a neighborhood, and may
be fully-automated or involve some human interpretation (Huising & Gomes Pereira,
1998). Kraus and Pfeifer (1998) used an automated, iterative technique that
interpolated a mean surface from the lidar cloud of xyz points and then successively
removed or down-weighted points with residuals higher than a specified threshold.
A relatively simple approach is to find local-minima relative to neighboring samples
at a specified scale and/or search configuration (Cobby et al., 2001; Petzold et al.,
1999). Resulting ground samples (i.e., local minima) must then be interpolated to
form a surface.
20
There have been relatively few studies that provide rigorous accuracy
assessments of lidar-derived DTMs under dense forest canopy with leaves, be they
simple or composite leaves in hardwood forests or needles in conifer forests.
Working with last-return lidar data flown over a leaf-on pine/deciduous forest
landscape, Hodgson et al. (2003) identified ground points through a combination of
proprietary software and human interpretation. A comparison of DTM elevation
against 1470 survey-grade field measurements had an overall RMSE of 0.93 m.
DTM error differed significantly by land use. Although RMSE was 0.33 m for low
grass, it increased to 1.22 m and 1.53 m for the more structurally-complex
scrubs/shrub and deciduous vegetation types, respectively. Furthermore, these
researchers found that in the dense, multi-layered shrub/scrub class, there was a
highly significant increase in DTM error of roughly 2 m from lowest (0-2 deg.) to
steepest (6-8 deg.) slopes, which the authors attributed to vertical inaccuracies over
relatively short horizontal distances under complex canopy. Cobby et al. (2001)
developed an automated ground-retrieval scheme for a floodplain environment that
included deciduous forests with leaf-on conditions. An initial DTM was interpolated
from local-minima cells retrieved from non-overlapping, 5 x 5-pixel windows (10-m
side) overlaid on a last-return DSM (2-m support). The final DTM was achieved by
tailoring the ground-retrieval algorithm to short and tall vegetation classes. While
terrain under short vegetation could be predicted with a 0.17-m RMSE (n=5), the
RMSE was 3.99 m (n=12) under deciduous forests on steeper slopes (10-15
degrees). Using last-return filtered proprietary methods, Reutebuch et al. (2003)
found that a DTM under conifer plantations had a 0.32-m RMSE and a +0.22-m
21
mean-signed error (overestimated, n=347 reference points). There was only a slight
0.15-m increase in mean error between clearcut and uncut forest classes.
Hodgson et al. (2003) used a triangulated irregular network (TIN) to interpolate
the DTM from ground points, while Cobby et al. (2001) used bilinear interpolation.
Additional interpolation techniques include co-variance linear weighting approaches,
such as inverse distance weighting described below (Reutebuch et al., 2003), which
can help smooth the high-frequency effects of spurious vegetation points, especially
when applied in iterative filtering schemes (Lohmann & Koch, 1999).
2.1.2. Geostatistical methods for DTM generation
I used geostatistical techniques to interpolate a DTM from 0.33-m support
ground cells derived from a vegetation-filtering algorithm. In the interpolation
process, cells were treated as xyz-coordinate samples (i.e., points). Two common
geostatistical interpolation algorithms, inverse distance weighted (IDW) (Bartier &
Keller, 1996; Isaaks & Srivastava, 1989) and ordinary kriging (OK) (Goovaerts,
1997; Isaaks & Srivastava, 1989; Lloyd & Atkinson, 2002a, 2002b), were
considered in this research. Both techniques involve a weighted linear combination
of neighboring data samples in estimating values at unsampled locations. In the case
of IDW, weighting of neighboring samples is based solely upon an estimation
location-to-sample inverse-distance function, along with a user-specified power
weight factor (Bartier & Keller, 1996). The further away a sample is from the
estimation location, and the greater the power weight factor, the less influence the
sample will have on the estimate value. OK uses a distance-covariance model to
22
weight samples based on their distance from the estimation location, as well as to
down-weight samples that are clustered in space—an effect which tends to smooth
the variance in the surface (Isaaks & Srivastava, 1989). The kriging covariance
model is generally developed with reference to an empirical semivariogram
(Goovaerts, 1997; Isaaks & Srivastava, 1989). Recently, Lloyd and Atkinson
(2002a, 2002b) reported that the more sophisticated OK technique was more
accurate than IDW when interpolating DTMs from photo-interpreted (Lloyd and
Atkinson 2002a) or lidar-derived (Lloyd and Atkinson 2002b) elevation points.
2.1.3. Lidar vegetation height and forest structure estimation
With discrete-return small-footprint systems, vegetation height is calculated as
the difference between the original footprint heights and the bald-Earth DTM. The
result is a set of estimated canopy heights with a footprint-scale support.
Alternatively, a DSM is interpolated from footprint heights and subtracted from the
DTM, thereby creating a canopy surface with height values recorded in square
pixels or cells, i.e., a digital canopy model (DCM).
Small-footprint lidar technology now permits the detection and segmentation of
individual tree crowns from fine spatial resolution DCMs (Brandtberg et al., 2003;
Persson et al., 2002). Metrics applied to either DCM cells or footprint heights from
within crown segments have been used to estimate the height and structure of
individual crowns (Brandtberg et al., 2003; Gaveau & Hill, 2003; Næsset & Økland,
2002; Persson et al., 2002; Popescu et al., 2003). Such measures also hold promise
for tree species classification (Brandtberg et al., 2003; Holmgren & Persson, 2004).
23
Persson et al. (2002) estimated individual conifer tree heights from DCM-cell
maxima with a model r2 of ~0.98 and RMS error of 0.63 m. Næsset & Økland
(2002) estimated conifer tree height with the maximum of first-return footprints
within crowns. The regression model explained 75% of the variance with a
prediction RMS error of 0.23 m. Tree height was slightly overestimated by 0.18 m
(3.15-m s.d.) in Næsset & Økland (2002), yet underestimated by 1.13 m in Persson
et al. (2002). Næsset & Økland (2002) concluded that to reduce underestimation,
high footprint density is needed to increase the probability of detecting the tops of
conifer crowns. I know of only two studies that estimated heights for hardwood
individuals. Working in West Virginia with forests in leaf-off winter conditions,
Brandtberg et al. (2003) estimated hardwood tree heights with first-return footprints
from within crown segments. A regression model explained 69% of the variance
(RMSE not reported). There was an underestimation of height for taller trees, which
the authors attributed to the low probability of the laser detecting the maximum
crown height with leaf-off conditions, as well as random error in field measurements
and inaccuracy in ground retrieval. In the United Kingdom, Gaveau and Hill (2003)
estimated leaf-on hardwood tree and shrub heights from a first-return DCM with a
model that explained 95% of the variance (1.89-m RMSE). The authors also
presented strong evidence that an underestimation of tree heights (-2.12 m meansigned error) resulted from laser pulses penetrating into the crown before reflecting a
detectable first-return signal. The depth of signal penetration depended on variation
of foliage and branches in the crown at the fine scale of a lidar footprint.
24
Stand-scale height of conifer-dominated forests have been predicted with varying
levels of success from small-footprint lidar metrics (Næsset, 1997, 2002, Næsset &
Økland, 2002). Several authors have shown that laser underestimation can be
minimized by comparing a quantile of ground measurements (i.e., mean, maximum
height) to a certain quantile of upper-most canopy returns in the stand (Magnussen &
Boudewyn, 1998; Næsset,1997, 2002). For example, Næsset (1997) found that the
mean of lidar maxima (footprints) from a grid-overlay of square cells (i.e., 15 x 15
m) estimated average conifer stand height with a mean-signed error of -0.40 m to
1.90 m (8-20 m tree height; 1.5-ha stands; RMSE not reported). For mixed conifer
and hardwood forests with 6 to 29-m tall trees, Næsset (2002) used multipleregression analyses to relate several lidar metrics to mean and dominant tree height
at plot (0.02 ha) and stand (average 1.6 ha) scales. Lidar data included first and last
pulses and metrics included the maximum, mean, density and various percentquantile measurements within the plots. Model cross-validation RMS errors for
young to mature forest plots ranged from 0.05 to 0.07 m (R2=0.82 to 0.95) and 0.07
to 0.08 m (R2=0.74 to 0.93) for mean and dominant plot-scale height, respectively.
Stand-scale heights were estimated as the mean of plot-scale lidar estimates, and
resulting models explained 92% and 87% of the variance for mean and dominant
tree height, respectively (RMSE not reported).
I am not aware of any studies that have used small-footprint lidar to estimate
stand-scale height for areas composed of evergreen, tropical tree species (i.e.,
hardwood trees, palms). However, one recent study by Lim et al. (2003) focused on
leaf-on hardwood forests in Ontario, Canada. Plot-scale (0.04 ha) Lorey’s mean
25
height was estimated using either the mean or maximum of lidar pulses (first and last
return). Regression models explained 66% and 86% of the variance, and it was
found that tree height was underestimated (RMSE not reported).
Fine-scale lidar data has also been used to estimate stand-scale biophysical
properties, such as volume, biomass, crown diameter and diameter breast height
(Lim et al., 2003; Popescu et al., 2003). Advances in analyzing large-footprint
waveforms have produced metrics that can estimate biophysical parameters over
dense conifer and broadleaf forests with considerable accuracy [see Lefsky et al.
(2002) for a recent review of the literature]. Drake, Dubayah, Clark et al. (2002)
used multiple-regression and metrics from LVIS waveforms to predict plot-scale
(025 to 0.5 ha) basal area, aboveground biomass, and quadratic-mean stem diameter
over a range of tropical forest types at the La Selva Biological Station, Costa Rica;
their resulting models explained 93%, 72% and 93% of the variance, respectively,
and models did not saturate with increasing forest height and complexity.
2.2. Methods
2.2.1. Topographic reference data
The DTM derived in this research was compared to 3859 in situ elevation points
(Table 2.1) that were surveyed in 1991 using optical-leveling techniques (Hofton et
al., 2002). In the comparison, co-located survey points and DTM cells were
determined by linking points to the nearest 1-m DTM cell centroid based on x,y
coordinates. Survey data covered the range of land-use and geomorphic conditions
found in the reserve. There were 1321 points measured at grid intersections, while
26
the remaining points did not have permanent markers and were located off the grid,
mainly in the northwest to southeast direction between grid intersections. All survey
points were transformed to the Universal Transverse Mercator (UTM), WGS-84
datum coordinate system from the local La Selva coordinate system using a leastsquares affine transformation based on 5 differential-GPS points, as used by Hofton
et al. (2002). DGPS measurements were taken from permanent towers or in open
areas to avoid obstruction by vegetation. Differential corrections were performed
using data from a base station at La Selva. The transformation had an overall RMSE
of 0.57 m (XRMS = 0.48 m, YRMS = 0.31 m). The vertical values of the survey points
were shifted from the local reference mean sea level (MSL) to the WGS-84 ellipsoid
MSL using the constant 11.44-m offset published by Hofton et al. (2002). This
offset was verified by selecting the 4 to 5 nearest FLI-MAP (discussed below) lidar
returns around three grid monuments in open areas with very flat, mowed lawn. The
average vertical offset between surveyed and lidar-derived elevations was also found
to be 11.44 m. Neither the FLI-MAP data nor the reference data were referenced
relative to a geoid, and so the elevation offset between the datasets is constant across
the study extent
DTM error was analyzed relative to slope and land-use factors. A total of 932
survey points at grid intersections in old-growth forest were further classified into
one of four slope classes based on field measurements from a separate forest
structure study (Clark et al., 1999; field data summarized in Table 2.1). Slope
classes included: 0-3, 3-10, 10-20 and > 20 degrees. A total of 2060 survey points
were classified into one of seven land-use categories based on an overlay operation
27
with an existing year 2000 land-use map (Fig. 2.1) derived from historic aerial
photographs, IKONOS imagery and survey maps (OTS, unpublished data). The
land-use categories, from short-simple to tall-structurally-complex vegetation, were:
Developed Areas, Pastures, Plantations (abandoned and current), Secondary forests
(1 to 34 years old), Selectively-logged Forest (including 50 year-old, abandoned
agroforestry areas), Swamp forests and Old-growth forests. To limit the
confounding effect of slope on DTM error (discussed in results), only survey points
on slopes less than 10 degrees were considered in the land-use analysis. A mask
delineating areas with slopes less than 10 degrees was created by sampling the
DTM-derived slope (Arc/Info 8.0.1, Environmental Systems Research Institute,
Redlands, California) after filtering with an averaging 3x3-cell filter.
2.2.2. Individual tree heights
In the field, the maximum height of a crown was estimated with a handheld laser
range-finder (Impulse-200LR, Laser Technology Inc., Englewood, Colorado) for
individual trees taller than 20-m. Several readings of the crown apex (i.e., highest
foliage) were taken per individual from different locations (if possible) and trees
were included in analyses if their measurement standard deviation was less than 1 m.
The mean of the readings for an individual established the tree’s height. Only the >
20-m height class was considered because field measurements were taken 3-4 years
after the FLI-MAP overflight, and rapid tree growth in smaller tree-height classes
during this time would confound the analysis.
28
The heights of 21 trees in pasture with isolated crowns were measured in the
field during July, 2001 (~ 9 species, Table 2.2). The crown center (centroid) for
each pasture tree was estimated using DGPS readings from a Trimble Pro XL GPS
and on-site base station (Trimble Navigation Limited, Sunnyvale, California) and
subsequent visual spatial adjustment of the DGPS points with reference to the DCM.
Between February and October 2000, the heights of 59 old-growth forest
emergent trees were measured (11 species, Table 2.2) in the field and their trunks
were geo-located relative to the closest grid marker using a tape measure and
compass. Trunk locations were then transformed to UTM coordinates (Hofton et al.,
2002) and crown centroids were identified through visual adjustment of trunk points
with reference to the DCM. I selected canopy-emergent trees in old-growth forest
for analysis because 1) they are major components of overall forest biomass (Clark
& Clark, 1996), 2) they were easy to locate unequivocally in the DCM, and 3) their
height growth in the 3 years between field measurement and the lidar overflight was
expected to be minimal.
2.2.3. Plot-scale stem heights
For plot-scale reference data, I used stem heights from a 1997 census of
agroforestry plantations within La Selva (Menalled et al., 1998). The 32 plantation
plots considered in this research were on 1-year (n=9), 4-year (n=9) and 16-year
(n=14) cutting cycles and had stems spaced 2-m apart at roughly equal densities at
the time of the census. Each plot had a total area of 0.04, 0.08, and 0.13 ha,
respectively, and the oldest trees in a plot were 6 years old in 1997. All trees were
29
one of three species: Hyeronima alchorneoides (Euphorbiaceae), Cedrela odorata
(Meliaceae), or Cordia alliodora (Borginaceae). Seven of the fourteen 16-yr
rotation plots included a mixture of one of the three dominant tree species as well as
a sub-canopy palm Euterpe oleracea (mean height 7.13 m) and a smaller understory
shrub Heliconia imbricata (mean height 1.53 m). The 2-dimensional area of each
field plot (i.e., a rectangle polygon) was digitized in a GIS and located with visual
reference to the DCM.
In the field, height measurements were recorded with a laser range-finder for
every individual stem in the plot. Plot mean height (Table 2.2) was calculated in two
ways: 1) as the average of heights for all stems in the plot (termed “all stems”), and
2) the average of heights for trees in the plot (termed “tree stems”). Note that the
difference between these two metrics lies in the inclusion of a mix of canopy (tree)
and sub-canopy (palm and shrub) stem heights in the plot mean-height calculation
for 7 of the 32 plots; the mean height of the other 25 plots, which did not include the
palms and shrubs, did not differ between the two methods.
2.2.4. Accuracy of field height measurements
I established the vertical accuracy of the laser range-finder in two ways. In one
case, five height estimates at 10, 20, 30 and 40-m distances from a 15.00-m pole in
an open area were found to have an overall vertical mean error of +0.18 ± 0.14 SD m
(slight overestimation). There was no significant effect on the mean height
measured with distance (Kruskal-Wallis non-parametric ANOVA, α=0.5). Next I
analyzed the stem height of actual trees. Crown apices were more difficult to
30
identify from the ground than the top of the 15-m pole. Four trees in open areas had
range-finder height estimates made before being cut down. The actual height of the
tree was then measured from its horizontal position on the ground, including the
stump. The four trees had an average ground-measured height of 27.3 m and a
range-finder estimate mean error of +1.45 ± 1.67 SD m (overestimation). In
summary, reference measurements using the laser range-finder have their own
associated error, mostly due to the human observer’s inability to identify the apex of
the tree crown, rather than distance from the target. The overall bias appears to be
an overestimation of tree height by the laser range-finder technique.
2.2.4. Small-footprint lidar data
The small-footprint lidar dataset used in this research (FLI-MAP) was introduced
in Chapter 1. Previous lidar research that used this dataset include analyses of LVIS
pseudo-waveforms synthesized from the DSM cells (Blair & Hofton, 1999) and a
comparison of LVIS to FLI-MAP sub-canopy elevation retrieval (Hofton et al.,
2002). In the latter study, FLI-MAP elevation readings over areas of structurallycomplex vegetation (e.g., old-growth forest) were found to be mostly from canopy,
not the ground, as is to be expected from a first-return sensor. However, in a 25-m
diameter circular area, there were at least some FLI-MAP hits that coincided with
survey elevations (Hofton et al., 2002), which may be due in part to sensor’s
relatively high sampling density (Huising & Gomes Pereira, 1998).
31
2.2.5. Ground retrieval and DTM interpolation
To create a DTM, FLI-MAP data processing involved two major steps: 1)
identifying ground-return cells in the DSM, and 2) subsequent geostatistical
interpolation of ground cells to form a DTM. I developed two simple groundretrieval algorithms that sought a balance between processing speed and the
minimization of elevation error over the range of vegetation and terrain conditions
found at La Selva.
2.2.6. Local-minima ground-retrieval scheme
The local-minima algorithm proceeded as follows: a grid of non-overlapping,
square cells was overlaid on top of the original DSM. Within each grid cell, one
local-minima DSM cell (0.33-m support) was selected and identified as a ground
return. This procedure resulted in a population of ground-return cells for each of the
five grid scales considered independently: 5, 10, 15, 20 and 30 m. At a relatively
coarse scale (e.g., a 25-m diameter circle), there is expected to be at least one lidar
pulse that penetrates to or near the ground surface. This pulse would register as a
cell of relatively low height in the DSM; and thus, the above local-minima scheme is
analogous to selecting the lowest return in a square footprint of a specified scale
(i.e., 5 m, 10 m, etc.). Ground-return cells identified at each scale were then used in
separate geostatistical interpolation schemes (described below) that generated DTMs
with a 1-m cell size. Samples from each DTM were compared to 3859 co-located
reference points. The overall RMS errors of the resulting DTMs were used as the
basis for the selection of a final ground-retrieval/interpolation scheme.
32
2.2.7. DTM interpolation schemes
The inverse distance weighted (IDW) and ordinary kriging (OK) geostatistical
techniques were assessed for the interpolation of the final DTM. In both procedures,
DSM ground-return cells (0.33-m support) were treated as representing a cloud of
xyz-coordinate points, with the cell’s centroid defining the x and y coordinates and
the cell’s DSM height defining the z value. Note that these points have less variance
than they would if derived directly from the original lidar footprints, instead of from
the DSM. This is because DSM cell values were smoothed from TIN-interpolation
and subsequent rasterization.
IDW interpolation of ground points was conducted using Arc/Info 8.0.1
(Environmental Systems Research Institute, Redlands, California). The IDW
interpolation moved in a circular window of 50-m radius and included at least 4 data
points. The IDW power parameter specifies how sample points are weighted with
distance from the interpolation node. A lower power relaxes the weighting of near
points, causing more smoothing in the interpolated surface (Isaaks & Srivastava,
1989). Initial experiments based on minimizing RMSE identified that a value of 2
was the most appropriate IDW power for this data set.
For OK, an isotropic normal-score transform variogram (Goovaerts, 1997) for
each set of ground points (i.e., 5 m, 10 m, etc. scales) was manually fitted using
Variowin v2.21 (Pannatier, 1996). These models contained a short- and long-range
nested structure and a very small nugget effect (Goovaerts, 1997; Isaaks &
Srivastava, 1989). Ordinary kriging of the normal-score data points was performed
33
using the GSLIB v2.0 software package (Deutsch & Journel, 1992) with a 250-m
search radius, and the inclusion of 4 to 12 data points at each interpolation node.
2.2.8. Iterative-addition ground-retrieval scheme
An additional ground-retrieval technique was tested that iteratively added in
ground points at successively finer scales (Fig 3). Point selection proceeds through
3 iterations, each with 3 steps (in Fig. 2.3; Iterations 1-3 are rows and steps are
columns). In the first step of the first iteration (Fig. 2.3; row 1, col. 1), a coarsescale grid (i.e., 20-m cells) is overlaid on the original DSM and one 0.33-m minima
cell is selected from within each 20 x 20-m cell. In Step 2 (Fig.3; row 1, col. 2),
DSM minima cells are converted to xyz points. In Step 3 (Fig. 2.3; row 1, col. 3), a
first-pass DTM is created from IDW interpolation with a 5-m support. The first
iteration is essentially the local-minima scheme as described above, but with a 5-m
IDW interpolation.
For Step 1 of Iterations 2 and 3 (Fig. 2.3; rows 2 & 3, col. 1), the coarse-scale
grid is reduced by 5-m (i.e., from 20 to 15 m, or 15 to 10 m) and overlaid on the
DSM; again, minima cells are identified, converted to xyz points, and then added to
the population of minima points from the previous iteration. In Step 2 of Iteration 2
& 3 (Fig. 2.3; rows 2 & 3, col. 2), minima-point z values are subtracted from the
DTM resulting from the previous iteration (rows 1 or 2, col. 3), and individual points
0.25-m higher than the DTM (marked with “+” in Fig. 2.3) are removed from the
population of minima points (Fig. 2.3; rows 2 or 3, col. 2). In Iteration 2, Step 3
(Fig. 2.3; row 2, col. 3) another preliminary DTM is interpolated with IDW for use
34
in the third iteration. At the end of the Iteration 3, a final population of ground
points is interpolated with IDW or Ordinary Kriging (OK) with a 1-m support (Fig.
2.3; row 3, col. 3).
The effect on the mean-signed error of the final DTM using a 0.25, 0.50 and 1.0m residual threshold was evaluated; and subsequently, 0.25 m was deemed a
conservative residual threshold in that it allowed finer-scale variation to be added
back to the population of ground-return points, while it minimized the risk of
including understory vegetation points (and subsequently inflating the mean-signed
error). To minimize processing time, an IDW interpolator with 5-m support was
chosen for generating intermediary DTMs (Fig. 2.3; rows 1 or 2 only, col. 3); future
research should test the impact of using a finer-scale interpolation (i.e., 1 m) or an
ordinary kriging interpolator on the accuracy of the final DTM product.
2.2.9. Sink removal
Roughly 0.5% of the DSM-derived, xyz points from the ground-retrieval schemes
contained very low, spurious points that created local sinks in the interpolated DTM.
Some sinks were prevented from entering the population of ground points by
imposing two restrictions on DSM local-minima cells: 1) that the minima cell not lie
60-m lower than its highest neighboring DSM cell (no trees or other objects in the
scene are expected to be taller than 60-m high), and 2) the cell not lie lower than the
minimum study site elevation (30 m). Sink xyz points that persisted through the
ground-retrieval process were automatically removed by calculating their sink depth
from a first-pass DTM interpolation. All sink points greater than 5-m in depth were
35
deleted from the population of ground points, and the DTM surface was reinterpolated.
2.2.10. Digital Terrain Model (DTM) accuracy assessment
For each DTM interpolated, error was calculated by subtracting DTM elevation
from the elevation measured at co-located field-survey points (DTM – reference;
sensu Hodgson et al., 2003). Calculated error statistics included mean-signed error,
mean absolute error (MAE), and RMSE (Isaaks & Srivastava, 1989).
The final DTM with the lowest RMSE was submitted to a more rigorous analysis
that sought to understand the variation of error across the TRF landscape. I asked
the following questions:
1) Does DTM error differ among slope inclination categories in old-growth
forest?
2) Does DTM error differ among land-use categories?
As reviewed in Section 2.1.1, relatively dense and structurally-complex
vegetation decreases DTM accuracy. Lidar range error also increases non-linearly
with greater surface-inclination angle (Baltsavias, 1999a); and, when dense
vegetation types occur in areas of steep terrain, these combined vertical errors
decrease the accuracy of the interpolated surface over relatively short horizontal
distances (Hodgson et al. 2003).
The square-root transformation of the mean absolute error (SqrtMAE) was used
in an ANOVA to test for statistical significance of the above hypotheses (sensu
36
Hodgson et al. 2003), with α=0.05. The square-root transformation of MAE was
chosen to adjust the positively-skewed distribution of MAE to a normal distribution,
as required by ANOVA. Since the minimum distance between reference points was
0.28 m, the classical ANOVA assumption of independence of residuals was suspect;
therefore, I opted to use the generalized least-squares ANOVA with a residual
variance-covariance weighting, formulated by Gotway and Cressie (1990). The
weightings take into consideration residual autocorrelation and reduce chances of
committing Type I errors. This technique requires a variance-covariance model of
the SqrtMAE residuals (SqrtMAE minus a trend component). The trend component
of the SqrtMAE points was estimated using ordinary kriging of the local mean
(Goovaerts, 1997). An isotropic variogram of the resulting SqrtMAE residuals was
modeled, and this model in conjunction with the data-to-data distances of the survey
points were used to estimate the SqrtMAE residual variance-covariance sub-matrices
for each class (e.g., Old-growth, Developed Areas, etc.), as required by the spatial
ANOVA (Gotway and Cressie, 1990).
2.2.11. Stem height accuracy assessment
A digital canopy model (DCM) was calculated by subtracting the final 1-m
support DTM, bald-Earth surface from the lidar DSM (i.e., DCM = DSM – DTM,
Fig. 2.2). The DCM maintained the 0.33-m support of the DSM. Individual tree and
plot-scale stem heights were estimated from the DCM using metrics that had
comparable spatial supports and calculations as those used in the field (sensu Drake,
Dubayah, Knox et al., 2002; Næsset, 1997).
37
Plot-scale DCM metrics considered the same 2-dimensional area as field plots,
i.e., same spatial support. In field measurements, plot mean height was calculated as
the average height of stems spaced 2-m apart. Following Næsset (1997), I devised a
comparable lidar metric that averaged the DCM local-maxima (i.e., maximum
height) from each 2 x 2-m cell in a grid overlaying each plot. Alternatively, I also
calculated the average of all heights (i.e., all DCM cells) in the plot extent. These
two lidar metrics are referred to as Mean2x2 and MeanALL, respectively.
Field measures of individual tree height located the highest leaf or twig in a
crown (i.e., crown apex). Similarly, I based DCM metrics for estimating tree height
on a sample of cells from within the crown. All DCM cells within 5 horizontal
meters of the crown centroid were sampled and tree height was estimated as the
maximum value of the cells (referred to as “Maximum”). Since crowns had
multiple peaks of high foliage, field measures may have missed the highest crown
apex. In consideration of this potential field error, I devised another metric that
averaged DCM cells whose values were above the 95-percent quantile (sensu Ritchie
et al., 1993; referred to as Mean95).
At individual tree and plot scales of analysis, height error was calculated as the
difference between the DCM and field height metrics (DCM – field, sensu Næsset,
1997). These errors, which are residuals from a 1:1 relationship, were summarized
with the statistics mean-signed error, MAE and standard deviation (Isaaks &
Srivastava, 1989).
I also evaluated the linear relationship between lidar and field height metrics
with regression analyses, which provided a coefficient of determination (r2) value
38
and a model RMSE (referred to as RMSEm). RMSEm was calculated from
prediction residuals (sensu Drake, Dubayah, Clark et al., 2002; Næsset, 2002,
Næsset & Økland 2002) and is the general error expected if the regression model
were applied to DCM-derived estimates, i.e., calibration using ground
measurements. Because I did not have an independent set of field data for model
validation, model prediction residuals needed for calculating RMSEm were acquired
through a common cross-validation procedure (Drake, Dubayah, Clark et al., 2002;
Næsset, 2002; Næsset & Økland 2002; Popescu et al., 2003).
2.3. Results
2.3.1. DTM generation and accuracy assessment
The five local-minima ground-retrieval scales tested in this research (i.e., 5, 10,
15, 20 and 30 m) had overall RMSE errors ranging from 2.29 to 5.09 m, using either
IDW or OK for surface interpolation (data not shown). The scale with the lowest
RMSE for both interpolation methods was found to be 20 m. For this tropical
landscape and lidar sampling density, 20 m appears to be the near-optimum scale to
identify ground returns with the local-minima approach, and so only the results from
this scale will be discussed (Table 2.3). This optimal scale is likely determined by
the average crown dimensions and canopy gap characteristics in old-growth forest,
which comprises 69% of the study area. By visual interpretation of colorized 1-m
IKONOS imagery, Clark, Read, et al. (2004) measured mean maximum crown
diameter to be 19.6 m for all trees within old-growth plots. In this research, mean
maximum crown diameter of the very largest, canopy-emergent trees in old-growth
39
forests were measured from the ground to be 27.2 m (n=36). At 15-m scales or
finer, there was a higher probability that a DSM local-minima cell could lie
completely within one large crown or many connected crowns. If these crowns are
vertically dense, or have understory trees underneath, the DSM local minima would
likely be from upper- or lower-canopy leaves or branches, not the ground. At 30-m
scales or greater, there was a higher chance of selecting a canopy-gap DSM cell with
a height from or near the ground; however, a trade-off exists in that local minima at
coarser scales fail to sample fine-scale topography. As shown in Figure 2.4 (green
points), the 20-m local-minima scheme identified DSM-cell minima (subsequently
converted to xyz points) on the periphery of dense vegetation, such as in the shortest
plantation plots, on the edges of emergent tree crowns and in more open swamps.
The 20-m scale, iterative-addition scheme could identify more of these local-minima
cells (xyz points), while still avoiding the areas of densest vegetation (Fig. 2.4, green
and red points together).
The population of xyz points for each ground-retrieval scheme were used to
model variograms needed for OK interpolation. The normal-score variogram from
the 20-m scale, iterative-addition dataset (method used for final DTM interpolation,
Fig. 2.5a) was found to have a zero nugget effect, exponential short-range structure
of 380-m range with a sill of 0.31, and an additional long-range spherical structure
spanning to a 2000-m range, comprising an additional 0.20 of the total sill
(maximum sill modeled in study extent was 0.51).
Considering all 3859 survey points distributed throughout the landscape, the
correlation (r) between DTM and reference elevation was +0.99 (p<0.0001) for IDW
40
interpolation and +1.00 (p<0.0001) for OK interpolation, using either groundretrieval scheme. In terms of RMSE, OK also performed better than IDW for DTM
interpolation. There was a 0.18-m RMSE difference between the best IDW and OK
interpolated surfaces (Table 2.3). The inclusion of finer-scale local-minima points
using the iterative-addition scheme (i.e., iteratively adding local-minima from 20,15
to 10-m scales) improved the OK-interpolation RMSE by 0.10 m, resulting in an
overall RMSE of 2.29 m (Fig. 2.5b; Table 2.3). All DTMs had a positive meansigned error, and so they tended to overestimate elevation. Although this
overestimation was up to 0.87-m higher with OK relative to IDW interpolation, the
mean absolute error (MAE) using OK was lower (Table 2.3). Indeed, the OK DTM
was significantly more accurate than IDW interpolation, using either groundretrieval technique (n=3859, paired t-test of sqrtMAE, p<0.0001 for both
comparisons). As expected, OK was found to reduce the variance of errors (Table
2.3, standard deviation). This smoothing of the variance across space tends to
minimize the influence of spurious understory vegetation or downed trunks that are
inevitably included in the DTM interpolation. In contrast, IDW does not exhibit this
desirable smoothing effect as strongly as OK. The OK iterative-addition method
was selected to generate the final DTM because the surface had a relatively low
error variance and the lowest overall RMSE and MAE. This DTM was free of
obvious underestimation (i.e., sinks) or overestimation (i.e., peaks) errors and
revealed remarkable fine-scale detail of geomorphology (Fig. 2.2b).
As expected, elevation error in the final DTM was not distributed evenly across
the landscape. Mean absolute error (MAE) was significantly different between slope
41
inclination classes in old-growth forest (p<0.0001). Mean-signed error was positive
(overestimation of DTM elevation) and increased 0.62 m from the lowest to the
steepest slope classes (Fig. 2.6, +1.07-m [0-3°] vs. +1.69-m [>20°] mean-signed
error). The difference in MAE between slopes 0-3 and 0-10 degrees was not
significantly different (paired t-test; p=0.68); however, the MAE on slopes ≤ 10
degrees was significantly less than on slopes 10-20 and > 20 degrees (Fig. 2.6, lines
connecting classes; p < 0.004 both comparisons). Under old-growth forest, RMS
error was lowest on slopes ≤ 10 degrees (2.21 m, 0-3° & 3-10° combined) and
highest on slopes greater than 20 degrees (3.09 m; Fig. 2.6), indicating that very
steep slopes had the largest overestimation bias and error variance (Isaaks &
Srivastava, 1989).
I next assessed DTM elevation error relative to land-use. The slope error
analysis indicated that slopes ≤ 10 degrees had statistically similar DTM errors, so I
limited land-use analyses to survey points on slopes ≤ 10 degrees to avoid the
confounding effects of slope error.
For all land-use classes on slopes less than 10 degrees, the overall RMSE for the
DTM was 1.72 m (Table 2.4). There was a highly significant difference in mean
absolute error between the land-use classes (p<0.0001). The classes with relatively
few trees or shrubs, Developed Areas and Abandoned Pastures, had a slight
elevation underestimation (mean-signed error) of -0.58 and -0.28 m, respectively
(Fig. 2.7; Table 2.4). As expected, DTM elevation was overestimated (positive
mean-signed error) under forest canopies (Fig. 2.7; Table 2.4). DTM error was most
severe in old-growth forests, which had extremely dense, multi-layered canopies.
42
Old-growth forests had significantly higher MAE (pair-comparisons with other
classes, p<0.05), the largest mean-signed error (+1.01-m overestimation), and
highest RMSE (1.95 m) relative to all other classes (Fig 7; Table 2.4). In contrast,
the classes with very little to no canopy cover, Developed Areas and Abandoned
Pastures, had the lowest RMS errors (1.02 and 1.10 m, respectively). Interestingly,
classes with continuous canopies of relatively low tree height, Secondary forests and
Agroforestry Plantations, had statistically similar absolute errors (Fig 7., lines).
2.3.2. Landscape view of canopy-surface heights
Differences in land-use depicted in Figure 2.1 are clearly observable as mean
height and textural variations in the DCM (Fig. 2.8). As expected, mean canopysurface height is greatest for old-growth, swamp and selectively-logged forests,
which have relatively low human disturbance (Table 2.5). Several linear strips of
low elevation running southwest to northeast indicate areas where no lidar footprints
were recorded due to gaps in the flight-line (Fig. 2.8). These areas were avoided in
tree-height analyses.
2.3.3. Individual- tree height estimation
In estimating old-growth tree height, the Maximum lidar metric had a
significantly lower mean absolute error than the Mean95 lidar metric (Table 2.6;
p<0.001, t-test). The relative MAE error for Maximum and Mean95, respectively,
was 8.1% and 8.7% of the mean height for individual emergent trees measured in the
field (Tables 2.2 & 2.6). Mean-signed error was negative, indicating that the DCM
43
metrics underestimate actual emergent tree heights in old-growth forest. Linear
regression models for predicting old-growth tree height explained 51% and 50% of
the variance (RMSEm of 4.15 and 4.19 m) for Maximum and Mean95, respectively
(Fig. 2.9 a & b).
Lidar metric underestimation can be seen as points clustering
above the 1:1-relationship line in the scatter plot (Figure 2.9 a & b).
As with old-growth heights, pasture tree heights were underestimated by the
DCM metrics (Table 2.6, negative mean-signed error). However, DCM height
estimates of pasture trees had a lower MAE, with 2.33 and 2.84 m for Maximum and
Mean95, respectively (Table 2.6). Mean absolute errors using Maximum and Mean95
were 7.4% and 9.0%, respectively, of the mean field height of pasture trees (Tables
2.2 & 2.6). For pasture trees, the Maximum lidar metric had a significantly lower
mean absolute error than the Mean95 lidar metric (Table 2.6; p<0.001, t-test). Height
errors were also less variable than those of old-growth trees (Table 2.6, standard
deviations; Fig. 2.9, 1:1-relationship line). For pasture trees, the linear models
relating lidar-derived to field-derived height explained 95% of the variance (using
either lidar metric) and relationships were stronger than those observed for oldgrowth tree heights (Fig. 2.9 c & d vs. Fig. 2.9 a & b). Model RMS errors for
pasture trees (2.41 and 2.48 m for Maximum and Mean95, respectively) were roughly
half of those expected in applying old-growth height models (Fig. 2.9).
In summary, direct estimation of individual tree height from the DCM was
more accurate (i.e., less absolute error) when using the Maximum lidar metric, which
calculated height as the maximum DCM cell within 5-m radius of the crown
centroid. Furthermore, pasture height estimates using either DCM metric had less
44
absolute error and stronger linear relationships with field measurements than those
from old-growth forest (Table 2.6; Fig. 2.9 a & c).
2.3.4. Plot-scale height estimation
In terms of mean-signed error, MAE and standard deviation of errors, the
MeanALL lidar metric out-performed the Mean2x2 method in estimating plot mean
height, calculated from all or tree-only stems (Table 2.7). Considering tree-only
stems, mean-signed error was slightly negative (-0.36 m) when using the mean of all
DCM cells in the plot (i.e., MeanALL metric), indicating a slight underestimation of
plot mean height. In contrast, I found that the Mean2x2 lidar metric overestimated
mean stem height (+1.55 mean-signed error). Mean absolute error was significantly
less for MeanALL relative to Mean2x2 when considering the plot mean height of all or
tree-only stems (p<0.0001, paired t-tests).
All plot-scale regression models were highly significant (p<0.001). Using the
MeanALL lidar metric, the mean height of tree stems in plantation plots was predicted
with a model r2 of 0.97 and RMSEm of 1.08 m (Fig. 2.10b—trees-only plots),
whereas this relationship dropped to an r2 of 0.84 and RMSEm rose to 2.26 m when
predicting mean height for all canopy and sub-canopy stems in plots (Fig. 2.10d—all
stem plots). A similar pattern was observed in using the Mean2x2 lidar metric (Fig.
2.10 a & c). Height was overestimated for the 7 all-stem plots, and so those points
fall below the regression line (Fig. 2.10 c & d—open circles) and decrease the
overall strength of the models. In general, the best plot-scale relationship was
between the MeanALL lidar metric and the trees-only field metric. Errors between
45
this combination of lidar and field metrics had the lowest 1:1-relationship bias (0.36-m mean-signed error), strongest linear relationship (r2 = 0.97) and lowest model
RMSE (1.08 m).
2.4. Discussion
2.4.1. Lidar ground retrieval in a tropical landscape
The overall 2.29-m RMSE of the DTM generated using an iterative groundretrieval and ordinary kriging interpolation scheme (Table 2.3) fell short of the
decimeter accuracies reported for other laser sensors (Cobby et al., 2001; Huising &
Pereira, 1998). However, those accuracies were achieved only under ideal
conditions favoring retrieval of ground points, such as in areas with relatively flat
terrain, without vegetation or with deciduous vegetation in a leaf-off state. In this
tropical wet forest landscape, areas exhibiting these characteristics are rare. The
“bald Earth” geomorphology in the study area is not flat due to its volcanic history
and leaf-on canopy persists throughout the entire year. In comparing areas of similar
slope (< 10 degrees), my DTM error is 0.79 m greater than the error (1.72 vs. 0.93-m
RMSE) observed by Hodgson et al. (2003) when analyzing a lidar-derived DTM
from a temperate-zone, deciduous (leaf-on) landscape. My research used a fullyautomated technique to retrieve ground xyz points from a DSM, which was
interpolated from the original lidar footprints. In contrast, Hodgson and colleagues
used proprietary software to identify ground-return footprints and then used a visualinterpretation step to remove spurious xyz points prior to DTM interpolation. Due
to my first-return lidar data, DSM-interpolation smoothing, fully-automated ground
46
retrieval, and extremely-dense vegetation, I would expect my DTM to have
relatively greater overall error than the Hodgson et al. (2003) study. Despite the
many challenges in using lidar over a TRF landscape, my observed DTM error is
well below the USGS maximum-permitted RMSE of 7 m for Level-3 DTMs, the
highest-quality nationwide products publicly offered in the United States (Hodgson
et al. 2003); and furthermore, this lidar-generated DTM is a vast improvement over
previously-available topographic data at this important tropical research site.
I found that DTM accuracy followed a gradient in human-disturbance intensity.
The DTM had the greatest RMS error (1.95 m) in old-growth forests. The structure
of these forests is composed of multi-layered leaves and branches maintained by a
fine-scale, tree-fall disturbance regime. In effect, this structure is a dense media
which acts to absorb or multi-scatter photons of near-infrared laser light, thereby
reducing the probability that photons reach the ground and return to the sensor;
consequentially, my ground-retrieval algorithm had a greater chance of mistaking
sub-canopy returns for ground returns in old-growth forests, and the accuracy of the
interpolated DTM decreased. In contrast, DTM accuracy was 0.93-m greater in
developed areas of the reserve (1.02-m RMSE), which had scattered overstory trees
and shrubs, mowed grass and cobble roads. Even under these more ideal groundretrieval conditions, the algorithm failed to entirely separate ground from overlying
vegetation height. When I considered just those survey points in developed areas
with mowed grass or roads, overstory vegetation > 3-m away, and on flat terrain,
RMSE was 0.58 m and mean-signed-error was -0.49 ± 31sd m (n=20). This error
can be considered the elevation error related to sensor artifacts when using first47
return laser pulses, i.e., footprint xyz positioning, scan angle (Baltsavias, 1999b;
Huising & Gomes Pereira, 1998). A last-return, small-footprint system would likely
provide better ground-retrieval performance in this densely-vegetated landscape.
Indeed, last-return systems are generally deployed for DTM-generating campaigns
(Hodgson et al., 2003; Lefsky et al., 2002) because there is a higher probability that
laser returns will come from ground reflections. Flying a lidar sensor with higher
sample density (i.e., more frequent postings) is also expected to increase the chances
of detecting the ground (Hodgson et al., 2003).
In structurally-complex old-growth forest, I found vertical errors to increase 0.67
m on the steepest slopes relative to flattest areas (Fig. 2.6, > 20 vs. 0-3-deg. class
RMSE). In comparison, Hodgson et al. (2003) found a 2-m greater RMSE on flat
versus steep slopes under structurally-complex, leaf-on vegetation (0-2 vs. 6-8 deg.
classes). Observed DTM vertical error has both horizontal and vertical error
components (Hodgson et al., 2003), such as footprint positioning and instrument
range errors, respectively. As mentioned above, these factors combined may
contribute 0.58-m to DTM elevation error. On steep slopes, vertical errors between
DSM-derived ground points translate into large DTM interpolation errors relative to
flat areas, which have more points of similar elevation that act to down-weight the
impact of a spurious point during interpolation. Also, most DTM error analyses
assume that errors in reference data are negligible. In this research, field-survey data
suffer from horizontal, planimetric errors (~0.57 m) due to the transformation
between the local and UTM WGS-84 coordinate and datum systems. On the
steepest slopes (44 degrees [Table 2.1]), planimetric error could introduce up to
48
0.40-m vertical error in reference data. When I considered only reference points on
slopes ≤ 10 degrees, RMS error decreased from 2.29 to 1.72 m (all land-use). This
0.57-m difference in DTM error can easily be explained by greater reference error
combined with instrument-related vertical error on steep slopes. However, my data
do not permit me to rigorously assess these differences in lidar- and referencerelated vertical errors.
Compared to previously published accuracies achieved with the large-footprint
LVIS sensor at LSBS (Hofton et al., 2002), FLI-MAP elevation estimation had
better overall performance across the entire landscape (5.64-m [LVIS] vs. 2.29-m
[FLI-MAP] RMSE). Also, the DTM from my research performed better than LVIS
on flatter areas under the range of vegetation conditions in the study area (1.72-m
[FLI-MAP, slopes ≤ 10º] RMSE vs. 2.42-m [LVIS, slopes ≤ 3º]). In the LVIS study,
the estimated elevation at centroids of 25-m diameter footprints were compared
directly to the nearest reference point, and so steep slopes compounded the vertical
error associated with planimetric discrepancies between centroids and survey points.
In contrast, the FLI-MAP DSM had a 0.33-m spatial support that permitted more
precise planimetric location of ground-retrieved samples. In addition, fine spatial
support provided denser ground sampling relative to the LVIS footprints. Due to
these benefits of using a small-footprint sensor, my observed elevation error was
relatively low compared to the Hofton et al. (2002) study, even on the steepest
slopes and under the densest vegetation in the landscape (2.42 to 3.09-m RMSE
from flat to steep slopes in old-growth forest).
49
As was found by Lloyd and Atkinson (2002b), ordinary kriging was a better
DTM interpolator of lidar ground-retrieved samples than the more conventional
IDW technique. Although the mean-signed error was smaller for IDW, OK
interpolations had smaller mean absolute errors, lower RMS errors, and smaller error
variance. The variance-dampening effect afforded by OK is desirable because it
tends to reduce the impact of spurious sub-canopy vegetation samples that have
inadvertently passed through the vegetation filter. There are other variants of
kriging that could be useful to DTM interpolation of lidar data. For example, cokriging (Goovaerts, 1997) the ground samples with a co-varying variable, such as
land-use type, optical reflectance data or a textural variable calculated from the
original lidar data, could provide a robust means of adjusting the data covariance
weighting to various landscape units.
2.4.2. Lidar estimation of tropical vegetation height
Individual tree heights were underestimated by the lidar metrics (Table 2.6,
negative mean-signed error). These results concur with those by Brandtberg et al.,
(2002), Gaveau and Hill (2003), and Persson et al. (2002). The FLI-MAP sensor
used in my research detected first-return signals from densely-sampled small
footprints, and so it is unlikely that this underestimation error is entirely from the
laser missing the upper-most reaches of the crown (Næsset & Økland, 2002). Based
on recent findings by Gaveau and Hill (2003) for leaf-on hardwood trees, I believe
underestimation is partly caused by laser pulses penetrating below crown surfaces
until inner-crown materials reflect a detectable first-return signal.
50
Some of the discrepancy between lidar and field estimates is also related to DTM
error. In abandoned pastures, the DTM tends to slightly underestimate elevation by
0.28 m on average and RMS error was 1.10 m (Table 2.4). In contrast, the DTM
error under old-growth forests was significantly greater. The DTM was
overestimated by 1.01 m on average and RMS error increased to 1.95 m (Table 2.4).
In abandoned pastures, broad areas without tree canopy permitted more laser energy
to reach the ground and return to the sensor, resulting in better DTM accuracy. In
old-growth forest, dense canopy causes DTM overestimation that effectively “clips”
trees at their base when the DCM is calculated. In terms of RMSE, DTM-related
error can account for up to 47% and 42% of the model error observed for old-growth
and pasture trees, respectively (Fig. 2.10, Maximum). I found that DTM error
increases on steep slopes in old-growth forests; however, slope was not linearly
related to the absolute difference between lidar- and field-derived old-growth tree
heights (Maximum metric [Fig. 2.9a], r2 = 0.003; p=0.68). I therefore conclude that
slope-related effects on tree height estimates are not severe.
Another source of random error in the predictive models is from field reference
data (Brandtberg et al., 2003; Gaveau & Hill, 2003; Persson et al., 2002). In my
research, field measurements of individual tree heights were taken 3 years after the
lidar mission, and many trees could have grown higher in that time. This growth
would result in the field estimates being higher than the lidar measurements, thus
adding to the underestimation bias seen here. Although my field data included only
trees with precise measurements (i.e., measurement std. dev. ≤1 m), high precision
does not ensure absolute accuracy. In this research, I found that directly-measured
51
stem height was overestimated by the laser range-finder. Range-finder estimation
precision was between 0.14 to 1.67 m (s.d. of errors). Again, this field measurement
error would create an additional source of apparent underestimation in lidar-derived
estimates and may account for up to 69% and 40% of model RMSE for pasture and
old-growth tree heights, respectively (Fig. 2.9, Maximum).
Given the difficulties in estimating elevation and field stem heights in oldgrowth forest, it is not surprising that regression models had a stronger linear
relationship for pasture trees than old-growth trees (r2 0.95 vs. 0.51; Fig. 2.9 a & c,
Maximum;). The strength of the relationship for pasture trees is greater than that
observed by Brandtberg et al. (2003) in estimating deciduous trees in leaf-off
conditions (r2 0.69), and the same as that observed by Gaveau and Hill (2003) in
estimating leaf-on hardwood trees and shrubs (r2 0.95). All studies used the firstreturns from small-footprint lidar data to estimate vegetation height and observed an
underestimation for tall trees. My research and the Brandtberg et al. (2003) study
used the local-maxima of samples from within a crown, while Gaveau and Hill
(2003) compared DCM cells directly to field points measured over crowns with high
xyz-coordinate precision. The two-return lidar data used by the other two studies, as
well as the leaf-off conditions in the Brandtberg et al. (2003) study, likely permitted
better ground retrieval and higher DTM accuracy relative to that which can be
expected with a first-return lidar flown over an evergreen TRF landscape. Relatively
high DTM error in old-growth forests likely explains why old-growth TRF tree
heights were not as reliably predicted as those of tropical pasture trees (my study)
and temperate-zone hardwood trees (Gaveau & Hill, 2003). Despite the limitations
52
of my elevation and field-height estimates, my pasture tree calibration model was
surprisingly strong relative to both the Brandtberg et al. (2003) and Gaveau and Hill
(2003) models. Hardwood trees in leaf-on conditions were underestimated by -1.58
m and -2.12 (mean-signed error) in my study and the Gaveau and Hill (2003),
respectively. Exposed leaves in hardwood trees favor the detection of the crown
surface because they readily transmit and multiple-scatter near-infrared laser light
(Grant, 1987). In contrast, the Brandtberg et al. (2003) model had weaker accuracy
because leaf-off crowns expose only twigs and branches to the sensor and allow
greater laser penetration into the canopy before returning a signal.
Regression models for estimating individual conifer tree height had RMS errors
of 0.23 and 0.63 m (r2 0.75 and ~0.98) in Næsset & Økland (2002) and Perrson et al.
(2002), respectively. In my research, I observed a 2.41-m model RMSE for pasture
trees (r2 0.95; Fig 9 c, Maximum) and Gaveau and Hill (2003) observed a 1.98-m
model RMSE for deciduous hardwoods and shrubs (r2 0.95). The two conifer
studies used first and last return, small-footprint lidar data. Assuming DTM error is
equal between the studies, it thus appears that lidar calibration models (i.e., linearregression models) have a lower RMS error in estimating conifer tree heights than
they do for leaf-on, hardwood trees. However, more comparative studies between
tropical and temperate species, tree conditions (i.e., stress, senescence), lidar sensors
and associated analytical methods are needed to confirm this accuracy assessment.
The linear correlation between lidar and reference measurements of plot mean
height of tree stems was very strong (Fig. 2.10 b). The 0.97 r2 exceeded the 0.380.49, 0.91, 0.82-0.95, and 0.91 r2 values reported for conifer-dominated stands by
53
Magnussen and Boudewyn (1998), Næsset (1997, 2002), and Næsset & Økland
(2002), respectively, and the 0.68 r2 value for leaf-on hardwood stands (Lim et al.,
2003) As was found in this and other lidar studies (Lim et al., 2003; Magnussen &
Boudewyn, 1998; Næsset, 1997), mean tree height at the plot scale was slightly
underestimated by the MeanALL lidar metric (Table 2.7). In my research, this
underestimation (negative mean-signed error) was 0.36 m, while it was 2.1 to 4.1 m
and 0.70 m in the Næsset (1997) and Magnussen and Boudewyn (1998) studies,
respectively. At the stand scale, conifer underestimation likely results because too
few footprints detect the upper-most twigs and branches of conical trees (Magnussen
& Boudewyn, 1998). In my research, laser penetration into the hardwood tree
canopy is likely an issue (Gaveau & Hall, 2003), and square DSM cells may not
always record the height of the highest surface material within the cell—an
underestimation that would persist in DCM heights. Furthermore, my field metric
considered the mean heights of tree stems (i.e., maximum height of individual
crowns), whereas the MeanALL lidar metric averaged all canopy-surface heights (i.e.,
DCM cells) in the plot, not just DCM crown-maxima, and this lidar measurement is
expected to be low relative to the field measurement (Magnussen & Boudewyn,
1998).
Although the Mean2x2 metric was designed to isolate DCM crown-maxima (i.e.,
stem heights of trees spaced 2-m apart), the metric overestimated mean tree height
by an average of 1.55 m (Table 2.7). These results are contrary to those of Næsset
(1997), who found that a local-maxima grid-overlay metric reduced mean-signed
error to insignificant levels. However, the Næsset (1997) and other “grid-overlay”
54
studies have used last-return lidar systems with relatively sparse sampling densities,
properties which combined lower the probability of recording a return from the
upper-most canopy. This problem may have been compounded by the geometry of
conifer crowns in the previous studies (Magnussen & Boudewyn, 1998; Næsset,
1997, 2002). In contrast to TRF tree crowns, which are broad and generally
hemispherical in shape, conical crowns expose much less upper-most canopy
material to a lidar sensor (Magnussen & Boudewyn, 1998). A grid-overlay lidar
metric can compensate for missing the crown by selecting only the highest heights in
the dataset. In the case of FLI-MAP, which has both a high sampling density and
first-returns, there is less chance of missing the upper-canopy, especially over these
agroforestry plots that contained even-aged stands with a relatively flat uppercanopy layer of broad leaves. In this case, the grid-overlay metric needlessly overcompensates the plot-height estimate upward relative to the reference height.
Given the dense upper-canopy layer in the older plots, FLI-MAP first-return
signals were not expected to be sensitive to sub-canopy stem heights. This subcanopy insensitivity was revealed clearly in linear-regression analyses. When
considering all stems in plots, regression models had weaker relationships and
higher model RMS errors than those built when considering just tree-stem heights
(Fig. 2.10). The averaging of canopy and sub-canopy stem heights from the field
lowered the mean height for seven of the reference plots, and since FLI-MAP was
sensitive mainly to the upper-canopy trees, the lidar-metric estimation was high
relative to the field measurement.
55
As was found by Næsset & Økland (2002) for conifer trees, the estimation of
height for individual hardwood trees (Table 2.6) was not as accurate as that found
for plot-scale estimation (Table 2.7). This finding is not surprising, because at the
plot scale, both lidar and field measurement errors are minimized by averaging many
co-located lidar and reference data points.
2.5. Conclusions
In this chapter, I found that a small-footprint, first-return lidar sensor can be used
to predict sub-canopy elevation with 2.29-m accuracy in a tropical landscape. The
accuracy of the elevation surface was significantly affected by vegetation cover,
with largest errors detected in areas with structurally-complex old-growth forests,
especially on steep slopes. The digital terrain model that resulted from this analysis
had a 1-m spatial support, which makes it an ideal input for other ecological or
management applications such as modeling of inundation zones, hill-slope
processes, and habitat associations.
I showed that small-footprint lidar systems have potential for the estimation of
individual tree height of tropical, hardwood species in leaf-on conditions. For oldgrowth forest emergent trees and isolated abandoned-pasture trees greater than 20-m
tall, individual tree heights could be estimated directly from a lidar canopy-height
surface (i.e., DCM) with mean absolute errors that were 8.1% and 7.4% of mean
field heights, respectively. Models for individual old-growth and pasture trees
explained 59% and 95% of the variance, with model RMS errors of 2.41-m and
4.15-m, respectively; however, as was found in other studies (Brandtberg et al.,
56
2003; Persson et al., 2002), it was difficult to separate lidar-related error from
reference measurement error.
Plot-scale, mean stem height for trees within plantation stands was estimated
from the DCM with much greater accuracy than for individual tree heights. The
best plot-scale models explained 87% of the variance when estimating the mean
height of all canopy and sub-canopy stems in plots, while 97% of the variance when
considering just canopy-tree stems. These results are encouraging in that a stand
(i.e., a plot in this study) is often an important scale for many broad-scale studies and
management decisions.
57
Table 2.1. Topographic reference data.
Count
Min
Max
Elevation (m)
3859
38.78 141.37
Slope (deg) a
932
0.00
44.00
a
Measured only in old-growth forest.
Mean
70.93
12.94
Median
62.38
12.00
Table 2.2. Height reference data for individual trees and plots (units in meters).
Count
Min
Max
Mean Median S.D.
Old-growth trees
59
31.00
56.39
45.04
44.34
5.76
Pasture trees
21
20.01
51.19
31.64
27.54 10.73
Plots (All stems)
32
0.38
17.53
6.19
4.70
5.27
Plots (Tree stems)
32
0.38
18.53
7.28
4.70
6.17
Table 2.3. Error of Digital terrain model (DTM) elevation estimates (n=3859, units
in meters).a
Interp.d (Ground retrievale)
RMSE b MAE c Mean S.D.
IDW (Local-minima)
2.47
1.78 +0.68 2.37
IDW (Iterative-addition)
2.47
1.75 +0.08 2.47
OK (Local-minima)
2.39
1.69 +1.10 2.13
OK (Iterative-addition)
2.29
1.60 |+0.97 2.07
d
e
Interp. (Ground retrieval )
Median
Min
Max
IDW (Local-minima)
0.69 -14.80
17.64
IDW (Iterative-addition)
0.29 -16.29
16.54
OK (Local-minima)
0.94 -14.77
18.62
OK (Iterative-addition)
0.83 -17.14
16.89
a
Error is calculated as DTM – field-survey elevation (deviation from 1:1
relationship).
b
RMSE is root mean-square error.
c
MAE is mean absolute error.
d
Interpolation schemes are: IDW = inverse distance weighted, or OK = ordinary
kriging.
e
Ground-retrieval schemes are: Local-minima = minimum DSM cell in a 20x20-m
grid, or Iterative-addition = iteratively adding local-minima from 20,15 to 10-m
scales.
58
Table 2.4. DTM error on slopes ≤10 degrees summarized by land use (in meters).a
Land-use Class
Mean ± SD (RMSE)
Developed Areas
-0.53 ± 0.88 (1.02)
Abandoned Pastures
-0.28 ± 1.08 (1.10)
Selectively-logged Forest +0.21 ± 1.61 (1.62)
Secondary Forest
+0.49 ± 1.36 (1.44)
Agroforestry Plantations +0.62 ± 1.08 (1.24)
Swamp Forest
+0.72 ± 1.48 (1.64)
Old-growth Forest
+1.01 ± 1.66 (1.95)
All classes
+0.66 ± 1.59 (1.72)
a
DTM interpolation from OK, iterative-addition scheme.
Table 2.5. Canopy-surface height summarized by land use.a
Mean
Land-use Class
No. Cells
Area (Ha)
(m)
S.D. (m)
Abandoned Pastures
2,126,838
23.6
3.3
5.0
Developed Areas
826,562
9.2
8.0
8.2
Agroforestry Plantations
3,287,534
36.5
12.8
9.0
Secondary Forest
5,395,687
60.0
12.9
8.2
Old-growth Forest
43,675,481
485.3
20.5
8.8
Swamp Forest
4,210,094
46.8
20.5
11.1
Selectively-logged Forest
3,271,990
36.4
20.8
11.0
a
Calculated by from the digital canopy model (DCM), 0.33-m cells.
Table 2.6. Estimation error of individual tree heights.a Units are in meters.
MAE b
Mean
S.D.
Min
Max
c
Old-growth Trees (n=59)
Maximum d
3.67
-2.11
4.11
-5.02
11.59
d
Mean95
3.94
-2.72
4.15
-4.40
12.10
Pasture Trees c (n=21)
Maximum d
2.33
-1.58
2.40
-5.11
2.68
Mean95 d
2.84
-2.30
2.46
-5.64
2.35
a
d
c
Error is calculated as: lidar – field height metric (deviation from 1:1 relationship)
b
MAE is mean absolute error.
c
Field metric is the maximum height in an individual crown (old-growth emergent
or isolated pasture trees).
d
Lidar metrics are: Maximum = maximum value of DCM cells; Mean95 = mean of
cells above 95% quantile.
59
Table 2.7. Estimation error of plot-mean stem heights.a Units are in meters.
MAE b
Mean
S.D.
Min
Max
Tree stems c (n=32)
Mean2x2d
1.74
+1.55
1.55
-0.52
4.92
MeanALLd
0.90
-0.36
1.09
-3.69
1.60
All stems c (n=32)
Mean2x2d
2.82
+2.63
2.78
-0.52
8.08
MeanALLd
1.64
+0.73
2.32
-3.69
5.91
a
d
c
Error is calculated as: lidar – field height metric (deviation from 1:1 relationship)
b
MAE is mean absolute error
c
Field metrics are mean of individual stem heights for tree species only (Tree stems)
or trees and other species (All stems).
d
Lidar metrics are: Mean2x2 = mean of maximum DCM values in a 2x2-m grid
overlaid on plot; MeanALL = mean of all DCM values in plot
60
!
La Selva
0
0.5
1
2
Km
Costa Rica
Lidar data extent
Rivers
Land Use
Developed Areas
Selectively- logged Forest
Old- growth Forest
Secondary Forest
Abandoned Pasture
K
Agroforestry Plantation
Swamp Forest
Figure 2.1. La Selva Biological Station study area land-use and lidar data extent.
61
A
Lidar
Height (m)
128.8
45.2
DSM
B
Elevation (m)
92.6
45.2
DTM
C
Vegetation
Height (m)
48.6
0.0
DCM
Figure 2.2. A 3-dimensional perspective of a 250 x 250-m subset of the lidar raster
products, all covering the same geographic extent in an old-growth forest: a) the
unprocessed lidar height surface (i.e., digital surface model, DSM), b) the estimated
sub-canopy elevation surface (i.e., digital terrain model, DTM), and c) the estimated
vegetation height surface (i.e., digital canopy model, DCM) resulting from the
subtraction of the DTM (b) from the DSM (a). In (c), canopy emergent trees are redtone, concave domes.
62
Iteration 1
20 m
5-m IDW
Iteration 2
15 m
5-m IDW
Iteration 3
10 m
or
1-m IDW
Step 1.
DSM grid overlay
find 0.33-m DSM minima
cell in each coarse-scale
grid cell (e.g., 20 x 20 m)
Step 2.
Calculate elevation
residuals (point-DTM)
1-m OK
Step 3.
Interpolate DTM
from filtered xyz points
remove xyz points (+)
with residuals
> 0.25 m
Figure 2.3. Conceptual flow of the iterative-addition ground-retrieval scheme.
Abbreviations are: digital surface model (DSM), digital terrain model (DTM),
inverse distance weighed interpolation (IDW), and ordinary kriging interpolation
(OK). Note: to create the figure, the same 60 x 60-m extent was placed on top of the
actual DSM, and ground samples outside the extent were included in the DTM
interpolation (Step 3, all iterations).
63
A
Height: 0
B
47 m
Height: 0
54 m
Figure 2.4. An example of ground-retrieved samples (i.e., xyz points) for a)
agroforestry plantations, and b) an old-growth forest with an open swamp. Points
shown are for the 20-m local-minima retrieval scheme (green dots) and the iterativeaddition scheme spanning 20, 15 and 10-m scales (green and red dots combined).
The underlying gray-scale surface is from the final DCM (Fig 2.2c; Fig. 2.8). Extent
is 210 x 180 m.
64
A
B
0.5
140
0.4
Survey Elevation (m)
Semivariance
120
0.3
0.2
OK iterative-addition method
n = 3859
r = 1.00
RMSE = 2.29 m
100
80
60
0.1
Model variogram
Empirical variogram
40
0.0
0
200
400
600
800
1000
40
Lag Distance (m)
60
80
100
120
140
DTM Elevation (m)
Figure 2.5. a) Empirical and modeled variograms from elevation points retrieved
from the iterative-addition algorithm (20-m starting scale). b) Relationship between
lidar-estimated ground elevation (sampled from the final DTM [Fig. 2.2b]) and fieldsurveyed elevation.
65
5
RMSE
Mean-signed Error
Observed Error (m)
4
187
3
323
138
284
2
1
0
0-3
3-10
10-20
20 +
Slope Class (Degrees)
Figure 2.6. Root-mean-square error (RMSE) and mean-signed error distribution by
slope class within old-growth forest. Numbers above the bars indicate the number of
reference points for each category. Lines connecting classes indicate groups of
homogenous mean absolute error (α = 0.05; t-tests account for heteroscedasticity
[Gotway & Cressie, 1990]).
66
RMSE
Mean-signed Error
Observed Error (m)
3
2
1
0
191
53
196
171
130
204
1115
-1
De
vel
Ab
S
Se
Pla
Sw
Old
c
and elec
tive onda ntatio amp F -grow
on
ed
r
n
e
l
th
ore
y
ys
dP
Are
st
ast logge Fores
as
ure
t
d
op
Figure 2.7. Root-mean-square error (RMSE) and mean-signed error distribution by
land-use class for all samples with slopes ≤ 10 degrees. Numbers below the bars
indicate the number of reference points for each category. Lines connecting classes
indicate groups of homogenous mean absolute error (α = 0.05; t-tests account for
heteroscedasticity [Gotway & Cressie, 1990]).
67
Vegetation Height (m)
Plantations
60
Abandoned
Pasture
0
Secondary
Forest
Old-growth
Forest
Swamp
Forest
±
0
0.5
1
Km
Figure 2.8. Landscape-scale overview of the final digital canopy model.
68
A
B
60
60
Old-growth Emergent Trees
n = 59
2
r = 0.51
RMSEm = 4.15 m
50
Tree Height (m)
Tree Height (m)
50
40
30
Old-growth Emergent Trees
n = 59
2
r = 0.50
RMSEm = 4.19 m
40
30
y = 8.64 + 0.85 x
y = 8.87 + 0.85 x
20
20
20
30
40
50
60
20
40
50
Mean95 Lidar Height (m)
C
D
60
60
60
Isolated Pasture Trees
n = 21
2
r = 0.95
RMSEm = 2.41 m
50
Tree Height (m)
50
Tree Height (m)
30
Maximum Lidar Height (m)
40
30
Isolated Pasture Trees
n = 21
2
r = 0.95
RMSEm = 2.48 m
40
30
y = 4.15 + 0.91 x
y = 4.79 + 0.91 x
20
20
20
30
40
50
60
20
Maximum Lidar Height (m)
30
40
50
60
Mean95 Lidar Height (m)
Figure 2.9. Individual-tree height regression models for old-growth emergent trees (a
& b) and isolated pasture trees (c & d). Reference data are the maximum height
within a tree’s crown. Lidar metrics include Maximum (a & c) or Mean95 (b & d).
Solid lines are the fitted models and dotted lines are the 1:1-relationships. RMSEm
is the root-mean-square error of the model (i.e., cross-validation prediction
residuals).
69
A
B
20
20
Tree stems only
n = 32
2
r = 0.96
RMSEm = 1.35 m
15
Plot Mean Height (m)
Plot Mean Height (m)
15
Tree stems only
n = 32
2
r = 0.97
RMSEm = 1.08 m
10
5
10
5
y = -0.41 + 0.87 x
y = -0.07 + 1.06 x
0
0
0
5
10
15
20
0
10
15
Lidar MeanALL Height (m)
C
D
20
20
20
All stems
n = 32
2
r = 0.87
RMSEm = 2.02 m
All stems
n = 32
2
r = 0.84
RMSEm = 2.26 m
15
Plot Mean Height (m)
15
Plot Mean Height (m)
5
Lidar Mean2x2 Height (m)
10
5
10
5
y = -0.06 + 0.7 x
y = 0.38 + 0.84 x
0
0
0
5
10
15
20
0
Lidar Mean2x2 Height (m)
5
10
15
20
Lidar MeanALL Height (m)
Figure 2.10. Plot mean height regression models. Reference data are the plot mean
of stem heights calculated from only tree stems (a & b; filled circles) or from trees,
sub-canopy palms and shrub stems (c & d; open circles). Plot-scale lidar metrics
include Mean2x2 (a & c) or MeanALL (b & d). Solid lines are the fitted models and
dotted lines are the 1:1-relationships. RMSEm is the root-mean-square error of the
model (i.e., cross-validation prediction residuals).
70
CHAPTER 3: Discrimination of tree species at multiple scales
3.1. Introduction
3.1.1. Hyperspectral discrimination of tropical tree species
Hand-held, airborne and spaceborne hyperspectral optical sensors measure
spectral information in over 100 narrow bands spanning the visible (VIS=437-700
nm), near-infrared (NIR=700-1327 nm), and two shortwave-infrared (SWIR1=14671771 nm; SWIR2=1994-2435 nm) regions of the electromagnetic spectrum (region
ranges adapted from Asner, 1998). It is anticipated that the automated classification
of tropical species may be possible with hyperspectral imagery that is both fine
enough to resolve individual tree crown (ITC) objects and also measures pertinent
discriminatory spectral features from 400 to 2500 nm (Cochrane, 2000); however,
this hypothesis has remained untested with an airborne or spaceborne hyperspectral
sensor. In this chapter, field spectrometer and high spatial resolution hyperspectral
data offer an unprecedented opportunity to explore the spatial-scale dependency of
spectral reflectance in the remote identification of tree species. Below I briefly
discuss important factors that influence plant reflectance at various spatial scales and
that may affect the automatic discrimination of tree species.
3.1.2. Reflectance properties of vegetation at leaf, branch and crown scales
Leaf-scale reflectance spectra are controlled by 1) leaf biochemical properties
(e.g., water, photosynthetic pigments, structural carbohydrates), which create
wavelength-specific absorption features, and 2) leaf morphology (e.g., cell-wall
71
thickness, air spaces, cuticle wax), which affects photon scattering (Asner, 1998,
Grant, 1987; Roberts et al., 2004; Woolley, 1971). VIS spectral variability among
species is low due to strong absorption by chlorophyll (Cochrane 2000; Poorter et al,
1995). High NIR transmittance and reflectance result from photon scattering within
leaf air-cell wall interfaces, such as in spongy mesophyll (Gausman, 1985; Grant,
1987; Woolley, 1971). In SWIR1 and SWIR2, water absorption tends to obscure
other absorption features produced by biochemical constituents (e.g., lignin and
cellulose) (Asner, 1998; Gausman, 1985).
Branch-scale spectra, such as from a high resolution pixel (e.g., < 4 m) or
measured in situ with a hand-held spectrometer, are a mixture of radiance
determined by the proportion, physical arrangement, and reflective and transmittive
properties of crown tissues, including leaves, stems, branches, fruits, and flowers.
Photon multiple-scattering among these components will tend to increase the
expression of leaf biochemical absorption features, especially within crowns with
large, densely-distributed and/or horizontally-oriented leaves (Asner, 1998). Finescale shadows cast within the branch may depress overall reflectance. Relative to
leaf scales, these factors are known to increase branch-scale spectral variability and
enhance separability of northern-latitude conifer and broadleaf trees (Roberts et al.,
2004). Fung and colleagues (1998) used laboratory-derived, branch-scale
hyperspectral data (400 to 900 nm, 90 evenly-spaced bands) and a linear
discriminant classifier to discriminate 12 subtropical tree species. An overall
accuracy of 84% was achieved and individual species Producer’s accuracies ranged
from 56 to 100%. Species spectral separability was attributed to the effect of leaf72
size variation expressed at the branch scale. Gong et al. (1997) found that a neural
network classifier applied to sunlit first-derivative spectra (6-8 cm spatial resolution,
in situ) could classify 6 conifer species with an average overall accuracy of 91%.
At the crown scale, the three-dimensional architectural arrangement of foliage
and non-photosynthetic components determines the amount of photon volumetricscattering and attenuation within the crown (Asner, 1998). van Aardt and Wynne
(2001) have shown that the VIS, NIR and SWIR1 regions are useful for
discriminating species of temperate forest conifer and hardwood species when using
in situ crown-scale hyperspectral data (sunlit sides of crowns). Spectral derivatives
provided the best overall classification accuracies, which were 84% for conifer
species and 93% for hardwood species. Cochrane (2000) provides the only
investigation of TRF crown-scale hyperspectral data for automated species
recognition (350-1050 nm data). The study used laboratory spectra from 11 tree
species to simulate branch and crown scales. Target species discrimination was
possible at crown scales, while it deteriorated at branch and leaf scales. Crown-scale
spectra were best separated in the VIS-NIR transition (i.e., the “red edge”) and NIR
regions. However, because the analysis used simulated branch and crown scale
spectra, it is not known how non-photosynthetic vegetation or volumetric crown
scattering will affect tree species spectral separability.
3.1.3. Challenges to tree species discrimination in tropical rain forests
Tropical rain forests pose challenging obstacles to ITC classification. TRF tree
communities are characterized by high diversity and relative rarity of individuals, so
73
large image extents are needed to find representative training samples. Many trees
occur below a dense overstory canopy, preventing their detection by a passive
optical sensor. In lowland tropical forests, a relatively constant growing season
fosters a diversity of phenological traits, and leaf flush and flowering may follow
annual or irregular cycles with no overriding community-scale patterns (Newstrom
et al., 1994). Therefore, strategically-timed over-flights to capture spectrallyimportant phenological events (sensu Key et al., 2001) may be done for only a few
tree species that have well-characterized phenology. Moreover, leaf-turnover and
flower display may be asynchronous among and within individual crowns of the
same species, thereby increasing conspecific variability in leaf- to crown-scale
spectra. For example, long-lived leaves within a crown may be covered with
epiphylls, which combined with leaf necrosis decrease VIS and increase NIR
reflectance (Roberts, Nelson et al., 1998). Depending on the density of leaves in a
crown, which may vary in time, radiance from understory shrubs, sub-canopy trees,
lianas, bark lichens, canopy soil, and epiphytes may mix with a target species
radiance and increase conspecific spectral variability. It is not yet clear whether
these spectral components will increase branch- and crown-scale within-species
variation to a level that inhibits among-species spectral discrimination (Castro-Esau
et al., 2004; Cochrane, 2000).
3.1.4. Objectives
In this chapter, I examine the relative trade-offs between spectral features, spatial
scale of measurement, and classification schemes for the automated classification of
74
individual TRF tree species using their reflectance properties. Field spectrometer
and airborne hyperspectral reflectance spectra (161 bands, 437-2434 nm) were
acquired from seven species of emergent trees in a lowland tropical rain forest,
permitting analyses at leaf, pixel and crown scales. My main objectives were to:
•
Determine if spectral variation among TRF tree species (interspecific) is
greater than spectral variation within species (intraspecific), thereby
permitting spectral-based species discrimination.
•
Identify the spatial scale and spectral regions that provide optimal
discrimination among TRF emergent tree species.
•
Develop an analytical procedure for the species-level (floristic)
classification of individual tree crowns using their reflectance spectra.
•
Assess the relative importance of narrowband hyperspectral versus
broadband multispectral information for species identification of TRF
trees.
3.2. Methods
3.2.1. Data sets and pre-processing
3.2.1.1. Canopy-emergent trees
To select my study species, I conducted field surveys and took advantage of a
Geographic Information System (GIS) database of tree locations from a long-term
tree demography study at LSBS (the TREES project; Clark et al., 1998). My
preliminary analysis involved 544 individual trees belonging to 27 species and led
75
me to focus my efforts on seven species (Table 3.1) of canopy emergents for which
there were sufficient individuals in the hyperspectral imagery for a representative
sample. Emergent trees with large, exposed crowns provided a large sample of
pixels that were less influenced by spectral shadowing or scattering by neighboring
trees, and they were easy to locate in the orthorectified hyperspectral imagery.
Furthermore, five of the seven study species (BAEL, DIPA, HYME, HYAL and
LEAM) are under analysis in the TREES project, providing opportunities to
generalize local-scale research (e.g., demographic changes) to broader spatial scales
using remote sensing.
A total of 214 individuals of the seven study species were identified in the
hyperspectral imagery through field surveys conducted between January 2000 and
July 2001 (Table 3.2; Fig. 3.1). The trunk coordinates in the LSBS grid system (see
Chapter 2) were surveyed by measuring the distance and angle of the trunk from the
nearest grid tube. These trunk coordinates were then converted from LSBS grid
coordinates to the UTM projection, WGS-84 datum coordinate system using a leastsquares affine transformation with RMSE of 4.8 m (OTS, unpublished data).
There is little long-term data on leaf and flowering phenology of the study
species. Some overstory tree species are deciduous and completely drop and flush
leaves, generally beginning in the first dry season, while others are evergreen and
continuously flush small amounts of leaves throughout the year (Table 3.1).
Hyperspectral imagery was acquired on March 30, 1998, at the end of the first dry
season, and all study trees were expected to have high mature leaf cover (i.e., high
leaf area index) except DIPA and LEAM (personal observation, April 12, 2004;
76
literature data [Frankie et al., 1974; O’Brien, 2001]; summarized in Table 3.1).
Although I do not have field observations from my study individuals during the
image acquisition, O’Brien (2001) estimated leaf cover of BAEL, DIPA, HYME and
LEAM individuals at LSBS that were 30-60 cm diameter above buttress and
unobstructed or emergent crowns. Data included March through April, 1998 and
showed that a relatively large proportion of DIPA and LEAM individuals had low
mature leaf cover, while BAEL and HYME individuals had higher mature leaf
cover. An example Balizia crown is shown in Figure 3.2a.
3.2.1.2. Leaf-scale spectra
A shotgun was used to shoot 152 leaf samples down from crowns of individual
study trees in August, 2002 (Table 3.3). Three to five individual trees from each of
the seven study species were selected for sampling, and 2-10 leaves per individual
were shot down from the upper, sun-exposed part of the crown. Leaf samples
included a range of maturity and health. Individual leaflets > 1 cm width were
sampled from separate leaves for species with compound leaves (CEPE, DIPA,
HYME) and leaflets were analyzed as leaves. For scale considerations discussed
below, BAEL compound leaves were analyzed rather than individual leaflets.
Bidirectional reflectance properties of the “leaves” (i.e., leaves or leaflets) were
measured in a darkroom at LSBS. Leaf samples were put in a plastic bag with a
moist paper towel and stored in a cooler with ice until refrigerated in the laboratory.
All samples were measured within 12 hours of collection. A single 150-W halogen
lamp was placed with a 25-degree incident angle and 53 cm above a matte-black,
77
5%-reflective box. An ASD FieldSpec spectrometer (Analytical Spectral Devices,
Boulder, CO, USA) sensor with an 8° fore-optic was positioned 7.1 cm at nadir
above the box center, yielding a 1-cm sensor field of view (FOV). The spectrometer
was optimized with a white Spectralon® panel (Labsphere, North Sutton, NH, USA)
placed in the box center, and the instrument was re-optimized using the panel after
measuring every 5 to 7 leaf samples. Leaf samples were placed in the box center
with adaxial (upper) surfaces to the sensor and radiance was measured 5 times per
leaf. Bidirectional reflectance of a single leaf sample will vary across its surface
depending on biochemical variation (e.g., chlorophyll concentration, leaf necrosis,
epiphyll cover), structural properties (e.g., cuticle texture, mesophyll depth), and
illumination and sensor geometry. To capture this potential spectral variation from
an individual leaf, leaf orientation and position relative to the sensor FOV were
varied with each of the 5 radiance measurements (i.e., the leaf was moved while the
sensor remained stationary). Radiance spectra from the Spectralon® panel were
used as a standard to convert leaf radiance to percent reflectance. The final leaf
reflectance spectrum was an average of the five reflectance spectra from each leaf.
Individual leaflets of BAEL (Balizia) leaves were smaller than the sensor FOV,
thus causing the black background to mix with the radiance signal. To counteract
this effect, each Balizia leaf was stacked on top of 3 other Balizia leaves to simulate
a dense layer of leaves. The leaf-stack position and orientation was haphazardly
varied with each radiance measurement.
78
3.2.1.3. Hyperspectral image pre-processing
Hyperspectral imagery came from the HYDICE sensor, which is described in
Chapter 1. HYDICE runs were delivered as 16-bit calibrated radiance data. The
morning data acquisition avoided afternoon cloud cover yet the 56.3° to 48.4° solar
zenith angles (92° to 94° azimuth angles) during the flight caused deep tree shadows
that are particularly noticeable in old-growth forest canopy gaps (Fig. 3.2b).
Individual tree crowns are clearly resolved in this high spatial and spectral resolution
imagery (Fig. 3.2b).
I orthorectified the LSBS sections of HYDICE runs 6, 9, 12 and 15 using the
Erdas IMAGINE OrthoBASE software package (Leica Geosystems GIS & Mapping,
LCC, Atlanta, GA, USA). Runs were segmented into 800-m long blocks and each
block was orthorectified with 21 to 75 ground control points collected by visually
matching emergent tree crown centers in HYDICE imagery to co-located crown
centers in a 0.3-m lidar digital canopy model (DCM; Chapter 2). Terrain distortions
in the imagery were corrected in the orthorectification processing with a 10-m
resolution digital terrain model (DTM: OTS, unpublished data), which was
originally derived from Laser Vegetation Imaging Sensor (LVIS) lidar data
(Rocchio, 2000). Orthorectification used a nearest-neighbor interpolator and georegistered the HYDICE imagery in the Universal Transverse Mercator (UTM),
WGS-84 datum projection of the DCM and DTM reference data (Fig. 3.1 shows
spatial extent of runs).
The ACORN v4.0 (Analytical Imaging and Geophysics LLC, Boulder,
Colorado) atmospheric correction package was used for calibrating radiance values
79
to surface reflectance. Although atmospheric water vapor can be calculated on a
per-pixel basis (Gao and Goetz, 1990), low signal-to-noise in principal water
absorption bands for HYDICE (Basedow et al., 1995) produced considerable spatial
error in water vapor estimates; and therefore, atmospheric corrections were
performed with a fixed precipitable water vapor of 32 mm. A tropical atmospheric
model was used with atmospheric visibility of 100 km. Water vapor and visibility
parameters were established based on visual assessment of old-growth tree spectra
and an empirical, minimum root-mean square error (RMSE) comparison with fieldcollected spectra.
Field spectra were measured in August, 2002 with an ASD
FieldSpec spectrometer and included gravel road, cement, tile, exposed soil, wood
planks, green metal roof-tops, and mowed lawn targets that were located within the
HYDICE runs.
An 8° fore-optic was positioned about 1.5 m above a target and
sensor radiance was converted to reflectance using an in situ white Spectralon®
calibration panel. Five individual reflectance measurements were averaged to create
a target spectrum over a 1 m2 area, and then several of these spectra were collected
over a homogenous area of the target and then averaged.
Wavelength calibration differed among runs by 0.61 to 2.67 nm per wavelength
(HYDICE metadata, SITAC). Atmospheric correction was performed using each
run’s respective band centers and full-width half-maximum (FWHM) parameters. A
common set of band center wavelengths were calculated by averaging bands centers
from the four runs, and reflectance values from each run were then linearlyinterpolated to this common set of center wavelengths. Band centers were spaced an
average distance of 6 nm in VIS, 14 nm in NIR, 12 nm in the SWIR1 and 9 nm in
80
SWIR2. Post-calibration reflectance artifacts (e.g., spikes near water absorption
features) were minimized with a 3-channel box-car filter. Bands with extreme noise
in spectral regions less than 437 nm and greater than 2435 nm, as well as bands in
the strong water absorption features 1313-1466 nm and 1771-1994 nm, were
removed from analyses.
An example comparison between final HYDICE reflectance and field-measured
ASD reflectance for a wooden-plank suspension bridge and a nearby tree crown
(Pentaclethra macrophylla) is shown in Fig. 3.3. The HYDICE reflectance spectra
were generally the same shape, but reflectance was much lower than field
spectrometer measurements in the SWIR1 and SWIR2 regions. This pattern was
observed in comparing HYDICE with other field spectra. NIR reflectance peaks
were high and water absorption features centered at 980 nm and 1200 nm were deep
relative to field spectra, especially for the wooden bridge.
These artifacts in
HYDICE derive from a combination of poor radiometric calibration, sensor noise,
atmospheric noise (e.g., water vapor absorption) and the difference in time between
HYDICE and field measurements. Field measurements were taken on August 2,
2002 (9:30 am) while HYDICE was acquired on March 30, 1998 (8 am).
If
atmospheric conditions were constant over the reserve, all HYDICE artifacts should
be common to all HYDICE spectra because the same parameters were used to
convert each pixel to reflectance. The mismatch between HYDICE reflectance and
expected reflectance affects my analysis in three ways: 1) in comparing HYDICE
spectra to laboratory leaf spectra, 2) possibly shifting band selection towards bands
81
with higher signal to noise, and 3) limiting the comparison of my results with other
sensors.
3.2.1.4. Simulated broadband, multispectral imagery
HYDICE reflectance spectra were convolved using sensor-specific spectral
response functions to simulate IKONOS, Landsat ETM+, and ASTER (Advanced
Spaceborne Thermal Emission and Reflection Radiometer) imagery. Although each
of these sensors has a different spatial resolution (i.e., 4 m IKONOS, 30 m ETM+,
15-30 m ASTER), the spatial resolution of simulated imagery was fixed at the 1.6 m
of HYDICE imagery. Also, by using simulated imagery, the same artifacts evident
in HYDICE were incorporated into simulated spectra. Therefore using simulated
multispectral imagery reduced the effects of spatial scale and radiometric artifacts on
inter-sensor comparisons.
3.2.1.5. Pixel- and crown-scale spectra from individual tree crowns
Field-surveyed trunk locations were overlaid on the orthorectified HYDICE
mosaic and the polygons representing the 2-dimensional area of the tree crowns were
manually digitized over the imagery. I used the DCM as a visual aid to determine a
crown’s shape in areas with shaded pixels. The average crown area for the study
species was 444 m2, with each crown comprising from 41-662 pixels (Table 3.2).
Hereafter I refer to the digitized crown polygons as individual tree crowns (ITCs:
Gougeon & Leckie, 2003).
82
My analyses differentiated between all (PixelALL) and sunlit-only (PixelSUN)
pixels within each ITC. Sunlit pixel spectra were designated as all pixels within an
ITC that had reflectance greater than or equal to the crown’s mean 800-nm (NIR)
reflectance (Gougeon, 1995). Crown-scale spectra were calculated by averaging
either all (CrownALL) or sunlit-only (CrownSUN) pixel spectra within each ITC.
3.2.2. Data analysis
3.2.2.1. Testing of within and among species spectral variability
Spectral separability of species should be optimal if different species have high
statistical distance in feature space and within-species variation is less than amongspecies variation. I tested the null hypothesis that within- and among-species
spectral variation are equal with a non-parametric multivariate analysis of variance
technique (NPMANOVA) first developed for use with ecological distance matrices
(Anderson, 2001; McArdle & Anderson, 2001). In remote sensing applications, the
spectral angle is a metric used for comparing the degree of similarity between two
spectra (Kruse et al., 1993). Unlike Euclidean distance, the spectral angle is
insensitive to linearly-scaled differences among spectra such as those caused by
illumination. In my implementation of the NPMANOVA, the distance between each
spectrum to every other within- and among-species spectrum was calculated using
the spectral angle and Euclidean distance, and distances were stored in N x N
distance matrices (N = number of observations). In the calculation of spectral
distance, spectra were analyzed using the entire 161-bands or limited to the VIS,
NIR, SWIR1, SWIR2 regions. A pseudo-F statistic was calculated as the ratio of
83
among to within species sums of squares (Anderson, 2001; McArdle & Anderson,
2001). The null hypothesis tested was that within and among species spectral
variation was equal, which would make the F-ratio close to one. The significance of
the F-ratio was tested against a null distribution of F created by 5000 random
permutations of the distance matrix (Anderson, 2001). For pixel-scale
NPMANOVA tests, 200 pixels for each species were randomly selected from
crowns for the respective species. For leaf- and crown-scale NPMANOVA tests, I
used all available spectra due to limited sample sizes. I performed NPMANOVA
tests using the DISTLM2 v.5 software program (Anderson, 2004).
3.2.2.2. Species classification schemes
I explored three popular supervised classification schemes for TRF tree
classification: spectral angle mapper (SAM), linear discriminant analysis (LDA), and
the maximum likelihood (ML) classifier. SAM is a spectral matching technique
(Kruse et al., 1993). The spectral angles between each sample spectrum and several
reference spectra are calculated to reduce the hyperspectral data cube from an ndimensional spectral space to a similarity space with dimensions equal to the number
of reference spectra (i.e., classes).
SAM classification was accomplished by
assigning each sample spectrum to the class with the closest similarity (i.e., lowest
spectral angle), and no maximum-angle threshold was used to minimize false
detections.
LDA is a common classifier that has been used in previous ITC
classification research (Gong et al., 1997; Fung et al., 1998; van Aardt & Wynne,
2001).
For LDA classification, the pooled within-class covariance matrix and
84
predictor variables (e.g., reflectance values) from training samples are used to build
classification equations, or discriminant functions for each class (Duda & Hart,
1973; Tabachnick & Fidell, 1989).
A class is chosen based on the highest a
posteriori probability calculated from the functions. The most important assumption
of LDA classification is that all classes share the same covariance matrix (i.e.,
homogeneity).
In the ML classifier, each class mean, standard deviation and
covariance matrix are estimated from the training data to evaluate a sample’s class a
posteriori membership probability (Duda & Hart, 1973). ML has been widely used
in ITC species classification (Gougeon, 1995; Key et al., 2001; Leckie, Gougeon,
Hill et al., 2003; Meyer et al., 1996).
Supervised classification schemes are often stymied by the large dimensionality
of hyperspectral imagery. Fine resolution spectral bands are often correlated and so
represent redundant information. Also, sensor noise such as stripes from bad
detectors or atmospheric attenuation may be greater in certain bands and this noise
may increase class variance and decrease class separability. With the ML classifier
in particular, it has been shown that the within-class covariance matrix can be poorly
estimated when there are few training samples relative to the data dimensionality,
leading to a decrease in classifier performance called the Hughes phenomenon
(Duda & Hart, 1973; Jackson & Landgrebe, 2001). Hence, the use of ML is limited
for hyperspectral remote sensing of forested areas because image dimensionality is
high while training data are expensive or difficult to acquire. A common solution to
this dilemma is to reduce data dimensionality through spectral feature (i.e., band)
selection. In this chapter, I used a forward stepwise selection method based on
85
discriminant analysis (Tabachnick & Fidell, 1989; van Aardt & Wynne, 2001). This
method was implemented using the SAS STEPDISC procedure (SAS Institute Inc.,
Cary, NC, USA) with the significance criteria set at α = 0.05 for all analyses except
the crown scale, which had criteria set to α =0.20.
Following feature selection, SAM, LDA and ML classifiers were applied to leaf-,
pixel- and crown-scale spectra to assess how the spatial scale of spectral
measurements affects species classification accuracy. At each scale, the ndimensional spectral space was varied to include the full-spectra dataset (161 bands),
spectral regions (i.e., VIS, NIR), or LDA stepwise-selected bands. Spectral regions
were sub-sampled to include only 10 bands per region. These bands were evenlyspaced with an average spacing of 23 nm (VIS), 55 nm (NIR), 25 nm (SWIR1), and
47 nm (SWIR2). The same set of classifiers was also applied to the simulated
broadband multispectral data. All LDA classification was accomplished using the
“MASS” package in the R statistical environment (R Development Core Team,
2004; MASS 7.2-12, R v2.0) while SAM and ML classification was performed in
ENVI v4.1 and IDL v6.1 (RSI, Inc., Boulder, CO, USA).
3.2.2.3. Crown-scale and “pixel-majority” ITC classification
A major objective of this chapter was to assess optical remote sensing for
operational, ITC species discrimination. Current research shows that tree species
discrimination is best accomplished by aggregating pixels into their respective
crowns for object-based (as opposed to pixel-based) classification using spectral and
spatial properties (Gougeon, 1995; Leckie, Gougeon, Hill et al., 2003; Meyer et al.,
86
1996). In this chapter, the species of ITCs in the HYDICE imagery were determined
by 1) the class assigned from crown-scale spectra, CrownALL or CrownSUN and, 2)
taking the majority value of the classified pixels (PixelALL or PixelSUN) within each
ITC, i.e., the “winner-takes-all” rule (Meyer et al., 1996). I refer to this latter
approach (2) as “pixel-majority” ITC classification.
3.2.2.4. Classifier training and accuracy assessment
For pixel-scale classifications, 300 randomly-selected training pixels were
sampled from the crown objects for each species (300 pixels x 7 species = 2100
training pixels). Pixels were sampled from the whole crown (PixelALL) or from
sunlit regions of the crown (PixelSUN), and each crown was sampled unless it had
fewer than 40 pixels. Each classifier (LDA, ML or SAM) was applied to the
remaining non-training pixels within each crown. I sampled 300 training pixels per
species to provide a robust estimation of ML class covariance statistics. For pixelscale testing, 300 non-training classified pixels per species were randomly selected
(300 pixels x 7 species = 2100 test pixels). A new set of 300 test pixels were
randomly selected for each classifier-band combination analyzed.
For crown-scale and pixel-majority analyses, classification training and testing
were performed with cross-validation due to the limited number of ITCs (Duda &
Hart, 1973; Krzanowski, 2001).
I sequentially left one ITC out and trained
classifiers with pixel- or crown-scale spectra from the remaining 213 ITCs (i.e.,
“leave-one-out” cross validation). For pixel-majority classification, each withheld
crown was classified based on the majority-class rule. Cross validation provides a
87
slightly biased estimate of true classifier accuracy (Krzanowski, 2001). Statistical
differences among classifications were tested with the Z statistic calculated from the
Kappa statistic and variance (Congalton, 1991). Leaf-scale classification was also
performed with a similar cross-validation procedure.
3.3. Results
3.3.1. Reflectance properties at different scales
3.3.1.1. Leaf-scale spectra
Leaf-scale spectra for the seven tree species (Fig. 3.4) showed typical patterns of
vegetation: low VIS reflectance caused by absorption by chlorophyll and other
pigments, high NIR reflectance due to multiple-scattering within the leaf structure,
weak NIR water absorption features at 980 and 1200 nm, and moderate reflectance
in SWIR1 and SWIR2 with peaks at 1650 and 2200 nm caused by dominant water
absorption features at 1400, 1900 and 2700 nm (Gausman, 1985; Roberts et al.,
2004).
There was considerable variation in reflectance within species, especially in the
NIR and SWIR (Fig. 3.4). Several factors can cause leaf spectral variation within a
given species, including epiphyll cover, herbivory, necrosis, maturation of the
mesophyll, and the concentration of chlorophyll and water. Seven percent of all
upper-canopy leaves sampled had epiphyll coverage. As seen for HYME leaves of
roughly the same age (Fig. 3.5a), epiphyll coverage tends to lower the green peak
and NIR reflectance, possibly due to more absorption of light by epiphylls covering
the adaxial leaf surface. Herbivory is another factor that affected 4% of the leaves
88
sampled. For LEAM leaves of the same age, some leaves had light brown-colored
leaf mines caused by an insect. As the percentage of these mines increased, there
was less VIS absorption (higher reflectance) likely due to lower amounts of
photosynthetic pigments from eaten leaf material, and increased NIR and SWIR
reflectance (Fig. 3.5b) likely due to lower water content and more exposed nonphotosynthetic leaf material (e.g., residual, dried leaf veins). Finally, leaf age is an
important factor because it determines time exposed to epiphylls and herbivory, as
well as internal leaf architecture and chemical properties. For TEOB, young thin
leaves had high red and low NIR, SWIR1 and SWIR2 reflectance relative to mature,
thicker leaves (Fig. 3.5c). Thin leaves are compact and have fewer air-cell wall
refractive discontinuities than mature leaves, causing lower NIR-SWIR reflectance
(Gausman, 1985). Also, lower chlorophyll content in young leaves likely accounts
for higher VIS reflectance (i.e., lower VIS absorption) in the blue (450 nm) and red
(680 nm) regions (Woolley, 1971; Gausman, 1985). As the leaf senesces, lower
concentrations of chlorophyll greatly reduce the amount of absorption throughout
the VIS, thereby increasing reflectance (Fig. 3.5c).
3.3.1.2. Pixel-scale spectra
Pixel-scale spectra revealed the same general patterns of NIR scattering and
chlorophyll and water absorption as seen in leaf-scale spectra (Fig. 3.6). Relative to
leaf-scale spectra, there was an overall reduction (darkening) of percent reflectance
in PixelALL and PixelSUN spectra (Table 3.4). Darkening of spectra is partly due to
fine-scale shadows within branches of leaves and other crown materials, especially
89
in the PixelALL samples. However, some of the darkening in the SWIR1 and SWIR2
regions of pixel spectra was due to poor HYDICE radiometric calibration. Roberts
et al. (2004) found that biochemical absorption properties of leaves were accentuated
at the pixel scale by the multiple-scattering of photons among leaves and other
crown tissues within the hyperspectral sensor’s FOV. HYDICE spectra show
evidence of this phenomenon; photon scattering can partially explain why NIR water
absorption features at 980 nm and 1200 nm were deeper in pixels relative to leafscale samples (Fig. 3.6).
The arrangement and density of crown tissues will govern the crown scattering
environment and the degree to which leaf biochemical properties are accentuated at
pixel or crown scales. Considering sunlit pixels (PixelSUN), DIPA and LEAM had
40.2% and 37.9% NIR reflectance, respectively, while the other species had between
43.8% (CEPE) to 51.5% (TEOB) mean reflectance (Fig. 3.6). Relatively low NIR
reflectance makes DIPA and LEAM appear purple in the image (Fig. 3.2b, RGB
1651 nm: 835 nm: 661 nm). Individuals of these deciduous species had low crown
foliage density (i.e., leaf area index, LAI) during HYDICE image acquisition in the
late dry season. With fewer leaves in the crown, there was less photon scattering
and subsequently, NIR reflectance was low relative to leaf-on species. The variation
in scattering environments among species with different crown LAI, explains why
the NIR standard deviation was much higher in pixel scales than in leaf scales (Table
3.4).
90
3.3.1.3. Crown-scale spectra
Crown scale spectra were an average of pixel spectra. The averaging of spectra
decreased CrownALL and CrownSUN variance in all spectral regions relative to
PixelALL and PixelSUN variance, respectively (Table 3.4, Fig. 3.7). As observed
with pixel scales, the mixture of pixels from bright and dark, shadowed parts of the
crown tended to lower average reflectance for CrownALL spectra relative to
CrownSUN spectra.
3.3.2. Among- and within-species spectral distances
Using spectral angle as a distance metric, among-species (interspecific) spectral
variability was significantly greater (p≤0.001) than within-species (conspecific)
variability for all spectral regions at leaf and pixel scales (Table 3.5). However,
species differences at the crown scale were mainly focused in the NIR region. Using
Euclidean distance, which includes variation due to illumination, among-species
variability was significantly greater (P≤0.001) than within-species variability at all
scales and all spectral regions. The greater separation of species with Euclidean
distance over spectral angle distance at crown scales indicates that crown-level
illumination differences among species, possibly due to varying crown LAI, tend to
increase species separability. There was no advantage to using just sunlit pixels at
crown scales (CrownALL vs. CrownSUN). However, at pixel scales sunlit samples had
greater separability (i.e., higher F-ratios) than when considering all samples. The
NIR, SWIR1 and SWIR2 regions of sunlit pixels had particularly high F-ratios.
91
3.3.3. Selected spectral features
3.3.3.1. Leaf-scale
For discriminating species, 90% of the 10 most important wavelengths selected
by the stepwise procedure were concentrated in the NIR and SWIR1 regions (Fig.
3.8a; Fig. 3.9a), where there was relatively large variation in percent reflectance
(Table 3.4). Balizia leaves had the highest NIR variability (6.7% s.d. for BAEL, 3.3
to 5.5% s.d. for other species). BAEL variability is likely caused by measuring their
reflectance from leaf stacks. Photons have more opportunity for scattering within
stacks of leaves, thereby increasing the NIR plateau, broadening water absorption
features, and increasing overall NIR variability. SWIR2 had lower variability than
NIR or SWIR1 yet SWIR2 comprised 25% of the bands selected when considering
20 bands (Fig. 3.9b). The VIS had the lowest spectral variability, and only 10% of
the bands selected were from VIS when considering 10 or 20 bands. The selected
VIS bands were in the blue absorption feature at 449 nm, the green peak at 568 nm,
and the red-edge at 719 nm (Fig. 3.8a).
3.3.3.2. Pixel-scale
Leaf-scale band selection is best compared to the PixelSUN samples because they
have a similar range of illumination. Of the 20 most important bands, there were 4
more NIR and 4 fewer SWIR1 bands selected for PixelSUN relative to leaf-scale
spectra (Fig. 3.8a-b). In terms of the percentage of bands selected (Fig. 3.9a & b)
and NPMANOVA F-ratios (Table 3.5), the NIR region was particularly useful in
pixel-scale species discrimination, especially with sunlit samples. Selected NIR
92
bands were concentrated in the red-edge and the plateaus surrounding the water
absorption features at 980 and 1200 nm (Fig. 3.8b), while less so at leaf scales (Fig.
3.8a). There was high NIR variability relative to other regions (Table 3.4), which is
likely caused by species differences in crown LAI. Low-LAI deciduous species
(e.g., DIPA, LEAM) had lower NIR reflectance relative to high-LAI species (e.g.,
HYAL, TEOB). These differences in crown architecture among species thus create
distinctive variation in maximum NIR reflectance that permits clearer NIR species
discrimination with pixel spectra relative to leaf spectra, which are not influenced by
crown architecture. SWIR1 may be less useful at pixel scales relative to leaf scales
due to a combination of greater water absorption and lower sensor signal to noise.
When considering the best 20 discriminatory bands, NIR was less important
when all sunlit and shaded pixels were analyzed, likely due to less photon scattering
in shaded pixels (Fig. 3.8b; Fig. 3.9b). With PixelALL samples, there were a similar
number of bands in the VIS, NIR and SWIR1 regions. VIS bands were spread
across blue-green edge (491 nm), yellow edge (575 and 619 nm) and red well (670
nm) features, while SWIR1 bands were evenly spaced across the region.
3.3.3.3. Crown-scale
With the stepwise band selection procedure, only 42 and 41 bands were
significant at the α=0.05 level for CrownALL and CrownSUN spectra, respectively.
The significance criteria was substantially relaxed to α=0.2 in order to select 60
bands for comparison with leaf and pixel scales. The twenty most important bands
(all at α=0.05) were concentrated in the VIS, NIR and SWIR2 regions for CrownALL
93
spectra (Fig. 3.8c; Fig. 3.9b). For CrownSUN spectra, there were 2 fewer bands in
VIS and 2 more bands in SWIR2 relative to CrownALL spectra. As with the pixelscale, NIR bands were clustered in the red edge and on the peaks bordering the water
absorption features.
3.3.4. Classification of tree species at leaf, pixel and crown scales
3.3.4.1. Hyperspectral, narrowband classification
The overall accuracies of tree species classifications with different narrowband
combinations and classification schemes are presented in Table 3.6. Leaf-scale
maximum likelihood (ML) analysis was limited to 10 bands due to the low number
of training samples per class (Table 3.3). Linear discriminant analysis (LDA) had
the highest accuracy (89.5%) for leaves when the analysis was isolated to the 10 best
bands, while ML with 10 bands was 8.6% less accurate (Z=6.66, p≤0.05). Balizia
was the only species that had no inter-class confusion with the LDA classifier and
10+ bands nor with LDA and NIR or SWIR1 bands, which can be attributed to its
distinct spectral leaf-stack properties. LDA accuracy was significantly higher with
20 bands (Z=3.15; p≤0.05) and overall accuracy reached 100% with 40 bands (no
significant difference between 20 and 30 to 60 bands).
Both band selection and NPMANOVA identified the SWIR1 region as important
to species separability at leaf scales, while the VIS region was not as important (Fig.
3.9, Table 3.5). In agreement with these findings, 10 SWIR1 bands provided the
highest classification accuracy of the spectral regions while 10 VIS bands produced
relatively low accuracy (Table 3.6, LDA or ML). Relative to LDA and ML
94
classifiers, the Spectral Angle Mapper (SAM) classifier had relatively poor
performance (< 51% overall accuracy) and differences among band combinations
will not be discussed.
At the pixel scale, ML generally had higher overall accuracy than LDA for most
band combinations, using all or sunlit-only pixels (Table 3.6). SAM had very low
performance, with no accuracy exceeding 50%. In the ML and LDA analyses,
PixelSUN classifications were significantly more accurate than when using PixelALL
samples (Table 3.6, arrows). Best pixel-scale performance was with 40-60 bands
from PixelSUN samples with either classifier (differences not significant). ML
classification with 161 bands (full-spectra) was significantly lower than when using
20 to 60 bands due to the Hughes phenomenon, while full spectra information did
not dramatically change LDA accuracy. Classification accuracies for both LDA and
ML were significantly higher with 10 bands from across the entire spectra relative to
selecting 10 bands from specific spectral regions.
The LDA classifier applied to crown-scale spectra produced some of the highest
species classification accuracies using airborne HYDICE imagery. For all band
combinations, LDA accuracies were not significantly greater with CrownALL versus
CrownSUN samples. LDA classification accuracy using 10 optimally-selected bands
was significantly greater than using all 161 bands or 10 evenly-spaced bands in the
VIS, NIR, SWIR1 and SWIR2 spectral regions.
The best accuracy was 92.1% when the LDA classifier was applied to 30
CrownALL stepwise-selected bands (Tables 3.6 & 3.7). However, 20 bands provided
only 2.8% less overall accuracy and were not significantly different than 30 bands.
95
Of the 30 bands used in the CrownALL classification, 30.0% were within the NIR
region while 23.3% were in each of the other 3 spectral regions. Forty-one percent
of misclassified ITCs involved confusion between DIPA and LEAM. This result
was expected because both species had similar reflectance properties due to low-LAI
crowns.
Although it was not possible to compare LDA and ML classifiers for data
dimensionalities > 10 bands due to limited training data, the cross-validation
accuracies between LDA and ML with 10-bands were significantly higher for
LDA—7.5% and 7.0% higher with CrownALL or CrownSUN, respectively. As with
other scales of analysis, SAM crown-scale classification accuracy was relatively
low. The highest SAM accuracy achieved was only 53.7%.
In all LDA and ML classification analyses, there were no minimum a posteriori
probability criteria set for assigning samples to a class; and therefore, there were no
unclassified samples. The criteria was omitted to permit a direct comparison of
LDA and ML with the SAM classifier. For crown-scale LDA classifications, I also
experimented with 50%, 75% and 90% probability thresholds, where a crown-scale
spectrum (i.e., ITC) was classified as “unknown” if its maximum class a posteriori
probability was lower than the specified threshold. As the probability threshold was
increased (i.e., made more conservative), more ITCs were labeled as unclassified
and the overall accuracy dropped (Fig. 3.10). The decrease in overall accuracy was
less severe as more bands were added to the analysis. With 30 bands, overall
classification accuracy was 8.9% significantly higher with no probability threshold
relative to a 90% threshold (Tables 3.7 & 3.8). A higher probability threshold acted
96
to increase omission errors by switching correctly-classified ITCs to “unknown”,
thereby decreasing the class Producer’s accuracies. In this example, 19 correctlyclassified ITCs did not meet the threshold criteria and so were left unclassified. On
the other hand, a more conservative threshold decreased commission errors, thereby
increasing class User’s accuracies. For example, the 90% threshold identified 3
ITCs that had been confused between DIPA and LEAM due to low a posteriori
probabilities. With the threshold set, these crowns were left unclassified and User’s
accuracy increased 2.5% and 11.3% for DIPA and LEAM, respectively (Tables 3.7
& 3.8).
In my classification analyses, only 7 of over 400 tree species were classified. In
an operational classification, a probability threshold or other technique will be
necessary to ensure that ITCs that are not target species remain unclassified. I tested
the crown-scale, threshold classifier against non-target ITCs. A total of 30 emergent
crowns comprising 14 non-target species were identified in the field and digitized
over the imagery, following methods used for the 7 study species. I then classified
these non-target ITCs with the 30-band, LDA classifier and a 90% a posteriori
probability threshold. The classifier was only trained with the 7 study species (214
ITCs). Five of the non-target ITCs were actually classified as “unknown,” whereas
17 were classified as HYAL, 6 as DIPA, 1 as BAEL and 1 as CEPE.
Misclassifications were not random. Crowns that were misclassified as DIPA had
very similar spectral properties as true DIPA crowns—appearing as purple ITCs in
false-color imagery (e.g., Fig. 3.2b)—suggesting that they had low-LAI crowns. In
contrast, ITCs misclassified as HYAL had spectral properties of the high-LAI
97
crowns typical of true HYAL crowns. The classifier thus appears most attuned to
differentiating ITC phenology and structure rather than clear species distinctions.
3.3.4.2. Multispectral, broadband classification
Classification accuracies from simulated broadband sensors (Table 3.9) followed
the same general patterns across scales as discussed for narrowband data: LDA and
ML classifiers outperformed SAM; there was an increase in accuracy with more
bands; and, LDA and ML accuracies were generally higher at crown scales relative
to pixel scales. As with narrowband analyses, the LDA classifier was particularly
strong at the crown scale. For CrownALL ASTER spectra (9 bands), overall accuracy
was 76.2%, which was 7.9% and 14.0% lower (both significant) than the accuracies
achieved with 10 and 30 narrow bands, respectively (Table 3.6). With CrownALL
LDA analyses, there was a non-significant 7% increase in overall accuracies with 6
ETM+ bands over 4 IKONOS bands, while there were 9.8% and 16.8% significant
increases with 9 ASTER bands over IKONOS and ETM+ bands, respectively.
3.3.5. Pixel-majority classification using within-crown pixels
3.3.5.1. Hyperspectral, narrowband classification
ITCs were next assigned a species label based on the majority class of pixelscale, narrowband spectra within each crown object (Table 3.6, pixel-majority).
Pixel-majority ITC classification accuracy was not significantly different using
PixelALL or PixelSUN within-crown spectra. The best overall pixel-majority
accuracies with each classifier-band combination were generally higher than those
98
achieved with comparable pixel-scale classifications, except for LDA with sunlit
pixels and 40, 60 or 161 bands (Table 3.6). The highest pixel-majority accuracy,
85.5%, was achieved with a LDA classifier applied to 30 bands selected from
PixelSUN spectra (Table 3.10), although accuracy was not significantly less with 10
and 20 band combinations, nor with PixelALL spectra.
For the 30-band LDA classification (PixelSUN), I assessed the percentage of
within-crown classified pixels that had the same species label as their corresponding
ITC. If a within-crown pixel and its corresponding ITC species labels were the
same, the pixel was considered “correctly classified”. If pixels were randomly
classified, then each crown would likely have only 14% (1 out of 7) of its pixels
correctly classified, and ITCs would be assigned to the class that had a pixel
majority by chance. I found that the mean percentage of correctly-classified pixels
within correctly-labeled ITCs was 89.9% (range 40.0% to 100.0%). Therefore,
although there were misclassified pixels within ITCs, I am confident that crowns
were not assigned a correct label by chance. For mislabeled ITCs, the mean of
correctly-classified pixels within each crown was 13.7% (range 0.0% to 47.8%)
This indicates that there was substantial spectral variation among individual crowns
of a single species—some crowns had very poor pixel classification accuracy, while
others had very high accuracy. For example, TEOB crowns were all correctly
labeled and individual crowns had an average 91.6% accuracy. In contrast, only
57% of LEAM crowns were correctly-labeled, and correct crowns only averaged
78.9% pixel accuracy. As expected, mislabeled LEAM crowns were mainly
dominated by DIPA pixels, the other low-LAI species.
99
As recommended by Meyer et al. (1996), I next set a 35% pixel-majority
threshold for determining the class of an ITC; that is, a crown was labeled
unclassified if the majority class comprised less than 35% of within-crown pixels.
With the threshold set and a LDA classifier applied to 30 PixelSUN bands, only 2
ITCs were affected. One ITC was a HYME that had been misclassified as a BAEL,
while the other ITC was a BAEL that had been misclassified as a DIPA. Since no
correctly-labeled ITCs were affected by the threshold, overall accuracy remained at
85.5%. However, the threshold helped decreased the commission error for BAEL
and DIPA by 2.3% and 0.9%, respectively, and so the threshold appears useful for
filtering ITCs with low accuracies.
3.3.5.2. Multispectral, broadband classification
Pixel-majority LDA and ML classification accuracies were significantly greater
for 10 to 30 narrowband, HYDICE imagery than for simulated IKONOS and ETM
broadband imagery, which had 4 and 6 bands, respectively (Table 3.9). Overall
accuracy with ASTER data (9 bands) and an ML classifier was 78.0% (PixelSUN),
which was not significantly lower than the pixel-majority accuracies achieved using
10 to 30 HYDICE bands with either the LDA or ML classifiers.
3.4. Discussion
3.4.1. Spatial scale and the spectral classification of TRF tree species
There was a decrease in classification accuracy from fine to coarser scales of
spectral measurement (i.e., leaf to pixel and crown scales). Some of this trend can
100
be explained by the differences in sensors used. Leaf-scale spectra had a relatively
high ratio of signal to noise because they were measured in a controlled laboratory
environment with a well-calibrated, high spectral resolution instrument. In contrast,
pixel- and crown-scale spectra had considerably more noise due to poor sensor
radiometric calibration and atmospheric effects.
Leaf spectral variability among individuals of a certain species, or even within a
single crown, was attributed to differences in internal leaf structure and biochemistry
(e.g., water, chlorophyll content, epiphyll cover and herbivory). Leaves have nonLambertian properties and physical differences in adaxial leaf cuticle (e.g., microtopography, wax, leaf hairs) affect first-surface specular reflectance, especially in the
VIS region with large incident and/or view angles (Grant, 1987). Another source of
variability among leaves of the same species was thus introduced by measuring
laboratory bidirectional (as opposed to hemispherical) reflectance with varying leaf
orientations.
Despite the multiple factors causing spectral variation, I found that leaf spectral
variability among species was significantly greater than that within species. Crossvalidation classifications confirmed that leaf-scale reflectance could discriminate
among species with >89% overall accuracy using as few as 10 optimally-positioned
bands. Important bands were concentrated in the NIR and SWIR1, where diffusereflectance dominates and variability is largely controlled by internal leaf structure
and water content (Gausman, 1985; Grant, 1987).
Pixel-scale measurements acquired with the airborne hyperspectral sensor were
dominated by leaf-scale spectral properties. Water absorption was enhanced at this
101
coarser scale by multiple-scattering of photons among leaves, stems and branches
(Asner, 1998; Roberts et al., 2004). In the NIR region, high levels of multiplescattering caused the 980 and 1200 nm water absorption features to deepen, and
overall NIR variability increased. The band-selection scheme for pixel spectra
identified important bands bordering the NIR water absorption features, possibly
detecting species differences in photon scattering caused by fine-scale crown
architecture (e.g., LAI). Bands in the visible part of the spectrum were also
important at pixel scales when considering all sunlit and shaded samples. The bluegreen edge, yellow edge and red well bands chosen may be sensitive to species
differences in spectral properties caused by their relative spectral mixing of leaf and
bark fractions; however, my data do not allow me to test this hypothesis. I originally
hypothesized that isolating sunlit regions of crowns for pixel-scale analysis would
lower within-species spectral variance and enhance species separability. This
hypothesis was confirmed by larger sunlit-sample F-ratios with spectral angle and
Euclidean distance metrics (Table 3.5). LDA and ML classifications also showed
significant improvements in accuracy with PixelSUN over PixelALL samples (Table
3.6). Gong et al. (1997) found similar results when classifying conifer spectra
acquired at 6-8 cm spatial scales.
At the crown scale, the architectural arrangement of crown components, such as
leaves and branches, controls the relative amounts of shading and complex,
anisotropic multiple-scattering of photons relative to the illumination and view
geometry—described by each crown’s bidirectional reflectance distribution function
(BRDF). Crown-scale spectra in my research were a linear average of within-crown
102
pixel spectra. The averaging of shadowed and bright pixels lowered mean crownscale reflectance, and as is expected by averaging, crown-scale variance was low
relative to pixel-scale variance. Band selection indicated that the NIR and SWIR2
were the main regions producing crown-scale separability among species. As with
pixel scale spectra, NIR reflectance is largely controlled by structural properties
(e.g., density and arrangement of leaves) that influence the photon scattering
environment and subsequent NIR reflectance. SWIR2 variability among species
may be related to two factors: overall differences in crown water concentration that
affects the expression of water absorption features at 1400, 1900 and 2700 nm
(Gausman, 1985; Roberts et al., 2004); and, ligno-cellulose absorption features that
may be expressed when high fractions of non-photosynthetic woody tissues are
exposed to the sensor (Asner, 1998; Curran, 1989), such as in a low-LAI deciduous
crown. Other methods for capitalizing on species-level differences caused by crown
structure and their influence on high spatial resolution imagery are discussed in
Section 3.4.2.
The SAM classifier was the least successful of the classifiers, regardless of the
spatial scale or spectral region considered. This result was surprising since there
were highly significant statistical separations of species with the spectral angle
metric at leaf and pixel scales (Table 3.5). SAM does not use second-order statistics
(e.g., covariance), and basing a classification on a single distance metric appears
ineffective given within-species spectral diversity.
LDA was highly accurate at all scales of analysis, indicating that spectral
covariance information—pooled for all species—is important for species
103
discrimination. The leaf-scale and crown-scale ML classifiers may not have
performed as well as LDA because of two factors. For one, the bands selected by
the stepwise procedure were optimized for LDA classification. Other band-selection
techniques, such as using Bhattacharyya distance (Haertel & Landgrebe, 1999), may
improve ML classification accuracy. Furthermore, ML requires a large number of
training samples for adequate estimation of the class covariance matrices from
higher-dimensional data that may contain redundant and noisy information. Only
with pixel-scale and pixel-majority classifications did I have a large enough sample
size to adequately assess ML with high dimensional data (i.e., 20+ bands). Although
ML performed slightly better at pixel scales relative to LDA, pixel-majority
classifications were generally higher with LDA. The apparent advantage of the ML
classifier at pixel scales may be spurious because only subsets of pixels from ITCs
were used for training and testing in the analyses, while pixel-majority
classifications used all ITC pixels through cross-validation. The crown-scale ML
classifications suffered from a lack of samples (individual trees) when estimating
class covariance matrices. Collecting large training sets is challenging in TRF
because target species have low densities, thus requiring large field surveys of forest
that is often difficult to access. The LDA classifier appears more appropriate for
TRF species discrimination because it strengthens the covariance estimation by
pooling information from all species.
Results from my pixel-scale LDA classification analysis can be compared to a
study by Fung et al. (1998), who used laboratory measurements of branch spectra to
classify 12 subtropical tree species with 84% overall accuracy (Producer’s
104
accuracies from 56 to 100%). LDA and 90 bands from VIS to NIR were used in the
analysis. In my analysis of sunlit pixels (i.e., branch scale) and 60 bands (VIS to
SWIR2 sampled), overall classification accuracy was 85% and Producer’s accuracies
ranged from 74% to 95%. Overall accuracy did not increase when using the fullspectrum of 161 bands. These results are encouraging since my airborne data suffers
from multiple factors that could confound species discrimination, such as mixed
pixels in training and testing data, variable illumination and viewing geometry, and
noise introduced by atmospheric conditions and non-target biological organisms
(e.g., lianas).
At all scales of observation, I noted an increase in accuracy with increased data
dimensionality to a certain level. Results using the LDA and ML classifier revealed
a general increase in accuracy up to 30 bands, while including more bands yielded
equal or lower accuracy.
My results confirm the benefits of hyperspectral over multispectral data for TRF
tree identification. At all scales, the best accuracies with hyperspectral data were
higher than those achieved with simulated multispectral imagery. Here I have
applied fairly conventional analytical techniques that select optimal bands and then
apply classifiers to narrowband reflectance values. Band selection was a necessary
analytical step to isolate the most important bands for reliable classifier parameter
estimation given my training data limitations (Duda & Hart, 1973). However, one
major advantage of contiguous hyperspectral bands is their continuous description of
spectral space, allowing measurements of the shape and position of key spectral
features, such as liquid water absorption features in NIR. Purely hyperspectral
105
analytical techniques exist and include spectral shape filters (Cochrane, 2000) and
analyses of spectral first- and second-order derivatives. First-order derivatives have
been shown to improve tree species classifications over the use of reflectance spectra
(Gong et al., 1997; van Aardt & Wynne, 2001). However, current research has
relied upon data from portable spectrometers in anticipation of high spatial and
spectral resolution data from future airborne or spaceborne sensors. Abiotic and
biotic noise, such as atmospheric water vapor, epiphytes, and lianas, will complicate
the radiance from a TRF canopy acquired by an airborne or satellite sensor, and
much research is needed to test hyperspectral-based classification techniques under
these challenging conditions.
3.4.2. Classification of individual tree crowns
Classifications of ITCs using crown-scale spectra had relatively high accuracies.
The maximum accuracy achieved was 92% with the LDA classifier and 30 bands.
Producer’s accuracies ranged from 70% (CEPE) to 100% (TEOB), and User’s
accuracies ranged from 81% (LEAM) to 100% (TEOB). My overall classification
accuracy is higher than the 65% accuracy Leckie and Gougeon (1999) observed for
temperate hardwood classification with crown-scale spectra, and it is similar to the
93% accuracy achieved by van Aardt and Wynne (2001) using in situ crown-scale
spectra to classify 3 hardwood tree species (second derivatives used, sunlit samples).
As reported in studies with conifer trees (Gougeon, 1995; Leckie, Gougeon, Hill et
al., 2003), there was no evidence that sunlit crown spectra could be more accurately
classified than by averaging the spectra from all pixels within the crown. My results
106
are encouraging considering that with visual interpretation of tropical tree species,
Clément and Guellec (1974) could only identify their target species with 73%
accuracy, and Myers and Benson (1981) visually-interpreted only 22% of their
species with >75% accuracy. I therefore conclude that spectral-based classification
of TRF tree species is possible, and accuracy is comparable or potentially greater
than from visual interpretation of aerial photographs. Furthermore, computer-based
classification permits the automation and removal of subjectivity from the process.
In the crown-scale LDA classification, there was inter-species confusion
between individuals of Lecythis and Dipteryx (LEAM and DIPA, Table 3.7). This
confusion is attributed to the deciduous phenology of these species—individuals had
very low crown LAI and similar spectral properties. Bark lichen, epiphytes, and
understory plants are also more likely to be exposed to the sensor in low-LAI
crowns, and spectra from these components could dilute tree species spectral
differences. Crown-scale spectra from DIPA and LEAM crowns were thus distinct
from other species, but confused between the two species.
Despite this confusion, DIPA crowns had 92.6% Producer’s accuracy and 89.3%
User’s accuracy (Table 3.7). Furthermore, the User’s accuracy could be increased to
91.8% by applying a LDA probability threshold (Table 3.8). These results are
encouraging because large Dipteryx trees have an important ecological function in
providing a major seed resource and nesting cavities for the endangered Great green
macaw (Ara ambigua) (pers. comm., Powell 2001). Deforestation in the Sarapiquí
region has mostly eliminated large Dipteryx trees outside of protected areas, thereby
contributing to a dramatic decline in the macaw population. Remote sensing
107
technology that can identify large Dipteryx crowns may contribute to macaw
conservation efforts by providing a rapid and cost-effective means to map macaw
habitat and migration corridors across the region.
In this chapter, another ITC classification technique was to label crown objects
using the majority class of classified within-crown pixels. Relative to crown-scale
LDA classification with 30 optimal bands, the pixel-majority classification scheme
had 8.5% and 2.4% lower accuracy with PixelALL and PixelSUN samples, respectively
(Table 3.6). However, one operational advantage of the pixel-majority approach to
ITC classification is that robust training statistics (e.g., the covariance matrix) can be
estimated from a relatively small number of individual crowns per species because
each crown contains many pixels. Furthermore, the pixel-majority approach allows
for the inevitable spectral variation within crowns that leads to misclassified
samples. I observed that correctly-labeled ITCs had a mean within-crown pixel class
accuracy of 90%. ITCs with more mixed-class pixels had more suspect species
labels and were often mislabeled. I found that a pixel-majority threshold (e.g.,
majority class must have ≥35% of within-crown pixels) could be used to improve
species User’s accuracy by excluding low-accuracy ITCs and reducing commission
errors; and, overall accuracy was left unchanged with the threshold. In contrast,
crown-scale spectra blur within-crown variation. Basing an ITC species label on a
single, crown-scale spectrum classification is risky due to relatively weak spectral
separation among species (Table 3.5, Euclidean distance). My research benefited
from distinct phenological differences among species, which undoubtedly helped
crown-scale species discrimination. The pixel-majority approach to classification
108
may prove useful if ITCs are more spectrally confused due to similar phenology and
crown architecture.
IKONOS and ETM+ overall accuracies were both lower than 67% for crownscale and pixel-majority ITC classifications. Since I used 1.6-m simulated
multispectral data, none of my tests considered the actual resolution of existing
sensors. The IKONOS multispectral sensor provides 4-m resolution images and so
could be amenable to either a crown-scale or pixel-majority classification scheme.
However, I found that 1.6-m resolution pixels with IKONOS-simulated bands only
provided 62% overall accuracy and DIPA User’s accuracy was 64% (LDA,
CrownSUN). However, nine ASTER bands could classify ITCs with 77% overall
accuracy with CrownSUN spectra, and DIPA User’s accuracy was 76%. ASTER has
2 VIS, 1 NIR, 1 SWIR1 and 5 SWIR2 bands, while IKONOS contains only 3 VIS
and 1 NIR bands. These additional SWIR2 bands in ASTER help discriminate
species. This conclusion is supported by my hyperspectral, narrowband analyses,
which also indicated that SWIR2 was important for species discrimination.
At crown scales, ten narrow bands from HYDICE imagery—with wavelength
positions in all spectral regions—produced 85% overall accuracy and DIPA User’s
accuracy was 93%. I therefore conclude that a high spatial resolution sensor with
10+ channels across the VIS to SWIR2 spectrum is necessary to classify TRF tree
species with reasonable accuracy (i.e., ≥85%). For a satellite sensor, finer spectral
resolution requires coarser fields of view due to limited surface photon flux.
Likewise, airborne sensors can cover a larger swath if flown at a higher altitude with
coarser spatial resolution pixels. The crown-scale results indicate that a spatial
109
resolution at the scale of tree crowns is adequate for species discrimination.
However, my technique calculated crown-scale spectra by averaging only those
spectra from within the ITC footprint, and in an operational situation, coarse-scale
square pixels (e.g., 10 to 30 m) may subsume background plant species, soil and
shadows, thereby producing mixed spectra and decreased classification accuracy.
Future studies should thus examine the sensitivity of classification accuracy to
spatial resolutions expected in operational circumstances (i.e., >1.6 m but less than
the scale of a crown).
I analyzed emergent trees for two reasons: to be certain crowns could be
identified in the imagery given georegistration errors, and to sample an adequate
number of pixels per crown. Higher spatial resolution sensors will allow the
investigation of more TRF tree species, especially co-dominant individuals that do
not have broad, emergent crowns. Leaf-scale analyses indicated that spectral
measurements with a very fine FOV are more accurately classified than spectrallymixed pixels from coarser spatial scales. For example, leaf spectra were classified
with 89% accuracy with just 10 bands and LDA classifier, while pixel and crownscale accuracies were < 85% for the same combination of bands and classifier (Table
3.6). However, these leaf-scale results are based on controlled laboratory conditions
with a few species; spectra measured at this fine of scale from an airborne sensor
will include atmospheric effects and spectral mixing from photon scattering among
various crown tissues and other plant species. It is thus unclear if airborne digital
sensors with very high resolutions (i.e., leaf scale) will allow species discrimination
with relatively few bands (e.g., Meyer et al., 1996).
110
3.5. Conclusions
My results confirm that species of tropical rain forest (TRF) trees can be
discriminated based on their spectral reflectance properties. Individual tree crowns
(ITCs) were successfully classified with 92% overall accuracy using 30 optimallyselected bands from crown-scale reflectance spectra and a linear discriminant
analysis classifier. Pixel-majority ITC classification, which labeled crowns based on
within-crown classified pixels, was not as accurate as crown-scale spectra
classification. At a fixed 1.6-m spatial scale, crown-scale ITC classification was
significantly more accurate with hyperspectral narrowband data (10 band HYDICE)
relative to accuracies achieved with multispectral broadband data (simulated
IKONOS, Landsat ETM+ and ASTER).
This chapter represents the first use of high spectral and spatial resolution
imagery acquired over TRF canopy for automated discrimination of individual tree
species. Similar to laboratory-based analyses by Cochrane (2000), my results
indicate that there are spectral differences among species that permit classification at
leaf to crown scales; however, there is also temporal and spatial spectral variation
within populations and even single individuals of TRF tree species that will
inevitably decrease classification accuracy. A major challenge is to develop
classification schemes that can maximize the spectral, spatial and temporal
information content of digital imagery while accommodating inherent variation
within species. In Chapters 4 and 5, I explore more elaborate classification
111
techniques for tree species discrimination that seek to fully exploit the hyperspectral
properties of the HYDICE data.
112
Table 3.1. Study tree species attributes (adapted from Frankie et al., 1974, O’Brien,
2001 and personal observation). Leaf cover is for late-March to early-April, and is
what would be expected for the majority of individuals for each species based on
available literature data and personal field observations.
Tree species
Code
Leaf
Leaf
3/30
[family or sub-family]
Phenology Exchange Leaf
Functional
Timing
Cover
Group
Balizia elegans
BAEL Deciduous
Annual
High
(Ducke) Barneby & Grimes
[Mimosoideae]
Ceiba pentandra
CEPE Deciduous
Annual
High
Gaertn.
[Bombacaceae]
Dipteryx panamensis
DIPA
Deciduous
Annual
Low
(Pittier) Record & Mell
[Papilionoideae]
Hyeronima alchorneoides
HYAL Evergreen Continuous High
Allemão
[Euphorbiaceae]
Hymenolobium mesoamericanum HYME Deciduous Sub-annual High
Lima
[Papilionoideae]
Lecythis ampla
LEAM Deciduous
Annual
Low
Miers
[Lecythidaceae]
Terminalia oblonga
TEOB Evergreen Continuous High
(Ruiz & Pav.) Steud.
[Combretaceae]
Table 3.2. Summary of characteristics of individual tree crowns from HYDICE data.
Species names defined in Table 3.1.
Tree
No. of Crown area, m2 All pixels/crown Sunlit pixels/crown
species crowns
mean (range)
mean (range)
mean (range)
BAEL
29
358 (108-699)
140 (42-273)
68 (19-131)
CEPE
10
766 (361-1695)
299 (141-662)
153 (62-338)
DIPA
81
519 (141-1167)
203 (55-456)
98 (28-227)
HYAL
34
388 (159-635)
152 (62-248)
78 (34-118)
HYME
14
479 (108-1009)
187 (42-394)
99 (23-185)
LEAM
21
349 (164-630)
136 (64-246)
67 (31-136)
TEOB
25
312 (105-543)
122 (41-212)
64 (21-110)
All
214
444 (105-1695)
174 (41-662)
87 (19-338)
113
Table 3.3. Laboratory leaf spectra summary. Species names defined in Table 3.1.
Tree species No. of samples No. trees sampled
(leaves per tree)
BAEL
16
3 (5-6)
CEPE
15
3 (5)
DIPA
30
5 (3-10)
HYAL
23
4 (6-10)
HYME
30
3 (6-10)
LEAM
14
3 (2-6)
TEOB
24
4 (5-7)
All
152
25
Table 3.4. Mean and standard deviation (s.d.) of percent reflectance across spectral
regions. Means and s.d. were calculated on a band-by-band basis and then averaged
for each spectral region.
VIS
NIR
SWIR1
SWIR2
Full Spectra
Leaf
5.8 (2.8)
41.3 (6.8) 26.3 (5.5)
12.7 (4.2)
20.9 (4.7)
PixelALL
2.7 (1.4)
34.8 (13.5) 15.1 (6.3)
5.5 (2.6)
14.4 (5.9)
PixelSUN
3.4 (1.1)
44.4 (10.3) 20.2 (4.8)
7.5 (2.3)
18.4 (4.6)
CrownALL 2.7 (0.8)
35.6 (6.9) 15.2 (2.9)
5.8 (1.7)
14.6 (3.0)
CrownSUN 3.5 (0.8)
45.0 (7.9) 19.7 (3.1)
6.9 (1.7)
18.5 (3.4)
114
Table 3.5. Non-parametric multivariate analysis of variance (NPMANOVA),
comparing among- and within-species spectral variation using spectral angle and
Euclidean distance. Values are F ratios. In all analyses, 5000 permutations were
used to test significance. Significance is: ns = not significant, * = p≤0.05, ** =
p≤0.01, *** = p≤0.001.
Crown Scale
Spectral Region
Bands Leaf Scale a Pixel Scale
Spectral Angle Distance
Full spectra
161
n/a
119.7 ***
5.4 *
Full spectra (sunlit) 161
16.1 ***
162.3 ***
7.7 ***
VIS
43
n/a
68.3 ***
1.3 ns
VIS (sunlit)
43
13.4 ***
103.3 ***
2.5 ns
NIR
46
n/a
113.0 ***
6.0 ***
NIR (sunlit)
46
12.6 ***
137.2 ***
8.3 ***
SWIR1
25
n/a
94.2 ***
0.8 ns
SWIR1 (sunlit)
25
26.6 ***
128.1 ***
1.4 ns
SWIR2
48
n/a
32.3 ***
3.0 ns
SWIR2 (sunlit)
48
10.2 ***
78.3 ***
3.6 *
Euclidean Distance
Full spectra
161
n/a
28.6 ***
3.7 ***
Full spectra (sunlit) 161
13.4 ***
180.3 ***
3.7 ***
VIS
43
n/a
13.1 ***
4.6 ***
VIS (sunlit)
43
4.1 ***
51.2 ***
4.5 ***
NIR
46
n/a
34.1 ***
2.8 ***
NIR (sunlit)
46
12.8 ***
195.7 ***
2.6 ***
SWIR1
25
n/a
16.4 ***
4.4 ***
SWIR1 (sunlit)
25
18.4 ***
133.7 ***
4.4 ***
SWIR2
48
n/a
61.2 ***
3.8 ***
SWIR2 (sunlit)
48
18.6 ***
157.4 ***
3.8 ***
a
Artificial illumination from halogen lamp
115
Table 3.6. Overall accuracy of classifiers using leaf-, pixel- and crown-scale narrowband
(HYDICE) spectra. Leaf-scale data were simulated HYDICE spectra from laboratory
measurements. “Pixel-majority” refers to ITC classification using a class-majority rule
applied to classified within-crown pixels. Arrows represent the direction of significant
improvement in overall accuracy between using all and sunlit-only samples (ns = not
significant at α=0.05).
Pixel d
Crown c
Pixel-majority c
Leaf c
Bands a
All
All
Sunlit
All
Sunlit
All
Sunlit
Linear Discriminant Analysis (LDA)
10 a
89.5
67.5 → 72.1
84.1 ns 84.6
74.8 ns 81.8
20 a
98.0
75.0 → 81.7
89.3 ns 85.5
81.3 ns 83.2
a
30
99.3
79.0 → 83.2
92.1 ns 87.9
83.6 ns 85.5
40 a
100.0
79.2 → 86.5
90.2 ns 89.7
84.1 ns 84.1
a
50
100.0
79.7 → 85.2
89.7 ns 91.1
83.6 ns 85.5
60 a
100.0
80.6 → 85.4
89.7 ns 90.2
84.1 ns 84.1
161 (Full)
88.8
80.9 → 85.5
61.2 ns 67.8
81.8 ns 83.6
b
VIS
60.5
45.1 → 53.1
65.9 ns 66.4
62.1 ns 67.3
NIRb
80.9
54.7 → 61.5
67.8 ns 67.3
60.3 ns 66.4
SWIR1 b
88.8
55.1 → 59.8
69.2 ns 68.7
64.0 ns 71.5
b
SWIR2
81.6
47.8 → 57.6
75.7 ns 76.6
64.5 ns 69.6
Maximum Likelihood (ML)
10 a
80.9
69.5 → 76.1
76.6 ns 77.6
77.6 ns 79.4
20 a
n/a
76.4 → 86.2
n/a
84.6 ns 82.7
30 a
n/a
79.5 → 86.7
n/a
82.7 ns 80.4
40 a
n/a
81.6 → 86.9
n/a
82.2 ns 80.4
50 a
n/a
81.2 → 87.3
n/a
79.0 ns 79.9
a
60
n/a
79.9 → 87.6
n/a
78.0 ns 77.6
161 (Full)
n/a
68.3 → 79.1
n/a
71.0 ns 71.5
VIS b
64.5
49.7 → 58.7
30.4 ns 27.1
62.6 ns 67.3
b
NIR
71.7
59.6 → 66.3
50.0 ← 37.9
69.2 ns 70.1
SWIR1 b
82.2
53.5 → 66.2
34.1 → 46.7
66.8 ns 69.2
b
SWIR2
71.1
49.2 → 58.1
39.7 ns 49.1
65.4 ns 71.0
116
Table 3.6 (continued).
Spectral Angle Mapper (SAM)
10 a
46.1
42.4 ns 44.1
45.8 ns 50.9
48.6 ns 42.5
a
20
50.7
37.9 → 48.9
46.3 ns 53.7
48.1 ns 56.5
30 a
48.0
39.8 → 48.6
46.7 ns 47.7
46.3 ns 50.9
40 a
48.7
39.2 → 46.4
45.8 ns 50.5
46.3 ns 52.8
a
50
47.4
38.9 → 47.0
46.7 ns 51.4
47.7 ns 53.3
60 a
46.1
38.0 → 48.4
46.3 ns 50.5
46.3 ns 50.0
161 (Full)
48.7
38.7 → 48.4
44.4 ns 48.6
43.9 ns 47.2
b
VIS
35.5
38.8 → 42.1
14.0 ns 12.6
41.6 ns 43.0
NIR b
38.2
42.3 → 47.1
37.4 ns 39.3
48.6 ns 49.5
b
SWIR1
37.5
33.5 → 38.0
45.3 ns 51.9
37.4 ns 41.6
SWIR2 b
36.8
31.6 → 36.8
45.8 ns 44.9
43.0 ns 39.3
a.
Bands selected using a Linear Discriminant Analysis (LDA) forward, stepwise
selection procedure. Significance criteria α =0.05 for leaf, pixel and crown objects, α =
0.2 for CrownALL and CrownSUN spectra).
b.
VIS, NIR, SWIR1 and SWIR2 regions have 10 evenly-spaced bands.
c.
Accuracy results are from cross validation of samples.
d.
Training and testing data were two mutually-exclusive sets of 300 randomly-selected
pixels per species.
117
Classification
118
Table 3.7. Error matrix for crown-scale classification using 30 bands and a Linear
Discriminant Analysis classifier with no a posteriori probability threshold (Kappa = 0.90).
Field Reference
Species
BAEL CEPE
DIPA HYAL HYME LEAM
TEOB Total
27
1
28
BAEL
7
1
8
CEPE
1
2
75
1
1
4
84
DIPA
1
33
34
HYAL
1
13
14
HYME
1
3
17
21
LEAM
25
25
TEOB
Unknown
29
10
81
34
14
21
25
214
Total
93.1% 70.0% 92.6% 97.1% 92.9%
81.0% 100.0%
Prod. %
Species names defined in Table 3.1.
118
User %
96.4%
87.5%
89.3%
97.1%
92.9%
81.0%
100.0%
92.1%
Classification
119
Table 3.8. Error matrix for crown-scale classification using 30 bands and a Linear Discriminant
Analysis classifier with a 90% a posteriori probability threshold (Kappa = 0.79).
Field Reference
Species
BAEL CEPE DIPA HYAL HYME LEAM TEOB Total User %
25
1
26
96.2%
BAEL
6
1
7
85.7%
CEPE
1
1
67
1
3
73
91.8%
DIPA
1
32
33
97.0%
HYAL
1
11
12
91.7%
HYME
1
12
13
92.3%
LEAM
25
25
100.0%
TEOB
2
3
10
2
2
6
25
Unknown
29
10
81
34
14
21
25
214
Total
83.2%
Prod. % 86.2% 60.0% 82.7% 94.1% 78.6% 57.1% 100.0%
Species names defined in Table 3.1.
119
Table 3.9. Classification results for simulated broadband spectra. Bands are for
IKONOS, Landsat ETM+ and ASTER sensors. Scales follow Table 3.6.
Pixel e
Crown d
Pixel-majority d
Leaf d
a
Bands
All
All
Sunlit
All
Sunlit
All
Sunlit
Linear Discriminant Analysis (LDA)
IKONOS a
44.7
42.7 → 49.2
59.3 ns 61.7
52.3 ns 59.3
ETM+ b
57.9
48.6 → 55.8
66.4 ns 66.4
60.3 ns 64.5
ASTER c
80.3
56.4 → 64.7
76.2 ns 77.1
66.8 ns 72.0
Maximum Likelihood (ML)
IKONOS a
61.2
45.3 → 55.1
50.5 ns 50.9
48.1 ns 51.9
b
ETM+
73.7
54.0 → 62.4
62.6 ns 62.1
62.6 ns 65.9
ASTER c
83.6
59.4 → 70.4
72.9 ns 72.4
77.6 ns 78.0
Spectral Angle Mapper (SAM)
IKONOS a
40.1
45.3 ← 33.6
20.1 ns 31.3
20.1 ns 31.8
ETM+ b
34.2
32.5 → 35.9
38.3 ns 39.7
34.1 ns 41.1
ASTER c
40.1
30.6 → 36.1
41.1 ← 31.3
38.3 ns 37.9
a.
4 bands - 483, 551, 663, 794 nm.
b.
6 bands - 479, 561, 661, 835, 1651, 2209 nm.
c.
9 bands - 555, 658, 805, 1655, 2166, 2207, 2264, 2333, 2394 nm.
d.
Accuracy results are from cross validation of samples.
e.
Training and testing data were two mutually-exclusive sets of 300 randomly-selected
pixels per species.
120
Classification
121
Table 3.10. Error matrix for pixel-majority classification using 30 bands, sunlit pixels, a Linear
Discriminant Analysis classifier, and no pixel-majority threshold (Kappa = 0.82).
Field Reference
Species
BAEL CEPE DIPA HYAL HYME LEAM TEOB
Total
User %
24
3
3
1
1
32
75.0%
BAEL
5
5
100.0%
CEPE
2
5
76
1
1
8
93
81.7%
DIPA
2
1
30
33
90.9%
HYAL
11
11
100.0%
HYME
1
12
13
92.3%
LEAM
25
25
100.0%
TEOB
1
1
2
Unknown
29
10
81
34
14
21
25
214
Total
82.8% 50.0% 93.8% 88.2% 78.6% 57.1% 100.0%
85.5%
Prod. %
Species names defined in Table 3.1.
121
La Selva
1
Km
Costa Rica
Tree crowns
HYDICE extent
Rivers
Land Use
Developed Areas
Selectively- logged
Old- growth Forest
Secondary Forest
Pasture
Plantation
Swamp
Figure 3.1. The La Selva Biological Station study site and extent of HYDICE
hyperspectral imagery. The 214 study crowns are labeled with points.
122
A
B
SPECIES
BAEL
CEPE
DIPA
HYAL
HYME
LEAM
TEOB
Figure 3.2. A) View of old-growth Tropical Wet Forest at the La Selva Biological
Station. The canopy-emergent tree in the foreground is Balizia elegans. B)
Example of 1.6-m spatial resolution HYDICE hyperspectral imagery over oldgrowth canopy (Red: 1651 nm [SWIR2], Green: 835 nm [NIR], Blue: 661 nm [Red])
with overlaid individual tree crown polygons. Map scale is 1:3000.
123
0.50
HYDICE Canopy
0.45
HYDICE Bridge
ASD Canopy
0.40
ASD Bridge
Reflectance
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
350
850
1350
1850
2350
Wavelength (nm)
Figure 3.3. Reflectance spectra from airborne HYDICE and field ASD
spectrometers for a wooden bridge (over dark water) and a Pentaclethra
macrophylla crown. The Pentaclethra ASD spectrum was acquired from the bridge.
124
Reflectance
Reflectance
Reflectance
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
BAEL (N =16)
850
1350
1850
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
DIPA (N = 30)
850
1350
1850
HYME (N = 30)
850
1350
1850
CEPE (N = 15)
850
1350
1850
2350
HYAL (N = 23)
850
1350
1850
2350
LEAM (N = 14)
850
1350
1850
2350
Reflectance
Wavelength (nm)
TEOB (N = 24)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
850
1350
1850
2350
Wavelength (nm)
Figure 3.4. Leaf-scale mean (bold line) and standard deviation (±1 S.D., thin line) of
reflectance by species. Species codes are listed in Table 3.1.
125
A
0.60
100%
Reflectance
0.50
50%
0.40
0%
0.30
0.20
0.10
0.00
400
650
900
1150
1400
1650
1900
2150 2400
Wavelength (nm)
B
0.60
20%
Reflectance
0.50
10%
0.40
0%
0.30
0.20
0.10
0.00
400
650
900
1150
1400 1650
1900
2150 2400
Wavelength (nm)
C
0.60
Senesced
Reflectance
0.50
Mature
Young
0.40
0.30
0.20
0.10
0.00
400
650
900
1150
1400
1650
1900 2150
2400
Wavelength (nm)
Figure 3.5. Leaf-scale variation in spectral properties. A) Hymenolobium
mesoamericanum leaves: percent area covered by a single species of epiphyll. B)
Lecythis ampla leaves: percent area of leaf herbivory (i.e., light brown-colored
mines) caused by a leaf-mining insect. C) Terminalia oblonga leaves: leaf aging
from young to senesced leaves. Note: all spectra in Fig. 3.5 were from upper-canopy
leaves and were included in leaf-scale spectral analyses except for the Terminalia
senesced leaf, which was collected on the ground.
126
Reflectance
Reflectance
Reflectance
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
BAEL (N =300)
850
1350
1850
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
DIPA (N =300)
850
1350
1850
HYME (N = 300)
850
1350
1850
CEPE (N = 300)
850
1350
1850
2350
HYAL(N = 300)
850
1350
1850
2350
LEAM (N = 300)
850
1350
1850
2350
Reflectance
Wavelength (nm)
TEOB (N = 300)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
850
1350
1850
2350
Wavelength (nm)
Figure 3.6. Mean (bold line) and standard deviation (±1 S.D., thin line) of
reflectance by species for sunlit pixels. Species codes are listed in Table 3.1.
127
Reflectance
Reflectance
Reflectance
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
BAEL (N = 29)
850
1350
1850
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
2350
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
DIPA (N = 81)
850
1350
1850
HYME (N = 14)
850
1350
1850
CEPE (N = 10)
850
1350
1850
2350
HYAL (N = 34)
850
1350
1850
2350
LEAM (N = 21)
850
1350
1850
2350
Reflectance
Wavelength (nm)
TEOB (N = 25)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
350
850
1350
1850
2350
Wavelength (nm)
Figure 3.7. Crown-scale mean (bold line) and standard deviation (±1 S.D., thin line)
of reflectance by species for CrownALL spectra. Species codes are listed in Table 3.1.
128
Leaf Scale
A
1.20
Reflectance
1.00
All 20
All 10
0.80
0.60
0.40
0.20
0.00
350
850
1350
1850
2350
Wavelength (nm)
Pixel Scale
B
1.20
Sunlit 20
Sunlit 10
Reflectance
1.00
All 20
All 10
0.80
0.60
0.40
0.20
0.00
350
850
1350
1850
2350
Wavelength (nm)
Crown Scale
C
1.20
Sunlit 20
Sunlit 10
Reflectance
1.00
0.80
All 20
All 10
0.60
0.40
0.20
0.00
350
850
1350
1850
2350
Wavelength (nm)
Figure 3.8. Mean reflectance spectra for the seven study species at A) leaf, B) pixel
and C) crown scales from Figures 3.4, 3.6 and 3.7, respectively. Pixel and crown
spectra were calculated from sunlit-only (Sun) and all (All) pixel spectra prior to
band selection. PixelSUN and CrownALL spectra are displayed in B and C,
respectively. Dots above spectra represent the best 10 and 20 bands selected by the
stepwise-selection procedure.
129
100
Percent
80
60
SWIR2
SWIR1
40
NIR
VIS
20
0
Leaf Pixel
(All)
Pixel Crown Crown
(Sun) (All) (Sun)
100
Percent
80
SWIR2
SWIR1
NIR
VIS
60
40
20
0
Leaf Pixel
(All)
Pixel Crown Crown
(Sun) (All) (Sun)
Figure 3.9. The percentage of 10 (A) and 20 (B) stepwise-selected bands within each
spectral region at leaf, pixel and crown scales.
130
Percent Accuracy
100
90
80
70
60
50
40
30
20
10
0
90%
75%
50%
No Probability
0
5
10
15 20
25 30
35 40
45 50
55 60
Number of Bands
Figure 3.10. Crown-scale classification overall accuracy with the addition of bands.
Bands were added based on their ranking using a stepwise procedure and the
classifier was linear discriminant analysis (LDA). The a posteriori LDA probability
required for a crown to be classified was adjusted to four different thresholds: 1) no
probability threshold and 50%, 75% and 90% thresholds.
131
CHAPTER 4: Classification of tree species with absorption features and binary
decision trees
4.1. Introduction
Digital data from high spatial resolution satellite or airborne optical sensors
provide a new means of mapping and monitoring individual tree crowns (ITCs) over
larger areas and with more frequency than traditional field-based techniques (Clark,
Read, et al., 2004; Gougeon & Leckie, 2003). Research in tropical forests has
shown that the spatial and temporal resolution of this imagery permits the remote
measurement of ITC growth and mortality, as well as canopy disturbances such as
gap formation from selective logging (Clark, Read et al., 2004; Clark, Castro et al.,
2004; Hurtt et al., 2003; Read et al., 2003). In temperate forests, high spatial
resolution multispectral imagery has been used to detect, delineate and classify the
species of tree crowns (Gougeon & Leckie, 2003; McGraw et al., 1998). In tropical
forests, however, within species (conspecific) variation in reflectance properties due
to crown architecture, phenology and other biotic factors severely limits the specieslevel discrimination of ITCs with low spectral resolution, multispectral data (Chapter
3).
Hyperspectral sensors measure the visible to shortwave infrared region (4002500 nm) of the electromagnetic spectrum in over 100 channels, and this increased
spectral resolution may permit species discrimination based on subtle differences in
their reflectance spectra. When viewed from above by an airborne sensor, spectral
132
variation among tree species is controlled by two principal factors: biochemistry and
structure. Leaves, woody branches and trunks, flowers and fruits each have their
own biochemical properties that create distinct absorption features in reflectance
spectra. Green leaves have photosynthetic pigments—mainly chlorophyll-a, -b, or c and caroteniods (carotenes, xanophylls) (Gates et al., 1965; Raven et al., 1992).
These pigments absorb light in the blue region of the electromagnetic spectrum (near
445 nm), but only chlorophyll pigments absorb light in the red region (645-680 nm).
Green leaves also have weak water absorption features in the near-infrared region at
970 nm and 1240 nm, respectively (Gao & Goetz, 1990; Gates et al., 1965). In
contrast, bark spectra typically have no photosynthetic pigment signal, but instead
have absorption features in the shortwave infrared region due to biochemical
compounds, such as lignin, cellulose and proteins (Curran, 1989; Elvridge, 1990).
The internal and surface structures of plant tissues also play significant roles in
multiple-scattering and transmission of photons within their canopies (Gates et al.,
1965; Grant, 1987). At a coarser scale, species-level differences in crown structure,
such as leaf and branch area, angular distribution, and clumping, affect the
expression of tissue biochemical absorption features in branch- to crown-scale
reflectance spectra (Asner, 1998).
I hypothesize that if tropical rain forest tree species differ in crown structure or
tissue biochemistry, then hyperspectral metrics that target key absorption features
should be useful for species-level individual tree crown classification. Several
recent studies have correlated hyperspectral reflectance bands or derived metrics
133
with canopy physicochemical properties to classify temperate forest composition
(Fuentes et al, 2001; Kokaly et al., 2003; Martin et al., 1998; Ustin & Xiao, 2001).
Cochrane (2000) noted differences in red-edge metrics for tropical rain forest tree
species when analyzing laboratory spectra, but there have been no studies that have
used hyperspectral-derived metrics for tropical tree species classification.
In this chapter, I explore hyperspectral metrics and a decision-tree classification
scheme for automatic classification of individual tropical rain forest tree species
from high spatial resolution imagery. My main objectives were to: 1) compute a
suite of hyperspectral metrics that respond to crown structure and absorption features
from photosynthetic pigments, water and other biochemicals; 2) at tissue, pixel and
crown scales, test for significant differences in mean response of these spectral
metrics among tropical tree species; and, 3) assess ITC classification accuracy using
binary decision trees applied to spectral metrics.
4.2. Background
4.2.1. Hyperspectral metrics
There are many methods for measuring the absorption features expressed in
hyperspectral data. One approach is to use a ratio-based index that contrasts an
absorption feature band with a band from a relatively stable spectral region (Elvidge
& Chen, 1995; Peñuelas et al., 1997b). More sophisticated techniques involve using
the multiple, contiguous reflectance bands within a hyper-spectrum to identify the
shape and position of key absorption features (Broge & Leblanc, 2000; Clark,
134
Swayze et al., 2003). Common metrics are based on first- and second-order
derivatives of reflectance spectra (Demetriades-Shah et al., 1990; Elvridge & Chen,
1995) and absorption feature area, width or depth (Clark, Swayze et al., 2003; Gong
et al., 2002; Kokaly et al., 2003; Pu, Ge et al., 2003). A first-order derivative
spectrum is the slope of the reflectance spectrum at every point. The first- and
second-order derivatives can be used to resolve overlapping absorption features and
identify inflection points (Demetriades-Shah et al., 1990). Below I discuss potential
spectral metrics related to key vegetation absorption features. In this discussion, I
refer to three regions of the electromagnetic spectrum: visible (VIS=437-700 nm),
near-infrared (NIR=700-1327 nm), and shortwave-infrared (SWIR=1467-2435 nm).
4.2.2. Absorption features related to photosynthesis
Vegetation indices are common in remote sensing and generally measure
properties of photosynthetic pigment absorptions.
I investigated the following
vegetation indices (formulas in Table 4.1): Simple Ratio Vegetation Index (SR;
Jordon, 1969; Tucker, 1979), Normalized Difference Vegetation Index (NDVI;
Rouse et al., 1973), Soil-Adjusted Vegetation Index (SAVI; Huete, 1988),
Photochemical Reflectance Index (PRI; Gamon et al., 1997), Atmospherically
Resistant Vegetation Index (ARVI; Kaufman & Tanré, 1992), Enhanced Vegetation
Index (EVI; Huete et al., 2002), and the Red-edge Vegetation Stress Index (RVSI;
Merton, 1998). The SR, NDVI, SAVI, ARVI, and EVI indices were developed for
broadband multispectral sensors (e.g., Landsat, AVHRR), but they can be
135
formulated with narrowband wavelengths from hyperspectral data (Elvidge & Chen,
1995; McGwire et al. 2000).
The popular SR and NDVI measure the contrast in red-light chlorophyll
absorption relative to high NIR reflectance caused by internal leaf and canopy
scattering (Tucker, 1979). At relatively coarse spatial scales (e.g., >900 m2), these
indices have been linked to forest biophysical parameters such as leaf area index
(LAI), aboveground biomass, and primary productivity (Elvidge & Chen, 1995;
Spanner et al. 1994), and have contributed to forest-type classification (Steininger,
1996), mapping of plant species richness (Oindo & Skidmore, 2002), and change
detection (Ferreira et al., 2003). The NDVI and SR are sensitive to background soil
and atmospheric contamination, and SAVI and ARVI were introduced to minimize
these two factors, respectively (Huete et al., 1985; Huete, 1988; Kaufman & Tanré,
1992). At coarse spatial scales, the EVI minimizes both soil and atmospheric noise
and is more sensitive to vegetation structure (e.g., LAI, canopy architecture) than
NDVI (Huete et al., 2002). The PRI is a narrowband ratio that is sensitive to
changes in reflectance at 531 nm caused by interconversions of xanthophyll cycle
pigments, which are related to photosynthesis light-use efficiency across species,
functional types, and nutrient treatments (Gamon et al., 1997; Peñuelas et al.,
1997a).
Few studies have investigated structural and chemical variation among plant
species as expressed through these ratio-based indices. Using field measurements of
6 tree species in Africa, Franklin et al. (1991) found strong relationships between
NDVI and crown biomass for each species. Nagler et al. (2004) found that NDVI
136
calculated from fine spatial resolution digital imagery was significantly different
among individual riparian species. They attributed variation among species to leaf
angle effects on light attenuation rather than LAI. In exploring narrowband indices
for species-independent chlorophyll estimation, Sims and Gamon (2002) noted that
relationships were weakened by species differences in leaf surfaces (e.g., waxes,
pubescence), causing variation in first-return reflection.
First-order derivative bands derived from hyperspectral reflectance spectra have
been used to estimate foliar content of chlorophyll, nitrogen, phosphorus, and
potassium in conifer and mixed-deciduous forest canopies (Gong et al., 2002; Martin
& Aber, 1997; Smith et al., 2002). Derivative analysis is useful for finding the
inflection point on the tail of an absorption feature. The wavelength of the red-edge
inflection point (RE-λ), between 680 nm and 740 nm, is correlated to chlorophyll
concentration, LAI and vegetation stress (Blackburn, 1998; Demetriades-Shah et al.,
1990; Horler et al., 1983; Pu, Gong et al., 2003), and it may be useful for
discriminating tropical rain forest tree species (Cochrane, 2000).
4.2.3. Water absorption features
Green vegetation spectral reflectance in the 900 to 2500 nm region of the
spectrum is dominated by liquid water absorption, and narrowband sensors can
exploit measurements in this region to estimate vegetation moisture content. Two
useful water indices are the Water Band Index (WBI; Peñuelas et al., 1997b) and the
Normalized Difference Water Index (NDWI; Gao, 1996), which measure NIR water
absorption features at 970 nm and 1240 nm, respectively (Table 4.1). These indices
137
are not influenced by chlorophyll absorption, and so they are complementary to
indices based on red absorption (Gao, 1996).
Near-infrared water absorption features are also useful for estimating leaf
equivalent water thickness (EWT), which is the amount of liquid water in vegetation
needed to account for an observed water absorption feature (Roberts et al., 1997).
The metric is calculated by fitting a Beer-Lambert model of light extinction through
a water-absorbing medium to the vegetation spectrum covering either the 950–1000
or 1150–1260 nm absorption regions (Roberts et al., 1997). Research has shown that
EWT and other hyperspectral-derived metrics correlate with moisture content and
structure of canopy components (Dennison et al., 2003; Roberts, Brown et al.,1998;
Roberts et al., 2003; Roberts et al., 2004; Serrano et al., 2000; Ustin et al., 1998).
4.2.4. Other biochemical absorption features
Several other biochemicals in vegetation create absorption features across the
400-2500 nm spectrum measured by hyperspectral sensors (Curran, 1989; Elvidge,
1990). Cellulose is a prevalent structural chemical that has NIR and SWIR
combination-band and overtone absorptions at 1.22, 1.48, 1.93, 2.10, 2.28, 2.34 and
2.48 µm (Elvidge, 1990). Lignin is another important chemical in plants, with
combination-bands and overtones at 1.45, 1.68, 1.93, 2.05-2.14, 2.27, 2.33, 2.38 and
2.50 µm. Other biochemicals include proteins, xylan, tannins and waxes (Curran,
1989; Elvidge, 1990). Typically, absorptions by these biochemicals are not isolated
but overlap due to broadening by multiple scattering and overlap with absorptions at
similar wavelengths (Curran, 1989). Green leaves contain 40-80% water by weight,
138
and thus water absorption tends to mask the more subtle SWIR absorptions caused
by other chemicals (Curran, 1989; Elvridge, 1990; Kokaly & Clark, 1999). Remote
sensing of vegetation has not focused on SWIR features for estimating biochemical
concentrations or discriminating species, largely because these regions are sampled
with only one to a few bands in multispectral sensors, if at all. Imaging
spectroscopy provides an unprecedented opportunity to exploit the information
content of SWIR features (Curran, 1989; Elvidge, 1990). In one novel study, bands
normalized to the depth or area of SWIR absorption features were found to be highly
correlated to biochemicals such as nitrogen, lignin, and cellulose in dry leaves
(Kokaly & Clark, 1999). Using similar techniques, Curran et al. (2001) found that
VIS, NIR and SWIR features were useful for predicting chlorophyll, water, nitrogen,
cellulose, lignin and other plant chemicals.
4.2.3. Spectral mixture analysis
Another analytical technique is spectral mixture analysis (SMA), which models
reflectance spectra as a linear combination of two or more dominant or “pure”
spectral components, or endmembers (reviewed in Keshava & Mustard, 2002).
SMA outputs are the fractional abundance of each endmember and a root-meansquare error (RMSE) model fit.
Endmembers are selected from a library of field,
laboratory, image-extracted or “virtual” spectra.
Typically one endmember is
designated as shade, which models variation in illumination.
Other bright
endmember(s) represent major spectral components—typically green photosynthetic
vegetation (GV), non-photosynthetic vegetation (NPV: litter, bark, branches) and
139
soil in vegetation mapping applications (Roberts et al., 1993). Although SMA can
be performed on multispectral data, the model solution is strengthened with
hyperspectral data because there are more high-signal bands, which may help
distinguish spectral mixtures based on the unique spectral shape (i.e., absorption
features) of their endmembers.
4.3. Methods
4.3.1. Canopy-emergent trees
Analyses focused on the classification of canopy emergent individuals of seven
tree species (Table 4.2). Chapter 3 provides details about the species and number of
study trees (Table 3.1) and how their ITC polygons were digitized on the HYDICE
hyperspectral imagery. As explained in Chapter 3, some overstory tree species are
completely deciduous, generally beginning in the first dry season, while others are
evergreen and continuously flush small amounts of leaves throughout the year
(Table 4.2). Hyperspectral imagery was acquired on March 30, 1998 (Chapter 1), at
the end of the first dry season, and all study trees were expected to have high mature
leaf cover except DIPA and LEAM (Table 4.2). However, BAEL and HYME had
relatively fine compound leaves and so their mature leaf cover was expected to have
a lower LAI relative to the broadleaf species CEPE, HYAL and TEOB. I thus
expected LAI for the study species to be low for DIPA and LEAM, moderate for
BAEL and HYME, and high for CEPE, HYAL and TEOB.
140
4.3.2. Bark spectra
Laboratory leaf and bark spectra were acquired for tissue-scale assessment of
metrics and SMA. I measured leaf spectra in a laboratory at LSBS (Fig. 4.1a;
detailed methods in Chapter 3). Bark specimens from the seven study species (Table
4.2), as well as from Cedrela odorata, Cordia alliodora, Laetia procera,
Pentaclethra macroloba, and Simarouba amara were sampled in the station vicinity
in April, 2004. Most samples were from the branches of young trees that were
retrieved with a shotgun, pole clipper, rope, or saw, while some DIPA trunk bark
was sampled by climbing emergent trees. Bark specimens had an average width of 6
cm and length of 20 cm. Fresh specimens were stored in sealed plastic bags in a
refrigerator at 6° C for 1.5 months and their spectral properties were measured with
an ASD full-range spectrometer (Analytical Spectral Devices, Boulder, CO, USA)
sensor with an 8° fore-optic. Specimens were blotted with a paper towel to remove
surface water and were placed in a 5%-reflective box and illuminated with a 250 W
halogen bulb. The fore-optic was positioned 10 cm at nadir above the box center,
yielding a 1.4-cm sensor field of view (FOV). To capture spectral variation from
each specimen, the sample orientation and position relative to the sensor were varied
with each radiance measurement (the sample was moved, the sensor remained
stationary). Radiance from a white Spectralon® panel (Labsphere, North Sutton,
NH, USA) placed in the box center was used as a standard to convert bark radiance
measurements to percent reflectance. A final bark reflectance spectrum was an
average of fifteen reflectance spectra from the specimen. The ASD spectrometer had
1-nm spectral sampling covering 350 to 2500 nm. I convolved leaf and bark
141
laboratory spectra to HYDICE band center positions (161 bands) using full-width,
half-maximum information for each HYDICE band.
For the seven study species, the bark library (Fig. 4.1b) contained 66 spectra: 15
BAEL, 9 CEPE, 10 DIPA, 8 HYME, 5 HYAL, 12 LEAM, and 7 TEOB. The leaf
spectral library contained 152 spectra: 16 BAEL, 15 CEPE, 30 DIPA, 23 HYME, 30
HYAL, 14 LEAM, and 24 TEOB. Spectral metrics were calculated from these bark
and leaf spectra. For spectral mixture analysis, I sought a more exhaustive bark
spectral library and included spectra from the study species, 12 spectra from five
other species (listed above), as well as 24 field spectra. Bark field spectra came
from tree trunks, which were measured at LSBS at 9:30-10:00 am local time in
August, 2002. The ASD FieldSpec sensor with 8° fore-optic was held about 1 m
from the trunk and measured radiance was converted to reflectance using an in situ
white Spectralon® calibration panel in full sun.
4.3.3. Calculation of spectral metrics
4.3.3.1. Narrowband, ratio-based indices
Narrowband vegetation and water-absorption indices were calculated using
formulas in Table 4.1 applied to reflectance spectra at tissue, pixel and crown scales
for HYDICE bands. The HYDICE bands chosen for the indices were closest to
those in the formulas presented in the literature.
142
4.3.3.2. Derivative-based metrics
Derivative analysis was used to measure the wavelength position and magnitude
of the blue edge (BE), green peak (GP), yellow edge (YE), red well (RW), red edge
(RE), NIR water absorption edges (NE1 & NE2), and the SWIR1 edge (SE) (Table
4.3; Fig. 4.1a) for HYDICE bands. Derivatives for these features were calculated
using a polynomial-fitting technique (Pu, Gong et al., 2003). Reflectance data were
retrieved for all bands in the region covering a spectral feature (e.g., red edge) and a
function was then fit to the wavelength and reflectance values according to a
polynomial equation and a least-squares minimization:
n
ρ = α 0 + ∑ α i λi
(1)
i =1
where ρ represents the HYDICE reflectance values at λ wavelength bands within the
absorption feature (Table 4.3), n is the polynomial order, and α represents the fitted
polynomial constants. A continuous derivative spectrum for each feature was then
calculated using the fitted polynomial function and used to estimate the wavelength
of the local minima or maxima for the edge features, corresponding to the inflection
point in the original reflectance domain. Similarly, the wavelength position at
which the derivative was zero identified the GP and RW features, which
corresponded to the peak and well in the reflectance domain, respectively. In
addition, the magnitude of the derivative (i.e., reflectance slope) was retrieved at
edge inflection wavelengths, while reflectance was retrieved at the GP and RW
wavelengths (Gong et al., 2002).
143
Green vegetation metrics based on derivative area were calculated following
methods outlined in Elvidge & Chen (1995) and Gong et al. (2002). First-order
derivatives of spectra were calculated using a 3-point, Lagrangian interpolation and
then the total area under the derivatives was calculated for the blue edge (BE-DArea;
491.0 -532.1 nm; SDB in Gong et al., 2002), yellow edge (YE-DArea; 549.6-581.9
nm; SDY in Gong et al., 2002) and the red well-red edge features (RWE-DArea;
626.7 to 798.2 nm; 1DL_DGVI in Elvidge & Chen, 1995). For the RWE feature, the
first-derivative area was also calculated after normalizing the first-derivative spectra
to the value at 626.7 nm (RWE-DNArea; 1DZ_DGVI in Elvidge & Chen, 1995). A
second-order derivative was then calculated from normalized first-order spectra and
the area calculated (RWE-2DNArea; 2DZ_DGVI in Elvidge & Chen, 1995).
4.3.3.3. Absorption-based metrics
Using methods adapted from Pu, Ge et al., 2003, I calculated the maximum depth
(D), wavelength position of maximum depth (λ), width (W), maximum depth x
width area (A1) and asymmetry (As) for photosynthetic pigment (blue and red),
water (NIR) and other biochemical (SWIR) absorption features (Table 4.3).
Absorption feature area (A2) was also calculated as the sum of the depths measured
at each band within the feature’s range (Table 4.3). I calculated EWT using a BeerLambert light-extinction model, with parameters fit to reflectance data between 865
and 1065 nm using a non-linear least-squares minimization routine (method detailed
in Dennison et al., 2003). The spectrum of water absorption coefficients required by
144
this technique was interpolated to HYDICE wavelengths from data in Palmer and
Williams (1974).
4.3.3.4. Spectral mixture analysis (SMA) fractions
Pixel- and crown-scale spectra were unmixed using a three-endmember model
composed of green vegetation (GV), non-photosynthetic vegetation (NPV) and
photometric shade. The GV endmember was selected by plotting crown pixels in a
red versus NIR scatter plot. Thirteen pixels with relatively low red and high NIR
were averaged to form one GV image endmember, and nine pixels with high red and
low NIR were averaged to form one NPV image endmember. A GV image
endmember was used in SMA because it contained pronounced absorption features
due to multiple-scattering within the canopy whereas laboratory-measured, leaf-scale
spectra had weaker absorption features. The NPV image spectrum was not “pure,”
but rather it was a mixture of spectral properties from NPV (e.g., bark) and GV (e.g.,
tree leaves, canopy epiphytes, moss) measured within the sensor’s instantaneous
field of view (IFOV). A library of 102 field and laboratory spectra was used to
select a relatively pure NPV endmember spectrum. I chose the NPV endmember
from the library using the criteria that it yielded a low root-mean-square error
(RMSE) and physically-reasonable fractions when unmixing the image NPV
endmember with the image GV endmember (Roberts, Batista et al., 1998). The final
NPV reference endmember was from a lichen-covered tree trunk acquired from a
suspension bridge at LSBS.
145
4.3.4. Tree species classification techniques
I explored the binary decision tree (DT) classifier (Breiman et al., 1984; Friedl &
Brodley, 1997; Pal & Mather, 2003; Roberts et al., 2002) for ITC classification.
Training data are used to build a DT framework composed of a sequence of binary
tests (nodes) applied to the predictor variables (i.e., spectral metrics) that assign
pixels to their final class at the end of the decision tree (terminal nodes). Decisiontree classifiers have an advantage over more traditional classifiers like the maximum
likelihood classifier, in that they make no assumptions about the data distributions
(i.e., they are non-parametric), can handle data with different scales, and can adapt to
noise and non-linear relationships inherent in remote sensing data (Friedl & Brodley,
1997). Also, DT rules are explicitly defined and are often interpretable in terms of
meaningful, physical factors that are important in discriminating classes. In this
chapter, I used the “Tree” package in the R statistical environment for DT analysis
(R Development Core Team, 2004; Tree v1.0-18, R v2.0). Parameters for growing
trees were “mincut” of 5, “minsize” of 10, “mindev” of 0.001, and “deviance” as the
criteria for splitting data into homogenous sets.
Similar to methods in Chapter 3, two ITC species classification schemes were
analyzed: crown-scale and pixel-majority. The crown-scale approach labeled ITC
species using classified crown-scale spectra. Crown-scale reflectance spectra were
first calculated from the average of all within-crown pixels (shaded and sunlit)
which provided one spectrum per crown. Spectral metrics (e.g., NDVI, red-edge
position, etc.) were calculated from crown-scale reflectance spectra. Crown-scale
DT used a leave-one-out cross-validation approach.
146
An ITC was classified by
removing the crown from the analysis and a DT was then built using the remaining
crowns as training data (n=213).
Initial analyses revealed that classification
accuracy was higher when DTs were constructed using a balanced training data set
of equal class sample size. Since each species class had a different number of
crowns, balanced training data were acquired with a random sample with
replacement of 50 crowns per class. No DT pruning was performed because low
sample sizes kept the decision rules relatively simple. The DT was then used to
classify the ITC that had been held out from the training set. This process of training
and classification was repeated 50 times per ITC and the final ITC label was chosen
using the majority class of the 50 DTs.
Pixel-majority classification was also performed with a leave-one-out cross
validation. In this case, each ITC was iteratively withdrawn from analysis and 2,000
pixels per species were randomly-selected without replacement from the remaining
ITCs (n=213). For each class, the sampled pixels were then divided evenly into
training and pruning data sets (i.e., 1,000 pixels each).
Decision trees were
automatically pruned with the R Prune.Tree routine, which constructs a nested
sequence of decision sub-trees by snipping off the least important splits based on a
cost-complexity parameter. Pruning data were dropped down each decision sub-tree
and deviances were calculated between the predicted and response classes. The
decision sub-tree producing the minimum deviance was chosen as the pruned
decision tree. Within-crown pixels from the withheld ITC were then classified using
the pruned decision tree and the ITC label was chosen using the majority class of the
within-crown classified pixels. For each ITC withheld from training, the process of
147
decision tree construction, pruning and ITC majority-class labeling was done 50
times—with a new set of mutually-exclusive, 1,000 training and 1,000 pruning
pixels per class randomly sampled each time. The final ITC label was then assigned
using the majority class of the 50 crown labels for each ITC.
4.4. Results
4.4.1. Differences in spectral metrics among species
A summary of the spectral metrics calculated in this chapter is presented in Table
4.4. Leaf-, bark-, pixel- and crown-scale summary statistics and tests are presented
for ratio-based (Appendix 2.1), derivative-based (Appendix 2.2), and absorptionbased (Appendix 2.3) metrics, while spectral mixture analysis fractions are presented
at pixel and crown scales (Appendix 2.4). I tested mean differences in spectral
metrics among species with an ANOVA and I used a Tukey’s Post Hoc Honestly
Significant Different procedure (Zar, 1996) for multiple comparison pair-wise tests
of mean differences between two species. The relative statistical strength of metrics
for distinguishing species was ranked primarily by the magnitude of the ANOVA F
statistic and secondarily by the number of significant pair-wise differences (21 total).
Out of 73 metrics calculated from laboratory leaf and bark spectra, 61 and 55 had
significant differences in mean values among species, respectively (Appendix 2.12.3). At pixel and crown scales, 76 and 69 out of 77 of the spectral metrics showed
significant differences in mean response among species (Appendix 2.1-2.4).
148
4.4.2. Photosynthetic pigment absorption metrics
In laboratory leaf spectra, 94% of photosynthesis metrics showed significant
differences among species. Metrics covering the red-edge feature—RVSI and REλ—were the most important of these metrics at this scale (Table 4.5, leaves). The
blue absorption wavelength (Blue-λ), depth (Blue-D), and area (Blue-A2, Blue-A1)
were the next high-ranking photosynthesis metrics (ranks 16, 19, 20, and 21,
respectively). For laboratory bark spectra, 65% of the photosynthesis metrics had
significant mean differences. As with leaf spectra, top-ranking metrics for bark were
also concentrated in the red-edge region: RWE-DArea, RE-Mag and RVSI (Table
4.5, bark). Blue-edge derivative metrics, BE-DArea and BE-Mag, were more
important than blue absorption-based metrics in bark spectra.
At pixel scales, the red absorption wavelength (Red-λ) was the only
photosynthesis metric that did not have a significant mean difference among species.
The metrics SR, Red-A1, Red-A2 and ARVI were among the top-ten metrics for
species discrimination at pixel scales (Table 4.5, pixels). The deciduous species
DIPA and LEAM had the lowest SR values, while high-LAI species (CEPE, HYAL,
TEOB) had the highest SR values (Fig. 4.2). Other highly-significant metrics
included Red-As, Red-W, NDVI, PRI and Red-D (ranks 11, 12, 13, 14, and 16,
respectively).
Half of the top-ranking metrics at crown scales were photosynthesis metrics
(Table 4.5, crowns). In particular, SR and YE-DArea had the highest ranks. The red
absorption feature area, asymmetry and width were the next set of important metrics
149
for crown-scale species discrimination, followed by the vegetation indices ARVI,
SAVI and EVI (ranks 14, 16, and 18, respectively).
4.4.3. Water absorption metrics
In leaf spectra, seven out of the top-ten metrics encompassed some aspect of the
NIR water absorption features (Table 4.5, leaves), which had mean locations at 994
and 1175 nm (NIR1-λ and NIR2-λ, Appendix 2.3). Important metrics describing
these features were the area, depth and the down-sloping edge inflection magnitude
(Table 4.5, leaves). All water metrics at leaf scales had significant differences
among species except for WE- λ, and leaves were more strongly separated with
NIR2 versus NIR1 absorption-based metrics. In contrast to leaves, there were only
two top-ranking NIR water absorption metrics for separating bark spectra (Table 4.5,
bark), although all water metrics had significant mean differences. The two indices,
NDWI and WBI, were more important for distinguishing bark spectra than the other
types of water metrics.
At pixel scales, NDWI was the top-ranking metric, followed by NIR1 area
metrics (Table 4.5, pixels; Fig. 4.2). Greater pixel-level water absorption was
observed in high-LAI species, especially the broad-leaved species CEPE, HYAL and
TEOB (Fig. 4.2; NDWI, NIR1-A1). The importance of water metrics decreased
substantially at crown scales. The NDWI was ranked 9 (Table 4.5, crowns; Fig.
4.2), but all other water metrics were ranked 38 or greater out of all metrics (77
total).
150
4.4.4. Shortwave-infrared biochemical absorption metrics
Only 8 out of 20 SWIR biochemical absorption metrics had significant mean
differences at leaf scales. The SWIR3 region was the best of the three absorption
features, while SWIR1 and SWIR2 absorption features were either not detected or
had minor significance in both leaf and bark spectra (Appendix 2.3). For leaves,
SWIR3 area (A1 and A2), depth, asymmetry and width had ranks 9, 11, 13, 17 and
18 out of all 73 metrics. All of these absorption metrics had top-ten ranks for
distinguishing species with bark spectra (Table 4.5; bark).
In contrast to leaves and bark, the SWIR1 and SWIR2 metrics had high rankings
at the pixel and crown scales. The SWIR1-As and SWIR1-W metrics had ranks 8
and 9 out of 77 metrics at pixel scales (Table 4.5, pixels). The SWIR2 absorption
feature area (A1 and A2), SWIR1-As and SWIR1-W had the strongest statistical
separation at crown scales (Table 4.5; crowns). At both pixel and crown scales, the
high-LAI species CEPE, HYAL and TEOB had low SWIR2-A1 (Fig. 4.2). In
contrast, low LAI species (DIPA, LEAM) to moderate LAI species (BAEL, HYME)
had comparatively high SWIR2-A1 values.
4.4.5. Pixel- and crown-scale spectral mixture analysis fractions
Mean values of SMA fractions from ITCs were roughly 40% GV, 15% NPV and
44% shade (+ 1% error; Appendix 2.4) at both pixel and crown scales. Canopies of
deciduous DIPA and LEAM had relatively high fractions of NPV and low fractions
of GV (Figs. 4.2 & 4.3). In contrast, leaf-on broad-leaf species CEPE, HYAL and
TEOB had relatively high fractions of GV and low fractions of NPV. Figure 4.3
151
reveals that there was variability in the crown-level proportions of GV, NPV and
shade among individuals of the same species (i.e., conspecific variability). For
example, BAEL and HYME (fine compound leaves, moderate LAI) had some
individuals with NPV and GV fractions similar to low-LAI crowns of DIPA and
LEAM, while other individuals of BAEL and HYME had fractions that resembled
the high-LAI crowns of broadleaf species HYAL and TEOB. As an index of nonphotosynthetic vegetation, NPV fractions were expected to correlate positively with
bark spectral properties and negatively with leaf properties (e.g., chlorophyll
absorption). This was indeed the case: NPV had negative linear correlations with
SR, Red-A1, Red-A2 and ARVI photosynthesis metrics (i = -0.75 to -0.78) and
NDWI and NIR1-A1 water absorption metrics (r = -0.71 and -0.32, respectively).
To quantify the influence of illumination variation on spectral metrics, I
calculated the Pearson’s linear correlation between each metric and the SMA shade
fraction. Pixel-scale metrics had correlations (r) ranging from -0.85 (BE-DArea) to
+0.60 (YE-DArea), with an average correlation of -0.09 (±0.30). Crown-scale
metrics had correlations (r) ranging from -0.63 (BE-DArea) to +0.30 (NIR2-A1),
with an average correlation of -0.02 (±0.23).
4.4.6. Decision-tree classification
4.4.6.1. Pixel-majority ITC classification
ITC classification accuracies were below 69% for DT pixel-majority classifiers
(Table 4.6). Including all 77 metrics produced the best overall accuracy (68.2%;
Table 4.6). From the analysis of all 77 metrics, the 10 top-ranked metrics were
152
selected according to the number of times the metric was found in decision trees
(Table 4.7; pixel scale). These 10 metrics were then applied in a separate DT
analysis. This approach produced slightly less overall accuracy than using all 77
metrics (Z = 0.27, not significant at α=0.05; Congalton, 1991). Of the groups of
spectral metrics considered, derivative-based metrics yielded the highest overall
accuracy. However, the 10 top-ranked metrics included indices, absorption-based,
derivative-based and SMA metrics (Table 4.7). In terms of the number of times a
metric appeared in decision trees, species were distinguished with a wide variety of
absorption features, including blue absorption (Blue-D), yellow edge (YE-DArea),
red-well (RW-λ), red-edge (RVSI), NIR water absorption (NDWI, NIR1-A1), and
SWIR biochemical absorption (SWIR1-W, SWIR2-A2). Even though they appeared
many times in DTs, the ten top-ranked metrics had average node depths between 3 to
5 tiers. A metric that consistently splits data at a primary (root) node would have an
average node depth of zero. Therefore, the most widely-used metrics were not
necessarily forming primary splits, but also played decisive roles closer to the
terminal nodes, where final species labels were determined. The lowest mean node
depth was only 3.5 (ARVI), which indicates that no single metric consistently split
the data at the root node.
4.4.6.2. Crown-scale ITC classification
Overall ITC classification accuracies using crown-scale metrics were slightly
higher than with the pixel-majority approach when considering absorption-based,
SMA fractions and all metrics (Table 4.6). The best crown-scale accuracy achieved
153
was 70.1% with all 77 metrics (Table 4.8). The selection of the 10 top-ranked
metrics (Table 4.7; Crown scale) slightly decreased overall accuracy relative to
using all 77 metrics (Table 4.6; Z = 0.58, not significant at α=0.05). Top-ranked
metrics were similar to those observed in pixel-scale DT analyses, except SWIR3-D
(depth) was more influential at crown scales (Table 4.7; Crown-scale). The SWIR2A1 metric had a mean node depth of 0.7, which is the lowest depth of all other
metrics and shows that the SWIR2 area was often the primary split separating
species.
An example decision tree using all 77 metrics showed that some individuals of
high-LAI species (Fig. 4.4 A) were separated from moderate- and low-LAI species
(Fig. 4.4 B & C) primarily with the SWIR2 area metric—lower SWIR2-A1, more
LAI (Fig. 4.2). The YE-DArea was negative in species with higher LAI (Fig. 4.2;
HYAL, HYME, TEOB) because derivative values were from a steeply down-sloping
edge in reflectance spectra. Low YE-DArea values were used by the decision tree to
separate moderate-LAI individuals (Fig. 4.4b) from individuals with lower LAI (Fig.
4.4c). Photosynthetic pigment, water and SWIR biochemical absorption features
distinguished individuals of BAEL, LEAM and DIPA (Fig. 4.4c). These species had
the most interspecific confusion (Table 4.8).
For comparison with previous research in Chapter 3, I also present the overall
accuracy achieved when using reflectance narrowband (hyperspectral) data in DTs
instead of spectral metrics. With crown-scale reflectance spectra, a linear
discriminant analysis (LDA) classifier had 92.1% overall accuracy using 30
optimally-selected bands and 61.2% accuracy with all 161 bands. I used the same
154
methods to select optimal spectral metrics for crown-scale LDA classification. The
best classification had an overall accuracy of 84.1% (DIPA User’s = 85.5%,
Producer’s = 87.7%) with the inclusion of 18 optimal metrics, 6% lower accuracy
than with 30 reflectance bands (Z=2.60, significant at α=0.05). With the DT
classification scheme used in this chapter, crown-scale overall accuracy was only
49.5% using either 30 or 161 bands (Table 4.6). Although the DT classifier was
20.6% more accurate when applied to spectral metrics over reflectance data, the
LDA classifier applied to reflectance data outperformed the best metric-based DT
classifier by 22.0% (Chapter 3; Z=6.17, significant at α=0.05).
4.5. Discussion
4.5.1. Crown structure and phenology detected by spectral mixture analysis
fractions
At pixel and crown scales, tropical tree spectra are mainly a mixture of GV
(e.g., leaves, green bark, epiphytes, lianas), NPV (e.g., non-photosynthetic bark,
fruits, flowers) and shade. The biophysical properties of the crown, such as leaf
density, angles and clumping, determine the fractions of these components that form
the mixed signal within an image pixel. Phenological variation in trees, such as leaf
aging and leaf drop, drought stress, and flowering, will also affect the multi-temporal
spectral signature of tree species and may help discriminate individual or
assemblages of plant species (Blackburn & Milton, 1995; Key et al., 2001). In
tropical forests, previous research has shown that individual deciduous trees with
155
low leaf cover have a clear spectral contrast with leaf-on trees in the visible and NIR
regions of the spectrum (Bohlman & Lashlee, in press).
As noted in Chapter 3, the DIPA and LEAM trees in our hyperspectral imagery
were mainly deciduous (i.e., low crown-level LAI) at the time of image acquisition.
Spectral mixture analysis fractions revealed that the DIPA and LEAM populations as
a whole had crowns composed of relatively high percentages of NPV materials
relative to other species (Fig. 4.3), which can be attributed to non-photosynthetic
branch and trunk tissues exposed to the sensor. BAEL trees also had relatively high
amounts of NPV. Although BAEL was expected to have high mature leaf cover
during image acquisition (O’Brien, 2001), I attribute high NPV in some BAEL
individuals to their architectural properties rather than deciduousness; BAEL trees
have compound leaves with fine leaflets and the overall architectural arrangement of
these leaves exposed conspicuous white bark when viewed from above by the sensor
(see Fig. 3.2a in Chapter 3).
ITCs with relatively low LAI, such as individual DIPA, were easily identified in
both the hyperspectral reflectance imagery (Fig. 4.5; natural color) and the images of
spectral metrics (Fig. 4.5; SMA fractions, NDWI, NDVI, SWIR2-A1). In the SMA
false-color image, ITCs with low-LAI were clearly seen as groups of pixels with
high NPV (red) in a matrix of high-GV canopy (green) and high-shade gaps (blue).
4.5.2. Species differences in photosynthetic-pigment absorption features
My analyses highlighted the importance of the red-edge for species
discrimination with leaf and bark spectra. These results are consistent with work by
156
Cochrane (2000), who also found this region useful for distinguishing tropical tree
species with their leaf spectra. The shape of the red edge is heavily influenced by
photon absorption and florescence from chlorophyll in tissues (Gates et al., 1965;
Zarco-Tejada et al., 2000).
Research has shown that high concentrations of
chlorophyll in leaves widens and deepens red-well absorption and shifts the red-edge
inflection wavelength (RE-λ) toward longer wavelengths (Blackburn, 1998).
The RVSI and RE-λ were the strongest metrics for discriminating leaf spectra.
The RVSI was designed to detect red-edge curve concavity or convexity relative to a
baseline (e.g., continuum removal).
The index was originally used to track
vegetation communities through time, not spectral properties of tissues. I found that
leaf-scale RVSI was positively correlated with RE-λ (r = +0.72), and so positive
RSVI values were related to a red-shift in leaf RE-λ.
In contrast to leaves, bark spectra had low to no chlorophyll and so there was
relatively low red-well absorption (Fig. 4.1b), increased red-well reflectance (RWRefl), decreased slope of the red-edge (RE-Mag), and a shift in the red-edge position
toward lower wavelengths—a blue-shift (Appendix 2.1 & 2.2).
There were
significant differences among species with bark red-edge metrics such as RE-Mag.
Some bark spectral variation in the visible to red-edge regions is likely related to a
strong ultraviolet-blue absorption wing from 400 to 900 nm created by lignin and
tannins (Elvidge, 1990), while overlapping red-well absorption may be caused by
algae, bryophytes, moss and lichen on the bark. I found that BAEL bark had the
highest mean RE-Mag, while TEOB bark had the lowest mean. BAEL had a lightbrown to red bark color, and no evidence of surface chlorophyll from other
157
organisms; these properties gave BAEL bark relatively high red-well reflectance and
a steep red-edge slope (Fig. 4.1b-BAEL). In contrast, TEOB bark samples were
darker brown with green mottling (providing evidence of chlorophyll in the tissues)
and their red-well reflectance was relatively low with a more shallow slope (Fig.
4.1b-TEOB). The RVSI of bark was not as strongly correlated with RE-λ (r =
+0.38) as with leaf spectra, but was mostly related to red-edge slope, RE-Mag (r
=+0.52).
The red absorption feature also had potential for discriminating species at pixel
and crown scales. The SR and red absorption area (Red-A1 and Red-A2) were
among the top-ranking variables in tests of mean differences among species (Table
4.5). A typical pixel from a high-LAI species, such as HYAL, revealed low red-well
reflectance due to chlorophyll absorption and high NIR reflectance due to multiple
scattering among transmittive leaves (Fig. 4.6); and consequently, the metrics SR,
NDVI and Red-A1 were relatively high. In contrast, a pixel from low-LAI DIPA
had a higher concentration of bark in the sensor IFOV, which translated into
relatively high red-well reflectance due to lower chlorophyll absorption and low NIR
reflectance due to less photon scattering within the crown (Fig. 4.6); and thus, the
metrics SR, NDVI and Red-A1 were relatively low.
I found that the ANOVA rankings (Table 4.5) were not necessarily predictive of
overall metric utility in DTs; for example, red-well wavelength position (RW-λ) was
a key metric in DTs at both pixel and crown scales (Table 4.7), but was ranked 43
and 51 out 77 metrics based on pixel and crown F statistics, respectively
(Appendices 2.1-2.4). This metric mainly separated BAEL from the other species
158
(Fig. 4.2, RW-λ). BAEL had the highest mean red-well reflectance relative to other
species, and this spectral shape tended to shift the red-well position toward smaller
wavelengths. The RW-λ was observed classifying a sub-set of BAEL in Figure 4.4
B, down at DT tier 3 (root tier is 0); and thus, RW-λ was important for lower-tier
decisions, even though overall mean species differences were weak (i.e., low
ANOVA F statistic).
4.5.3. Species differences in water absorption features
The NIR1 and NIR2 water absorption features dominated top-ranking metrics for
leaf spectra. Internal leaf structure, such as the air-cell wall interfaces in the spongy
mesophyll, heavily scatters NIR photons and increases the expression of absorption
features (Gates et al., 1965; Gausman, 1985). Species differences in water
absorption depth, width and area may thus result from a combination of factors that
affect internal leaf photon scattering and absorption, such as leaf thickness, water
content and the distribution of air-cell wall interfaces.
Bark water absorption metrics were more variable than those for leaves. For
example, species standard deviations of NDWI ranged from 0.004 (DIPA) to 0.021
(HYME) for leaves while 0.032 (CEPE) to 0.118 (DIPA) for bark. This variability
in bark spectra made species less distinct relative to leaves (in terms of F statistic
and significant pairs). Bark laboratory specimens were cleared of large bryophytes
and mosses, which when wet would further increase water absorption variability
within species. In my laboratory data, much of the within-species bark variability
comes from spectral differences between branch and trunk bark.
159
One striking
example was DIPA. Branches sampled from young DIPA trees had smooth, green
bark. Spectra from these branches had clear chlorophyll absorption features at 680
nm (Fig. 4.1c), and so they can not be considered as “pure” NPV. In contrast, trunk
bark samples from mature DIPA crowns were woody, rough, and brown (i.e., nonphotosynthetic). The green bark spectra had more notable NIR water absorption
features than the woody trunk bark (Fig. 4.1c). These data indicate that bark spectral
properties can vary considerably within a population of tropical tree species.
Williams (1991) found that first-year twig spectra for conifer and hardwood species
showed evidence of chlorophyll and NIR water absorption, and the depth of these
features decreased considerably in second-year twigs from the conifer species, which
were more woody. It thus appears that bark chlorophyll and NIR water absorption
features are correlated, likely because photosynthesis requires water.
Considering all species, there were significant (p≤0.01) differences in means of
NIR water absorption metrics between leaf and bark laboratory spectra. The ratiobased indices, WBI and NDWI, had significantly greater values in leaves than bark;
however, absorption-based metrics such as EWT and area (NIR1 or NIR2) were
significantly higher in bark relative to leaves. Bark NDWI tended to be negative
because NIR reflectance at 862 nm was low relative to the absorption feature at 1239
nm, while the opposite trend was found in leaves. This finding is similar to Gao
(1996), who found that the NDWI metric was negative in non-photosynthetic
materials (e.g., dead grass). The larger NIR water absorption area in bark is seen
clearly in Figure 4.1 when comparing BAEL, DIPA and HYAL mean spectra in bark
(a) to leaves (c). Large NIR water absorption features, especially in the 1200 nm
160
region, are unusual in dry bark spectra from temperate regions (Elvidge, 1990;
Roberts et al., 2004). However, my bark samples from tropical rain forest trees were
not dried and evidently contained a great deal of moisture, especially in green bark.
Bark water absorption not only manifests as broad NIR features but it also obscures
SWIR biochemical absorption features that are otherwise prominent in drier bark
material (Elvidge, 1990).
The depth, width and area of NIR water absorption features tended to increase
from the centimeter spatial scales of leaf and bark laboratory spectra to the 1.6-m
scale of HYDICE image pixels (Appendix 2.3). The depth of these features can be
partly attributed to poor radiometric calibration and signal-to-noise in the HYDICE
sensor (discussed in Chapter 3). However, the three-dimensional stacking of leaves
and branches within the crown, which increases the path length of photons and
increases their multiple-scattering and absorption, is expected to accentuate tissuelevel biochemical and biophysical properties (Asner, 1998; Curran, 1989; Roberts et
al., 2004). Water absorption metrics calculated at pixel and crown scales should
thus respond to the combination of water content in GV and NPV tissues and the
structural distribution of these tissues within the crown (Dennison et al., 2003; Gao,
1996; Roberts et al., 2004). The water absorption signal should be greater with more
leaf area relative to branch area because leaves are more transmittive, thereby
permitting more photon scattering. This effect can be observed in Figure 4.6, where
a leaf-off DIPA pixel spectrum had lower NDWI and NIR1 and NIR2 water
absorption area relative to a leaf-on HYAL pixel spectrum. Furthermore, the pixels
within DIPA crowns formed dark ITCs in the NDWI image (Fig. 4.5).
161
In terms of the F statistic and number of significant paired comparisons, species
differences in NIR water absorption metrics were greater at pixel scales than at leaf
scales. This finding suggests that pixel-scale biophysical differences among species
affect NIR water absorption in a way that amplifies species separability. This
scaling effect may explain why several NIR water absorption metrics were among
the primary variables for separating species in DTs at both pixel and crown scales
(Table 4.7). Also, results indicate that both NIR1 (e.g., NIR1-A1, WBI) and NIR2
(e.g., NDWI) absorption features were equally important variables for species
classification.
4.5.4. Species differences in shortwave infrared absorption features
There were also strong differences among species in metrics covering the
SWIR3 absorption feature in both leaf and bark spectra. The SWIR3 absorption
feature results from a combination of overlapping absorptions by protein, nitrogen,
starch, lignin, cellulose, and sugars (Curran, 1989; Elvidge, 1990). In HYDICE
data, the deepest part of the feature had a mean wavelength of 2294 nm and 2308 nm
for bark and leaves, respectively (SWIR3-λ, Appendix 2.3), which is a wavelength
region predominantly associated with absorption from N-H and C=O bond stretching
within the amino acids of proteins (Osborne & Fearn, 1986). As mentioned, the tails
of strong water absorption features in the SWIR region centered at 1400 and 1940
nm can mask the expression of more subtle SWIR absorption features in green
vegetation (Curran, 1989; Elvridge, 1990; Kokaly & Clark, 1999). There were
+0.20 and -0.38 correlations between NIR2-A1 and SWIR3-A1 in leaf and bark
162
spectra, respectively.
If the NIR2 absorption feature area is considered as an
indicator of tissue water concentration, it appears that absorptions by bark
biochemical constituents (e.g., lignin, cellulose, proteins) are negatively associated
with water content. It is unclear why there is a weak positive correlation between
leaf water absorption and SWIR3 biochemical absorption.
Biochemical assays
acquired at the time of spectral measurement would be necessary to establish
conclusive links between species biochemistry and observed reflectance patterns
(e.g., Kokaly & Clark, 1999).
Although the SWIR3 absorption feature had strong statistical separation among
species bark spectra, the importance of the feature diminished at pixel and crown
scales. This may be due to low signal-to-noise in the HYDICE sensor and low
reflectance in the SWIR3 region (Basedow et al., 1995; Chapter 3). In contrast to
SWIR3, the SWIR1 and SWIR2 absorption features were detected in only 25% and
7% of laboratory leaf spectra, respectively, and in 70% and 15% of bark spectra,
respectively. However, metrics incorporating the area, asymmetry and width of
SWIR1 and SWIR2 features were among the top-ranked discriminatory metrics at
pixel and crown scales. The SWIR1 feature is associated with lignin (primarily),
starch, protein and nitrogen absorption and the SWIR2 feature is associated with
protein and nitrogen (primarily), starch and cellulose absorption (Curran, 1989).
One explanation why SWIR1 and SWIR2 metrics were important for species
discrimination in pixel and crown spectra is that overlapping absorptions broadened
due to multiple scattering within the crown, thereby permitting feature detection in
pixel and crown-scale spectra. This SWIR absorption broadening should increase as
163
the fraction of bark tissues increases and as the fraction of leaves decreases. A lower
fraction of GV should correlate with lower SWIR water absorption and allow the
detection of bark SWIR1 and SWIR2 features. At crown scales, SWIR1-A2 and
SWIR2-A2 had positive correlations with NPV (r = +0.65 and +0.74, respectively)
and negative correlations with GV (r = -0.70 and -0.78, respectively), suggesting
that crowns with lower LAI and more exposed NPV had more strongly expressed
SWIR1 and SWIR2 features. Furthermore, low- to moderate-LAI DIPA, LEAM and
BAEL had the highest mean SWIR2 area, while high-LAI CEPE, HYAL and TEOB
had the lowest mean area (Fig. 4.2). This pattern is illustrated in Figure 4.6, where a
typical DIPA pixel spectrum had low SWIR1 and SWIR2 area relative to a HYAL
pixel. DIPA pixels formed bright ITCs in the SWIR2-A1 image (Fig. 4.5). This
image also reveals a pattern of noise in SWIR2, seen as diagonal stripes. The
absorption feature may have been more important at crown scales because averaging
many within-crown pixels will reduce spurious spectral variability due to sensor
noise. This reduction in variance can be seen as lower standard deviations at crown
scales relative to pixel scales in Figure 4.2 (SWIR2-A1).
4.5.5. Individual tree crown classification with decision trees
Despite the many statistically significant differences detected among species in
the various spectral metrics, overall ITC classification accuracies were no greater
than 71% using pixel-scale or crown-scale spectra. Some derivative-based and
absorption-based metrics had null values when absorption features could not be
detected, such as when calculating SWIR2 metrics when they were masked by water
164
absorption. These null values were not necessarily random values, and so they
should be included in the classifier. An advantage of the DT classifier was that it
could form splits on null values (coded as very negative numbers), whereas null
values are excluded in a traditional classifier such as maximum likelihood. Overall
pixel-majority DT classification accuracies were up to 12% lower when null values
were excluded from the analysis, indicating that there was an advantage to using null
values in DTs.
The crown-scale DT classifier had better performance than the pixel-majority
technique (Table 4.6). There are several reasons to explain higher crown-scale
accuracy. Crown-scale DTs had fewer nodes than pixel-scale DTs used in pixelmajority classification because there was lower variability in crown-scale spectra
(Appendix 2.1-2.4; Fig. 4.2), as they were averages of pixel-scale spectra. Despite
pruning, pixel-scale DTs still had complex structures that made class decision rules
more unstable. Furthermore, important crown-scale metrics changed less often in
DT depth (i.e., low average node depth) relative to pixel-scale metrics. Again, lower
crown-scale variability permitted more stable structural relationships among metrics
in separating species, thereby leading to greater ITC classification accuracy.
The best crown-scale accuracy was achieved by including all 77 metrics in
decision trees. The top 10 metrics used in decision rules included 4 absorptionbased metrics, 2 derivative-based metrics, 3 indices, and 1 SMA fraction (Table 4.7).
Red, NIR and SWIR features were all characterized by these top-ranking metrics. It
thus appears that a wide array of information extraction techniques applied to the
165
whole 400-2500 nm spectrum should be considered when classifying tree species
from hyperspectral imagery.
The crown-scale DT classification using 77 metrics had an overall accuracy of
70.1%. The 80.2% Producer’s and 79.3% User’s accuracies for DIPA (Table 4.8)
are also encouraging for mapping this species of high conservation value (see
Chapter 3). Confusion of DIPA with BAEL and LEAM (Table 4.8) was likely
because these species had low crown LAI at the time of image acquisition, which
would result in similar fractions of GV and NPV tissues and comparable photon
scattering environments.
4.5.6. Comparison with past research and recommendations for future research
As found in Chapter 3, the analysis of crown-scale data provided better
classification accuracy than using pixel-scale data. Both DT and LDA classifiers
thus indicated that crown-scale hyperspectral data may be sufficient for tropical rain
forest species classification at the canopy scale. Airborne hyperspectral sensors,
such as AVIRIS and HyMap, are typically flown with altitudes that provide 4- to 20m pixels. Since canopy-level tree crowns in old-growth forest at La Selva generally
have 20-27 m diameters (Chapter 2), a single crown could contain 1 to < 40 pixels,
depending on crown size and sensor resolution. By delineating ITCs and averaging
their 1.6-m pixels, crown-scale spectra in this research had minimal spectral mixing
with neighboring vegetation and soil. These factors would be expected to diminish
species signal and subsequent separability as pixel size increases.
166
Consistent with my results in this chapter, previous analyses found that
optimally-selected bands using LDA were located across the VIS, NIR, and SWIR
regions (Chapter 3). Several of the 20 most important reflectance bands were found
in SWIR1, SWIR2, and SWIR3 absorption features; and thus, both DT and LDA
classification schemes identified the SWIR region of the spectrum as important for
species discrimination. I recommend that future research use imaging spectrometers
that measure the full 400 to 2400 nm spectral range. Engineering improvements in
sensor signal-to-noise in the SWIR region should lead to a greater ability to
discriminate species with distinct phenology or structure.
Overall accuracy was 22% greater when using optimal reflectance bands and
LDA (Chapter 3) relative to spectral metrics and decision trees. LDA and
reflectance bands had 6% higher accuracy than LDA with spectral metrics. These
results indicate that future research should explore methods to exploit reflectance
spectra rather than developing new spectral metrics.
Why did LDA applied to optimal reflectance bands outperform the DT or LDA
classifier with spectral metrics? One explanation is that illumination variation drives
spectral differences among species, such as relatively low NIR reflectance from lowLAI DIPA relative to leaf-on HYAL. Illumination is best captured by reflectance
spectra as opposed to spectral metrics, whose brightness variation is minimized from
continuum removal, derivative analysis or band ratios. If illumination was the
dominant factor in distinguishing species, I would have expected the shade fraction
to be among the top metrics in ANOVA tests or in decision trees. Instead, the shade
fraction was ranked at 60 and 59 out of the 77 metrics in terms of F statistics for
167
pixel and crown scales, respectively, and it was ranked at 56 and 28 for pixelmajority and crown-scale classifications in terms of times found in decision trees.
Also, shade was not included as an optimal metric in LDA classification. These
results suggest that illumination variation is not a principal factor in discriminating
species. In general, spectral metrics were not well correlated with shade, indicating
that species were mostly separated according to biochemical absorptions rather than
illumination. However, of the top-ten metrics in decision trees, the NPV, RVSI and
YE-DArea metrics were correlated to shade at pixel and crown scales (Table 4.7; r =
-0.63 to +0.60). Illumination may thus account for some variation in metrics, but it
does not appear that it aids species discrimination.
My comparison of results with Chapter 3 indicates that LDA is a superior
classifier, regardless of using spectral metrics or reflectance bands. LDA considers
all predictor variables simultaneously to build a decision space using predictor
variable covariance. In contrast, the decision structure of a DT is built one variable
at time and decisions at terminal nodes are dependent upon those made at higher
nodes. The DT is thus more susceptible to misclassification errors when data
variability causes many different possible splits, as is the case when using highlyvariable spectra from ITCs in tropical rain forest. Despite better LDA performance,
I found decision trees useful for both ranking the importance of spectral metrics and
visualizing the decision rules in species discrimination.
168
4.6. Conclusions
A major goal of this chapter was to assess airborne hyperspectral imagery for
the species-level classification of individual tree crowns (ITCs) in a tropical rain
forest. The spectral metrics used to train the decision-tree (DT) classifier were
expected to capitalize on the detailed spectral information contained in the imagery.
The overall accuracy of ITC classification was 22% lower than when using the LDA
classifier (Chapter 3). Despite this weaker accuracy, the decision tree classifier was
a useful tool for ranking important spectral metrics and for visualizing the
hierarchical decision space used in the classification.
Statistical tests and classifier variable-selection analyses revealed that there were
detectable differences among species absorption features. Properties of the red
chlorophyll, NIR water and SWIR biochemical absorption features were all found to
differentiate species from centimeter scales of leaf and bark tissue to pixel or crown
scales measured by the airborne spectrometer. Tree structure and phenology at the
time of image acquisition were driving factors that influenced spectral separability of
species. For example, low leaf area crowns had low NIR water absorption due to
more exposed non-photosynthetic tissues and a relatively weak photon-scattering
environment. Several spectral metrics explored in this chapter, such as the NPV
fraction, NDWI, and SWIR2 area, revealed the immense potential of hyperspectral
metrics for mapping deciduous crowns in tropical rain forest canopies. The
capability to detect deciduous crowns could permit the study of species phenology
across large spatial extents and may prove a useful tool for monitoring changes in
tree phenology, stress or mortality due global climate change.
169
These results also have important implications for studies that seek to estimate
tropical forest biochemical and biophysical parameters over broad spatial scales
using spectral metrics derived from airborne or satellite imagery. Attempts have
been made to estimate aboveground biomass using a positive correlation with
metrics such as NDVI or other indices (Foody et al,. 2001; Thenkabail et al., 2004).
Although leaf-off deciduous trees may represent a substantial fraction of standing
live biomass, their unique crown spectra will influence metric values measured at
coarse spatial scales. For example, the NDVI value in a 30 x 30-m Landsat pixel
will go down with increasing fraction of deciduous crowns (e.g., Fig. 4.5-NDVI).
The fractional abundance of deciduous trees within a coarser scale pixel should be
considered in prediction models, especially if the imagery is acquired when a large
proportion of trees are leaf-off. This generally occurs in the dry season when there
is also the greatest opportunity to acquire cloud-free imagery.
170
Table 4.1. Formulas for narrow-band, ratio-based indices (ρ is reflectance at a
specific wavelength in nm). Wavelengths chosen are the closest HYDICE
wavelengths to the formulas in the cited literature.
Vegetation Indices
Simple Ratio
Tucker, 1979
ρ798
Jordan, 1969
SR = ——
Ρ679
Normalized Difference Vegetation Index
ρ798 - ρ679
NDVI = —————
ρ798 + ρ679
Soil-Adjusted Vegetation Index
1.5 * ρ798 - ρ679
SAVI = ———————
ρ798 + ρ679 + 0.5
Turner, 1979
Rouse et al., 1973
Huete, 1988
Photochemical Reflectance Index
ρ532 - ρ568
PRI = —————
ρ532 + ρ568
Gamon et al., 1997
Enhanced Vegetation Index
ρ798 - ρ679
EVI = —————————————
1 + ρ798 + 6 * ρ679 – 7.5 * ρ482
Huete et al., 2002
Atmospherically Resistant Vegetation Index
ρ798 - 2 * ρ679 + ρ482
ARVI = ——————————
ρ798+ 2 * ρ679 - ρ482
Kaufman & Tanre, 1992
Red-Edge Vegetation Stress Index
ρ719 + ρ752
RVSI = ————— – ρ730
2
Merton, 1998
171
Table 4.1. (continued).
Liquid Water Content Indices
Water Band Index
ρ902
WBI = ——
ρ973
Peñuelas et al.,1997b
Normalized Difference Water Index
ρ862 - ρ1239
NDWI = ——————
ρ862 + ρ1239
Gao, 1996
172
Table 4.2. Study tree species attributes (Adapted from O’Brien, 2001 and Frankie
et al., 1974). Leaf cover is for late-March to early-April, and is what would be
expected for the majority of individuals for each species based on available
literature data and personal field observations.
Tree species
Code
Leaf
Leaf
3/30
[family or sub-family]
Phenology Exchange
Leaf
Functional
Timing
Cover
Group
Balizia elegans
BAEL Deciduous
Annual
High
(Ducke) Barneby & Grimes
[Mimosoideae]
Ceiba pentandra
CEPE Deciduous
Annual
High
Gaertn.
[Bombacaceae]
Dipteryx panamensis
DIPA Deciduous
Annual
Low
(Pittier) Record & Mell
[Papilionoideae]
Hymenolobium mesoamericanum HYME Deciduous Sub-annual
High
Lima
[Papilionoideae]
Hyeronima alchorneoides
HYAL Evergreen Continuous
High
Allemão
[Euphorbiaceae]
Lecythis ampla
LEAM Deciduous
Annual
Low
Miers
[Lecythidaceae]
Terminalia oblonga
TEOB Evergreen Continuous
High
(Ruiz & Pav.) Steud.
[Combretaceae]
173
174
Table 4.3. Definition of spectral features analyzed (See Fig. 4.1a). Minimum and maximum
wavelengths are for HYDICE band centers and the number of bands within the range is
controlled by HYDICE band spacing.
Spectral Feature
Abbrev.
Min. λ
Max. λ
No.
Polynomial
(nm)
(nm)
Bands
Order
Derivative-based analyses (polynomial fitting)
Blue Edge (up-sloping)
BE
491.0
532.1
9
5
Green Peak
GP
532.1
581.9
9
5
Yellow Edge (down-sloping)
YE
549.6
581.9
6
5
Red Well
RW
643.3
698.7
7
5
Red Edge (up-sloping)
RE
679.3
751.5
8
5
NIR Water1 Edge (up-sloping)
NE1
958.2
1075.3
9
3
NIR Water2 Edge (down-sloping)
NE2
1105.1
1164.9
5
3
SWIR Edge (up-sloping)
SE
1494.5
1638.0
12
5
Absorption-based analyses
Blue Absorption
Blue
460.5
537.7
17
n/a
Red Absorption
Red
588.9
751.5
19
n/a
NIR Water Absorption 1
NIR1
902.0
1075.3
13
n/a
NIR Water Absorption 2
NIR2
1105.1
1253.9
11
n/a
SWIR Absorption 1
SWIR1
1650.5
1771.9
11
n/a
SWIR Absorption 2
SWIR2
2045.8
2221.6
19
n/a
SWIR Absorption 3
SWIR3
2240.2
2365.9
15
n/a
174
Table 4.4. Summary of hyperspectral metrics organized by methods (in
bold) and spectral region and dominant absorption feature (in italics).
Indices
Absorption-based
Derivative-based
SMA
Visible – Photosynthetic pigments
SR
Blue-λ,D,W,A1,A2,As
BE-λ,Mag
NDVI
Red-λ,D,W,A1,A2,As
GP-λ,Refl
SAVI
YE-λ,Mag
PRI
RW-λ,Refl
EVI
RE-λ,Mag
ARVI
BE-DArea
RVSI
YE-DArea
RWE-DArea
RWE-DNArea
RWE-2DNArea
Near infrared – Water
WBI
EWT
NE1-λ,Mag
NDWI
NIR1-λ,D,W,A1,A2,As
NE2-λ,Mag
NIR2-λ,D,W,A1,A2,As
Shortwave infrared – Other biochemicals
SWIR1-λ,D,W,A1,A2,As
SE-λ,Mag
SWIR2-λ,D,W,A1,A2,As
SWIR3-λ,D,W,A1,A2,As
Full-spectrum – All absorption features
GV
NPV
Shade
RMSE
λ = wavelength, Mag = derivative magnitude, Refl = percent reflectance,
DArea = area under 1st derivative, DNArea = area under normalized 1st
derivative, 2DNArea = area under 2nd derivative, D = depth, λ =
wavelength, W=width, A1 = area calculated using width and depth, A2 =
area calculated using tabulation, As = Asymmetry.
175
Table 4.5. Ranking of metrics based on F statistics and significant pair-wise
comparisons.
Rank
Bark
Leaves
Pixels
Crowns
1
SWIR3-λ
RVSI
NDWI
SWIR2-A2
2
SWIR3-D
NIR2-A1
NIR1-A1
SR
3
NDWI
NIR2-A2
SR
SWIR2-A1
4
SWIR3-As
NIR2-D
NIR1-A2
SWIR1-As
5
SWIR3-W
NE2-Mag
Red-A2
SWIR1-W
6
WBI
RE-λ
Red-A1
YE-DArea
7
RWE-DArea NIR1-A1
ARVI
Red-A2
8
RE-Mag
NIR1-A2
SWIR1-As
Red-As
9
SWIR3-A2
SWIR3-A1
SWIR1-W
NDWI
10
RVSI
NIR1-D
NPV
Red-W
176
177
Table 4.6. Accuracy of individual tree crown classification using leave-one-out crossvalidation and a decision tree classifier. Pixel-majority: pixel metrics were classified and then
aggregated into crown objects using a majority class rule. Crown-scale: crown-scale metrics
classified.
Pixel-majority
Crown-scale
Vars
Species
DIPA
Species
DIPA
Metrics
Overall
User/Producer
Overall
User/Producer
Accuracy
Accuracy
Accuracy
Accuracy
Indices
9
53.3
76.2 / 59.3
49.5
68.5 / 61.7
Derivative-based
21
65.4
89.3 / 61.7
56.5
78.3 / 66.7
Absorption-based
43
62.6
0.0 / 67.9
65.9
81.1 / 74.1
SMA fractions
4
44.4
78.6 / 27.2
46.7
64.7 / 40.7
All metrics
77
68.2
82.4 / 69.1
70.1
79.3 / 80.2
Top-ranked metricsa 10
67.3
75.0 / 70.4
67.3
75.0 / 70.4
Reflectance bandsb
30
55.6
87.5 / 34.6
49.5
64.7 / 54.3
Reflectance bands
161
54.2
89.7 / 32.1
49.5
66.7 / 61.7
a
Top-ranked metrics were those most common in trees when using all metrics.
b
Selected using step-wise linear discriminant analysis (Chapter 3).
177
Table 4.7. Ten top-ranked spectral metrics for decision trees
(DTs) at pixel and crown scales. Metrics are ranked by times
found in DTs. Shade corr. is the Pearson’s linear correlation of the
metric with the SMA shade fraction. There were a total of 10,700
DTs built at each scale of analysis, and some metrics appeared
more than once in DTs.
Pixel scale
Rank
Metric
Times in DTs
Shade Corr.
1
RW-λ
95961
-0.02
2
NDWI
81305
0.18
3
NIR1-A1
76031
0.09
4
SWIR1-W
44616
-0.05
5
NPV
41137
-0.52
6
RVSI
40241
-0.63
7
SWIR2-A2
34957
0.02
8
SWIR1-As
33623
-0.05
9
YE-DArea
32499
0.60
10
Blue-D
30283
-0.36
Crown Scale
Rank
Metric
Times in DTs
Shade Corr.
1
RW-λ
21486
0.21
2
RWI
17004
0.18
3
NDWI
16618
0.09
4
SWIR3-D
14372
0.17
5
YE-DArea
13030
0.28
6
NPV
12128
-0.42
7
SWIR2-A2
9836
-0.16
8
RVSI
9800
-0.36
9
SWIR2-A1
9531
-0.15
10
SWIR1-W
7056
0.12
178
179
Classification
Table 4.8. Error matrix for crown-scale classification using all spectral metrics (Kappa = 0.62).
Species
BAEL
CEPE
DIPA
HYAL
HYME
LEAM
TEOB
Total
Producer’s
BAEL
11
8
2
6
2
29
37.9%
CEPE
4
2
2
2
10
40.0%
DIPA
5
3
65
2
1
5
81
80.2%
Field Reference
HYAL HYME LEAM
2
2
1
1
1
1
6
28
2
1
8
14
2
34
14
21
82.4% 57.1% 66.7%
179
TEOB
4
1
20
25
80.0%
Total
21
13
82
37
16
21
24
214
User’s
52.4%
30.8%
79.3%
75.7%
50.0%
66.7%
83.3%
70.1%
Percent Refelctance
0.8
A
0.7
0.6
RE
0.5
0.4
GP
0.3 Blue
BE YE
0.2
0.1
NIR1
NE2
NE1
NIR2
SWIR1
SE
SWIR2
SWIR3
BAEL
CEPE
DIPA
HYAL
HYME
LEAM
TEOB
RW/Red
0
350
600
850
1100 1350 1600 1850 2100 2350
Wavelength (nm)
Percent Refelctance
0.8
B
0.7
NIR1
NIR2
SWIR1
0.6
SWIR2 SWIR3
RE
0.5
0.4
RW
0.3
0.2
BAEL
CEPE
DIPA
HYAL
HYME
LEAM
TEOB
0.1
0
350
600
850
1100 1350 1600 1850 2100 2350
Wavelength (nm)
Percent Reflectance
0.8
0.7
C
DIPA
0.6
0.5
Trunk
0.4
Branch
0.3
0.2
0.1
0
350
600
850
1100 1350 1600 1850 2100 2350
Wavelength (nm)
Figure 4.1. Laboratory spectra for study species. A) Mean spectra of leaf
samples, B) Mean spectra of bark samples, C) Dipteryx (DIPA) trunk and
branch bark spectra. All spectra are convolved to HYDICE bands.
180
1.0
0.8
0.8
0.6
0.6
GV
NPV
1.0
0.4
0.2
0.4
0.2
0.0
0.0
BAEL CEPE DIPA HYAL HYME LEAM TEOB
BAEL
0.14
NIR1-A1
NDWI
DIPA
HYAL HYME LEAM TEOB
20.0
0.10
0.06
0.02
15.0
10.0
5.0
-0.02
-0.06
0.0
BAEL CEPE DIPA HYAL HYME LEAM TEOB
BAEL CEPE DIPA
669.0
25.0
668.0
20.0
667.0
15.0
SR
RW-λ
CEPE
666.0
665.0
HYAL HYME LEAM TEOB
10.0
5.0
664.0
0.0
BAEL CEPE DIPA HYAL HYME LEAM TEOB
BAEL CEPE DIPA HYAL HYME LEAM TEOB
-0.001
8.0
YE-DArea
SWIR2-A1
10.0
6.0
4.0
2.0
0.0
-0.006
-0.011
-0.016
BAEL CEPE DIPA HYAL HYME LEAM TEOB
BAEL CEPE DIPA HYAL HYME LEAM TEOB
Figure 4.2. Species mean (bars) and standard deviation (error bars) for
selected spectral metrics calculated from pixel-scale (white; n=300) and
crown-scale spectra (gray; n=214).
181
NPV
Species
BAEL
CEPE
DIPA
HYAL
HYME
LEAM
TEOB
GV
Shade
Figure 4.3. Spectral mixture analysis (SMA) fractions of green
photosynthetic vegetation (GV), non-photosynthetic vegetation (NPV) and
shade for individual tree crowns. Ternary diagram values are the individual
tree crown fractions from crown-scale spectra.
182
SWIR2-A1 < 3.7
|
YE-DArea < -0.007
SWIR2-A2 < 1.4
183
NDWI < 0.09
YE-DArea < -0.01
SWIR3-D < 0.06
NPV < 0.09
PA
DI
AM
LE
PA
DI
L
PA
EL
BA
EVI < 0.084 SE-Mag < 0.0007
SAVI < 0.49
E
BA
B
NE2-λ < 1114.23
DI
HY
ME
BA
EL
HY
ME
NIR1-As < 0.93 SWIR1-D < 0.07
AM
LE
PE
CE
PA
DI L
E
A
RWI < 0.10
SWIR2-W < 97.74
NDWI < -0.01
RW-λ < 666.2
BA
PA
DI
ME
HY
TEOB
TEOB HYAL
NDWI < 0.02
CEPE
HY
AL
RWI < 1.1
HYAL
BA
EL
SWIR1-W < 39.6
C
Figure 4.4. An example decision tree constructed to classify tree species with all 77 crown-scale spectral metrics.
Species codes are listed in Table 4.2. Node depth is scaled to the change in deviance relative to the parent node.
183
Color
SMA
DIPA
HYAL
NDWI
NDVI
SWIR2-A1
HYAL
GV = 54%
NPV = 0%
Shade = 46%
DIPA
GV = 21%
NPV = 33%
Shade = 46%
NDWI = 0.08
NDVI = 0.91
SWIR2-A1 = 4.1
NDWI = 0.01
NDVI = 0.69
SWIR2-A1 = 8.9
Figure 4.5. A 300 x 300-m section of hyperspectral data and derived metrics
over individual Dipteryx (DIPA) and Hyeronima (HYAL) trees, delineated
with yellow polygons. The natural color image displays reflectance bands
482, 550, and 679 nm. The SMA fraction image displays (Red = NPV, Green
= GV, Blue = Shade). High values of NDWI, NDVI and SWIR2-A1 are
white. Mean values of SMA fractions, NDWI, NDVI and SWIR2-A1 are
shown for the DIPA and HYAL individuals.
184
Percent Reflectance
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
DIPA
HYAL
350
600
850 1100 1350 1600 1850 2100 2350
Wavelength (nm)
Metric
GV
NPV
Shade
NDVI
SR
RVSI
Red-A1
YE-DArea
NDWI
NIR1-A1
NIR2-A1
SWIR1-A1
SWIR2-A1
DIPA
19.9
61.2
18.9
58.3
3.8
1.1
49.9
0.39
-1.8
7.3
11.8
4.1
6.1
HYAL
82.1
12
16.7
89.9
18.9
2.7
101.8
-1.53
7.8
10.3
14.9
2.9
1.4
Figure 4.6. A reflectance spectrum and associated metrics for a single pixel
from the Dipteryx (DIPA) and Hyeronima (HYAL) crowns delineated in Fig.
4.5. Values for SMA fractions, NDVI, RVSI, YE-DArea and NDWI are
multiplied by 100.
185
CHAPTER 5: Classification of tree species with multiple endmember spectral
mixture analysis
5.1. Introduction
Species-level mapping of individual tree crowns (ITCs) using remote sensing
technology has immense potential for extending our understanding and monitoring
of species distributions and organization in the face of global climate change and
increased pressure on forest resources (Chapter 3; Gougeoun & Leckie, 2003). A
variety of spaceborne and airborne sensors offer high spatial resolution (<4 m),
multispectral digital imagery in which ITCs are readily delineated by manual or
automated methods.
Some progress has been made toward automated, digital
classification of ITC species in relatively species-poor temperate and boreal forests
(McGraw et al., 1998; Gougeon & Leckie, 2003).
However, my analyses in
species-rich tropical rain forest (TRF) indicate that multispectral data are insufficient
for discriminating even a limited number of tree species (Chapter 3). Spectral
variability among species is greatly limited due to dominant biochemical controls on
photon absorption by chlorophyll and water in tissues (Price, 1992; Cochrane, 2000).
Subtle species differences in tropical tree spectra do exist at leaf to crown scales, but
harnessing this information requires the higher spectral resolution offered by
imaging spectrometers, or hyperspectral sensors (Cochrane, 2000).
The spectral properties of materials within the sensor’s instantaneous field of
view (IFOV) mix to create the spectrum within an image pixel. In tree canopies, the
186
dominant materials of concern are photosynthetic leaf and non-photosynthetic bark
tissues, which have strongly contrasting spectral properties (Asner, 1998; Roberts et
al., 2004; Williams, 1991). Variation in pixel spectra among and within species is
largely related to differences in relative proportions of leaf and bark tissues within
the sensor IFOV, as controlled by crown architecture and phenology. Leaf and bark
spectra may also mix with the spectral characteristics of soil, understory vegetation,
neighboring trees, strangler trees, lianas, epiphytes, lichen, or moss. The extent to
which these factors can be considered noise or signal is related to species specificity.
For example, lichens may only affect trees with a certain bark texture, and so lichens
may add a unique spectral component to a species or functional type. Complicating
matters, individual trees within a population may exhibit a site-specific spectral
response to nutrient levels, water balance, light conditions, and herbivory. These
factors increase conspecific (within-species) variability and make automated TRF
tree species discrimination more challenging.
One promising technique for the analysis of mixed pixels is Spectral Mixture
Analysis (SMA), which models image spectra as a linear combination of two or
more dominant or “pure” spectral components, or endmembers (reviewed in
Keshava & Mustard, 2002).
Endmembers are selected from a library of field,
laboratory, image-derived or “virtual” spectra. A shade endmember is generally
used to account for spectral illumination variation. Other bright endmember(s)
represent major spectral components, such as green photosynthetic vegetation (GV:
leaves, photosynthetic bark), non-photosynthetic vegetation (NPV: litter, bark,
branches) and soil (Roberts et al., 1993).
187
Mixed reflectance spectra (ρ’λ) are
modeled as the sum of the fractional abundance of each endmember according to the
following formula:
n
ρ λ = ∑ f i ∗ ρ iλ + ε λ
,
(1)
i =1
where n is the number of endmembers in the mixing model, fi is the fractional
abundance of each endmember, ρiλ is the reflectance spectrum of each endmember,
and ελ is a residual error spectrum. SMA output is the fractional abundance of each
endmember, and all fractions are generally constrained so that they sum to 100%.
The root mean squared error (RMSE) of the model fit can by calculated according to
the formula:
b
RMSE =
(ε λ )
∑
λ
2
=1
(2)
b
where b is the number of bands.
The model solution is strengthened with
hyperspectral data over multispectral data because there are more high-signal bands
than model endmembers.
A single linear-mixture model with two to three endmembers will not adequately
model every pixel in a spectrally-complex image (Dennison & Roberts, 2003a;
Roberts, Gardner et al., 1998), such as from a TRF canopy. Multiple Endmember
Spectral Mixture Analysis (MESMA) tackles this problem by automatically
selecting endmembers on a per-pixel basis from a library of spectrally-diverse
endmembers (Dennison & Roberts, 2003a; Roberts, Gardner et al., 1998).
188
MESMA is also a promising classifier. In this mode, the class of the spectrum
being unmixed is determined based on the class label of the optimally-assigned, nonshade endmember in the SMA modeling process (e.g., bright endmember from a
particular species). Although there have been no tropical applications involving
MESMA classification, it has been used to map dominant chaparral vegetation types
in southern California, USA (Dennison & Roberts, 2003a; Roberts, Gardner et al.,
1998; Roberts, Dennison et al., 2003), soils (Okin et al., 2001), and snow cover and
grain size (Painter et al., 1998). Best results (88.6% overall accuracy) were achieved
when using a regionally-specific image endmember library and when assessing
accuracy at the scale of a vegetation patch rather than at the image pixel scale
(Dennison & Roberts, 2003a).
Conducting MESMA with a large set of candidate endmembers can be
computationally intensive when applied to a high spatial and spectral resolution
image because multiple models must be evaluated on a per-pixel basis. Methods are
thus necessary to reduce the endmember library to contain only those spectra that are
pertinent to species discrimination. In particular, the spectral library should include
those endmembers that specialize in modeling conspecific spectral variability, while
excluding more generalist endmember spectra that model a broad range of pixels
from other species or materials within the scene.
One approach to MESMA
endmember selection includes using expert knowledge of spectral mixtures within a
scene to guide the collection of a limited number of pure field or image spectra that
produce known fractional abundance (Okin et al., 2001; Painter et al., 1998).
Bateson et al. (2000) developed a spectrally-diverse library of image endmembers by
189
including bundles of spectra arrayed in the convex hull of spectral space. Asner and
Heidebrecht (2002) conducted SMA in an arid landscape using endmember bundles
simulated with Monte Carlo sampling of field spectra. Roberts, Gardner et al.
(1998) explored an alternative method where a large spectral library of reference
endmembers was searched using a maximal coverage procedure to find endmembers
that both mapped a large spatial area in the image while had minimal model overlap.
Finally, Dennison and Roberts (2003a) developed an automated technique that
selected endmembers that best modeled other spectra in an image-derived spectral
library. The single candidate endmember that produced the minimum endmember
average RMSE (EAR) for models within its class was selected to represent the class
in MESMA classification. This technique was later expanded to include two optimal
endmembers per class for multi-temporal characterization of chaparral vegetation
(Dennison & Roberts, 2003b).
Despite its ability to incorporate within-class variability, researchers have noted
that MESMA, as a classifier, suffers from many of the same problems that plague
other classifiers—high conspecific variability, lack of sharp interspecific (among
species) spectral contrast, mixing with background material, and sensor noise (Okin
et al., 2001; Roberts, Gardner et al., 1998). Furthermore, it is assumed in SMA that
the non-linear mixing of photons among spectral components is minimal. Nonlinear mixing does occur in vegetation mainly because leaves permit light
transmission and photon multiple-scattering, especially in NIR wavelengths (Roberts
et al., 1993). Non-linear mixing will tend to increase fraction estimation error and
190
possibly decrease MESMA classification accuracy by selecting the wrong
endmember as optimal in a linear mixing solution.
The objective of this chapter was to assess MESMA for the species-level
classification of ITCs in a tropical rain forest. Starting with a large spectral library
of image and laboratory spectra, I developed a new automated approach to select
optimal endmembers for two- and three-endmember MESMA models. In particular,
the method sought to find a parsimonious set of endmember spectra that could
simultaneously model the greatest number of conspecific spectra with little overlap,
while modeling the fewest spectra from other species (extraspecific).
5.2. Methods
5.2.1. Canopy-emergent trees
My analyses focused on the classification of canopy emergent individuals of
seven tree species (Table 4.1). Chapter 3 provides details about the species and
number of study trees and how their ITC polygons were digitized on the HYDICE
hyperspectral imagery. As explained in Chapter 3, some overstory tree species are
completely deciduous, generally beginning in the first dry season, while others are
evergreen and continuously flush small amounts of leaves throughout the year
(Table 4.1). Hyperspectral imagery was acquired on March 30, 1998 (Chapter 1), at
the end of the first dry season, and all study trees were expected to have high mature
leaf cover except DIPA and LEAM (summarized in Table 4.1). However, BAEL
and HYME had relatively fine compound leaves and so their mature leaf cover was
expected to have a lower LAI relative to the broadleaf species CEPE, HYAL and
191
TEOB. I thus expected LAI for the study species to be low for DIPA and LEAM,
moderate for BAEL and HYME, and high for CEPE, HYAL and TEOB.
5.2.2. Laboratory bark spectra
Pure NPV endmembers (i.e., bark) were difficult to locate in the hyperspectral
image because these components rarely occupied a whole image pixel, and so
laboratory-measured endmembers were used in 2EM+Shade models (see Section
5.2.4). Bark specimens from the 7 study species were sampled in the station vicinity
and their reflectance was measured in a laboratory dark room with an ASD
FieldSpec spectrometer (Analytical Spectral Devices, Boulder, CO, USA), which
has 1-nm spectral sampling covering 350 to 2500 nm (detailed methods in Chapter
4). These ASD bark spectra were convolved to HYDICE band center positions (161
bands) using full-width, half-maximum information for each HYDICE band. The
final laboratory library contained 66 bark spectra. DIPA had considerable withinspecies variation due to some specimens having green bark (Chapter 4; Fig. 5.1).
5.2.3. Scales of analysis
As in Chapters 3 and 4, two ITC species classification schemes were analyzed:
crown-scale and pixel-majority. The crown-scale approach labeled ITC species
using classified crown-scale spectra.
Crown-scale reflectance spectra were
calculated as the mean of all within-crown pixels, providing one spectrum per
crown, and MESMA was then applied directly to crown-scale spectra. The pixelmajority classification scheme applied the MESMA classifier to pixel-scale
192
reflectance spectra, and the ITC species label was assigned to the class representing
the majority of classified pixels within the each crown.
5.2.4. Multiple endmember spectral mixture analysis (MESMA) classifier
The MESMA classifier was evaluated using 1) multiple two-endmember models
composed of an image spectrum and photometric shade (1EM+Shade), and 2) with
multiple three-endmember models composed of a GV image spectrum, a NPV (bark)
laboratory spectrum, and photometric shade (2EM+Shade). Photometric shade is a
spectrum of zeros.
Prior to MESMA, endmembers were selected for each species from a spectral
library. The goal of the selection process was to find a set of endmembers from the
library that modeled the maximum number of conspecific spectra while modeling the
minimal number of extraspecific spectra. For 1EM+Shade MESMA, the spectral
library included 300 pixel spectra per species from study crowns in the image (7
species x 300 pixels = 2100 total pixels). Pixels were drawn randomly from every
crown without regard to their brightness, but sampling from any given crown was
terminated if it had 40 or fewer available pixels. The sample size of 300 pixels per
species was arbitrary yet ensured that crowns had a large number of independent test
pixels after samples were removed from the classifier.
Endmember selection proceeded as follows: every pixel spectrum was used to
unmix every other pixel spectrum in the spectral library. Each model was
considered valid according to the following constraints. The sum of SMA fractions
was constrained to sum to 100%; however, valid models could have non-shade
193
endmember fractions between -6% and 106%. This fraction constraint was
empirically determined by Halligan (2002) to allow for model error while still
providing acceptable MESMA classification accuracy. Highly shaded pixels can
have low signal-to-noise and can produce unreliable models; and thus, I imposed a
constraint that the modeled shade fraction be less than 80% (Roberts, Gardner et al.,
1998). Model fit was also considered unacceptable if overall error (RMSE) was
greater than 2.5% (Roberts, Gardner et al., 1998). Some spectra could have spurious
features in model residual spectra, such as spikes due to poor sensor calibration or
extreme illumination-view geometry, yet still yield acceptable RMSE. To
distinguish between spectra with these erroneous features from residuals resulting
from the presence or absence of an absorption feature in the candidate endmember, I
imposed a constraint that no seven contiguous bands have reflectance above 2.5%
(Roberts, Gardner et al., 1998). Since HYDICE band spacing was variable across
the full spectral range, this 7-band threshold applied to a spectral region between 23nm to 90-nm wide.
Any given candidate spectrum could be modeled by multiple spectra from its
class (e.g., conspecific in the case of species) as well as with spectra from other
classes (e.g., extraspecific). For each species, I sought endmembers that had the
highest percentage of successful conspecific models and the lowest percentage of
successful extraspecific models. This was accomplished by finding the candidate
endmember within each class that maximized the COunt-Based Index (COBI)
according to Equations 3 through 5:
194
n
COBI con =
COBI extra
∑ Mod
i =1
n
i
Mod = 0: not modeled
Mod = 1: modeled
(3)
p
t
⎛ o
⎜ ∑ Mod j ∑ Mod k
Mod l
∑
⎜ j =1
k =1
l =1
+
+
+
...
⎜
o
p
t
⎜
=⎝
c −1
⎞
⎟
⎟
⎟
⎟
⎠
(4)
COBI = COBI con − COBI extra
(5)
where c is the number of extraspecific classes excluding the candidate species (i.e., c
= 6) and Mod is a SMA model of the candidate endmember unmixing other n
conspecific (i.e., n = 299 excluding itself) and o, p, … t spectra per extraspecific
class. In this chapter, there were 6 extraspecific classes, each with 300 spectra, and
so o, p, q, r, s, and t were equal to 300. The candidate endmember was either
successful in modeling the spectrum (Mod = 1), or not (Mod = 0). The endmember
selection scheme is depicted in Figure 5.2 for three hypothetical species (A, B & C),
each with 4 candidate endmember spectra. These candidate spectra (in columns) and
photometric shade were used in two-endmember SMA to model spectra in rows.
195
Successful SMA models are indicated as black (conspecific) or gray (extraspecific)
boxes, while unsuccessful models are white boxes. The COBI was calculated for
each candidate endmember column (Eq. 3-5), and the best species endmember had
the highest COBI. Ties among candidate endmembers with the same COBI were
decided by selecting the endmember with the lowest endmember average RMSE
(EAR; Dennison & Roberts, 2003a), calculated from all n conspecific models within
the endmember’s column. In this example, Species A‘s candidate endmember #3
was selected because it had the highest COBI within its class (Fig. 5.2). Species B
had endmembers with the least amount of extraspecific spectral similarity (positive
COBI indices) while Species C had the most extraspecific spectral overlap (zero to
negative COBI indices).
Once an optimal endmember was chosen for its class, the spectrum and all
spectra successfully-modeled were removed from the analysis and then COBIs were
re-calculated for the remaining spectra. The new COBIs were then used to select
another endmember for each class. By removing spectra modeled by the first
selected endmember, the second-iteration endmember had reduced spectral
similarity with the first-iteration endmember. This iterative-selection process was
repeated for each class until COBI was zero or less, indicating that remaining
candidate endmembers modeled a greater percentage of extraspecific spectra than
conspecific spectra. Final endmembers were not necessarily pure GV spectra, but
rather bright spectra that could model other unique mixtures within each species.
For example, an endmember from a deciduous leaf-off crown could have been a
bright NPV-dominated spectrum.
196
A similar COBI approach was used for selecting jointly-optimal endmembers for
2EM+Shade MESMA models composed of shade, GV and NPV. Bark spectral
endmembers for each species were selected from the laboratory library of 66 spectra
(Section 2.2.2). Even though this library included some green and presumably
photosynthetic bark, I considered it a “pure” NPV spectral library. Candidate image
endmembers representing GV were selected by first calculating the NIR minus red
contrast in all 2100 pixels in the image spectral library and then selecting those
spectra that were in the top 90th percentile for each class. This reduced library
included 217 image spectra (31 pixels per species). Every combination of GV
(image) and NPV (laboratory) spectra were matched within each species, forming
2046 2EM+Shade candidate models, which were then used to unmix the original
library of 2100 image spectra. The COBI endmember selection scheme, as
described for 1EM+Shade models, was then implemented for simultaneous selection
of GV and NPV endmember pairs for each species.
Pixel-majority MESMA classification of ITCs proceeded by modeling pixels
within all crowns using 1EM+Shade or 2EM+Shade models. Pixels included in the
image spectral library were excluded from the classification analyses. For
2EM+Shade models, only combinations of GV and NPV endmembers from the same
species were considered (e.g., a GV image endmember from BAEL and an NPV
laboratory endmember from BAEL). The best model for each pixel was chosen
using the same constraints used in COBI endmember selection: non-shade fractions
were between -6% and 106%, the shade fraction was less than 80%, no 7 contiguous
model residuals were above 2.5% reflectance, and the model RMSE was less than
197
2.5%. If multiple models met these constraints for a given pixel, then the model
with the lowest RMSE was selected. A species label was assigned to each pixel
using the class of the best-fit model for the pixel. Pixels with no acceptable models
were labeled as unclassified. For the pixel-majority ITC classification, the species
label was assigned based on the majority class of classified pixels within each
crown. Image and laboratory endmembers selected for pixel-scale MESMA were
used in crown-scale MESMA classification. In this case, 1EM+Shade and
2EM+Shade models were used to model crown-scale spectra.
5.3. Results
5.3.1. Endmember selection
A matrix of 2100 candidate endmembers (columns) modeling the same spectra
(rows) using 1EM+Shade SMA is shown in Figure 5.3. Models meeting the
constraints (Section 2.2.4) were colored black, and white otherwise. Spectra were
sorted primarily by species and secondarily by spectral brightness based on their
800-nm reflectance, with spectra ranked from dark (left or up) to bright (right or
down). As expected, the brighter candidate spectra within a species successfully
modeled more spectra than darker spectra, producing within-species triangular
patterns in the modeled/un-modeled matrix (Fig. 5.3). Part of this trend is due to the
relatively liberal -6 to 106% fraction constraints on successful models. I expected
that candidate endmembers from leaf-off deciduous species (DIPA and LEAM)
would model spectra between those two species but not among leaf-on species
because leaf-off crowns had distinct spectral properties such as lower NIR
198
reflectance and shortwave infrared absorption features from exposed NPV (Chapter
4). Contrary to what was expected, many deciduous candidate endmembers modeled
spectra from across all species, regardless of phenology (Fig. 5.3). Also, most dark
spectra of HYAL and HYME were successfully modeled by candidate endmembers
from other species, suggesting potential confusion of HYAL and HYME by other
species endmembers during actual MESMA classification (Fig. 5.3). However,
Figure 5.3 only depicts spectra that meet model constraints (black dots), and does
not take into account the RMSE selection criteria used for final model selection in
MESMA.
The first COBI-selected endmember per species modeled 41% (BAEL) to 81%
(TEOB) of conspecific image spectra in the library (Table 5.1). A second and third
COBI endmember modeled an additional 13% (LEAM) to 39% (BAEL) of
conspecific spectra. The fourth and greater endmembers per species modeled a low
percentage of conspecific spectra. Although training pixels came from every crown,
the final set of endmembers for each species came from five (TEOB) to nine (DIPA)
different crowns (Table 5.1, Crown ID).
A spectrum’s 800-nm (near infrared) reflectance was used to gauge spectral
brightness (Table 5.1). No deciduous leaf-off endmembers had 800-nm reflectance
greater than 58.0%, while the leaf-on species endmembers covered a wider range of
brightness (e.g., up to 86.6% for HYAL). A near-infrared reflectance of 86.6% is
unusually bright for a plant image spectrum. The presence of extremely bright
spectra in the image resulted mainly from poor radiometric calibration of the
hyperspectral data and solar geometry that made the eastern sides of the crowns
199
relatively bright (HYDICE scene was acquired early in the morning at 7:55-8:27 am
local time, 56.3° to 48.4° solar zenith, and 92° to 94° solar azimuth). Extremely
bright endmembers are expected to properly model other pixels within the crown if
their spectral mixtures vary only by brightness and not shape.
Leaf-off DIPA image spectra in the library showed clear separation from a leafon species, TEOB (Fig. 5.4). Relative to DIPA, TEOB had more 680 nm and 1650
nm absorption (lower reflectance) due to chlorophyll and water, respectively. The
scatter plots also depict the mixing lines connecting three COBI-selected
endmembers per class (Fig. 5.4). These endmembers were not necessarily the
brightest pixels in the library. Also, some DIPA spectra strayed into the TEOB
spectral space determined by 680 and 800 nm bands, indicating the presence of
green tissues in their spectral mixtures. Endmembers 3, 5, 8, and 9 for DIPA formed
mixing lines in the TEOB 680-800 nm space (Fig. 5.4, EM3 is shown), and it was
expected that these endmembers would be misclassified as TEOB or other leaf-on
species in a MESMA classification.
5.3.2. Pixel-majority ITC classification
When MESMA was applied to pixels with 1EM+Shade models, overall
classification accuracy was 36.3% with one optimal endmember per species, and
accuracy increased 4.4% with the inclusion of 5 to 9 COBI-selected endmembers per
species class (Table 5.2). However, including multiple endmembers per class did
not necessarily improve class Producer’s accuracies. For example, the best CEPE
Producer’s accuracy included 5-9 endmembers per species, while the best LEAM
200
Producer’s accuracy included just one endmember per species. When all 2100
image spectra in the library (300 spectra per species) were included in 1EM+Shade
MESMA, overall accuracy and most Producer’s accuracies were higher than when
using COBI-selected endmembers.
The best pixel-majority ITC classification using COBI-selected endmembers had
59.8% overall accuracy, with 1EM+Shade models with 3 endmembers per species
(Table 5.3). With the 2EM+Shade endmember selection, COBI values were
negative after selecting one pair of GV-NPV endmembers for HYAL, HYME and
TEOB, while DIPA had up to 4 GV-NPV endmember pairs. The 1EM+Shade
classification with 3 endmembers per species was 6.5% more accurate than the
equivalent 2EM+Shade classification with 1-3 endmembers (not significant, Z=1.86,
α=0.05;Congalton, 1991).
Example ITCs for each species are shown for 1EM+Shade MESMA with 3
endmembers per species (Fig. 5.5). All ITCs in this example were correctly-labeled.
Some ITCs had spectra with similar shapes but differed in brightness, and so a single
endmember modeled most of the crown. For example, 96% of the pixels in the
CEPE crown in Figure 5.5 were modeled by its third optimal endmember (EM3),
which was relatively bright (Table 5.1). In contrast, the DIPA crown had 34% of its
pixels modeled by EM1 and 19% of its pixels modeled by EM3 (Fig. 5.5). These
two DIPA endmembers had NPV and GV spectral properties, respectively (Fig. 5.4).
Only 48% of the HYME crown’s pixels were correctly classified, which may be a
result of broad spectral overlap with endmembers from other species (Fig. 5.3,
discussed in Section 5.3.1). However, the pixel-majority vote was HYME because
201
no other species had a greater number of pixels classified within the crown. The
error matrix for this classification (Table 5.4) revealed misclassification of
deciduous DIPA (mostly leaf-off) individuals with BAEL individuals, a leguminous
species with fine compound leaves (Chapter 3). Some BAEL individuals were also
mapped as HYME, another leguminous, composite-leaved species. Crowns of
TEOB were confused with HYAL, which were both evergreen species with broad
leaves.
When all 2100 image spectra in the library (300 spectra per species) were
included in 1EM+Shade models, pixel-majority MESMA accuracy increased
dramatically to 90.2% (Tables 5.3 and 5.5). Much of the confusion in the
classification was with DIPA crowns classified as BAEL, as seen with MESMA and
3 endmembers per species. Overall accuracy did not significantly improve (Z=0.43)
with 2EM+Shade models using 2100 image spectra matched by species with 66
laboratory bark spectra (19,800 unique models; Table 5.3, pixel-majority).
5.3.3. Crown-scale ITC classification
In general, overall accuracies at the crown-scale were lower than with the pixelmajority technique (Table 5.3). The best COBI-based MESMA accuracy was only
54.7% with 1EM+Shade models composed of three image endmembers per species.
When including all 2100 image spectra in 1EM+Shade models, overall accuracy
reached 67.3%.
202
5.3.4. Model constraints
Roberts, Gardner et al. (1998) implemented MESMA with model fractions
constrained from -1% to 101%. The reference spectral library used in that study had
high radiometric quality and tended to be bright relative to image spectra, which
include shadowing in the IFOV. Working in a temperate forest landscape, Halligan
(2002) found that MESMA classification accuracy improved when fraction
constraints were relaxed to accommodate extremely bright or dark endmembers in
an image spectral library. Concurring with those results, I also found that fraction
constraints of -1% to 101% on COBI and MESMA produced weaker classification
accuracies than those presented for -6% to 106% fraction constraints (data not
shown).
5.4. Discussion
5.4.1. Count-based index (COBI) selection of image endmembers
The primary goal of the COBI endmember selection scheme was to find the set
of spectra that model the most within-species spectra with minimal ability to model
spectra from other species. In other words, I sought endmembers from the spectral
library that were conspecific ”specialists” as opposed to interspecific “generalists”
(Roberts, Gardner et al., 1998). It was apparent from Figure 5.3 that this objective
would be difficult to achieve for this forest type since most bright candidate
endmembers were generalists, regardless of species. This is because vegetation
spectra had similar spectral shapes due to dominant biochemical controls on photon
absorption within tissues (Price, 1992).
203
Despite this broad spectral overlap among spectra, the COBI technique was
successful in selecting at least one specialist endmember for each species. The firstiteration endmembers modeled a large proportion of conspecific spectra in the
library (41% to 81%). Species-level differences in pixel- and crown-scale spectra
are related to crown structure and phenology, which controls the spectral mixing of
leaves, bark and shadows within the sensor IFOV (Asner, 1998; Roberts, Ustin et al.,
2004). Specialist endmembers should thus respond to unique aspects of crown
structure and phenology for each species.
5.4.2. MESMA classification
With one COBI endmember per species, MESMA was able to correctly classify
a third of ITC pixels. About half of ITCs were correctly classified with the pixelmajority and crown-scale classifiers (1EM+Shade models). These results indicate
that there are unique spectral properties of species at both pixel and crown scales
that can be discriminated by MESMA with one image endmember per species.
Despite this encouraging finding, the overall accuracy of the final ITC
classification was still too low for operational use. Including up to three COBIselected endmembers in 1EM+Shade MESMA increased pixel-scale accuracy,
translating into improved pixel-majority accuracy. Evidently, the 2 to 3
endmembers identify other unique spectra for species at pixel and crown scales. For
example, the first three DIPA endmembers included a spectrum with characteristics
of NPV (EM2), one spectrum similar to a leaf-on TEOB spectrum (i.e., high GV,
EM3), and a spectrum in between these two spectra (EM1). Additional endmembers
204
(4+) tended to decrease overall accuracy, likely a result of additional endmembers
having spectral overlap with extraspecific spectra, which leads to increased class
confusion in MESMA.
MESMA with 1EM+Shade models generally had higher overall accuracies than
with 2EM+Shade models. Relative to 1EM+Shade models, including one laboratory
and image spectrum per species in 2EM+Shade models decreased overall accuracy
by 2% with pixel-majority while 11% with crown-scale analysis. I concluded that
including laboratory NPV spectra in models does not help discriminate species. One
explanation is that laboratory NPV spectra may differ substantially from image NPV
spectra due to differences in measurement scale, scattering environments,
atmospheric contamination, and radiometric calibration. In this situation, laboratory
NPV spectra would be too dissimilar from image NPV to be effective in the mixture
models with image GV endmembers. Unfortunately, pure NPV endmembers were
difficult to acquire for leaf-on species, and even NPV spectra from leaf-off species
had distinct chlorophyll signals (i.e., red absorption well) from internal photon
scattering and mixing with epiphytes, lianas, mosses or lichen on branches. In
contrast, image endmembers included the whole range of GV-dominated to NPVdominated spectral mixtures and incorporated sensor calibration artifacts and nonlinear mixing from photon scattering that is absent in laboratory spectra. My results
indicate that, although they may not be “pure”, image endmembers provide a
considerable level of species discrimination at pixel and crown scales. When mixed
pixels are modeled with slightly-mixed GV and NPV endmembers, there may be a
variety of models that have adequate fits. In this case, the fractional abundance of
205
NPV relative to GV in a crown may help distinguish an ITC species, rather than the
endmembers’ species label.
Regardless of MESMA model complexity, the overall accuracies with COBI
endmembers were lower than the 86% (Chapter 3) and 68% (Chapter 4) overall
accuracies with pixel-majority linear discriminant analysis (LDA) and decision tree
(DT) classifiers, respectively. MESMA applied to crown-scale spectra had even
lower accuracies relative to those using LDA and DT classifiers (Chapters 3 & 4). I
was able to boost MESMA accuracy to a relatively high level (90%) only by
including the full library of image spectra as potential 1EM+Shade models.
However, one caveat with this result is that endmember-selection pixels and testing
pixels came from the same crowns. Although the endmember-selection pixels were
excluded from ITC classification analyses, there were still spatial autocorrelation
effects—pixels from the same crown have similar spectral properties. It is likely
that the accuracies reported for the full-library MESMA are optimistic due to a lack
of complete independence between library and testing data. This bias was negligible
when considering only nine or fewer endmembers per species because ITCs
contained 41 to 662 pixels, no crown had fewer than 40 testing pixels, and no more
than 2 endmembers came from the same crown (Table 5.1). In Chapters 3 and 4, DT
and LDA analysis were performed with a cross-validation approach and so
classifiers did not have autocorrelated spectra in training sets.
206
5.4.3. Recommendations for future research
Despite the potential bias when using the full spectral library to classify ITCs, it
appears that many more than five endmembers per species are necessary to
adequately use MESMA as an ITC classifier in this TRF environment. These results
suggest that future research using MESMA for species classification should focus on
methods to increase conspecific endmember variability while minimizing
extraspecific endmember similarity. However, increasing the number of mixture
models does have a computational cost. For example, in the per-pixel analysis of
2100 1EM+Shade models using IDL code (RSI, Inc., Boulder, CO, USA) running on
a 3.2 GHz Intel Pentium 4 processor, 35,043 ITC pixels were processed in 3 hrs, 14
min (181 pixels/min). The 19,800 2EM+Shade models took 29 hrs and 45 min. to
process (20 pixels/min). To process the entire HYDICE mosaic of 3,774,300 pixels
would have taken 14.5 and 133.5 days for evaluating all 1EM+Shade and
2EM+Shade models, respectively. In an application over a large spatial extent,
segmenting the hyperspectral image into ITCs of interest (e.g., Gougeon & Leckie,
2003) would greatly improve processing time by excluding canopy gaps or nontarget vegetation from the modeling process. As with all remote sensing
applications, limitations on processing time are also expected to decrease as
computer processing speed increases per unit cost according to Moore’s Law
(Moore, 1998).
I have identified two improvements for the COBI selection technique that may
be useful in future studies. For one, the extraspecific COBI formula (Equation 4)
was implemented with an equal class weighting scheme. Weights could be
207
multiplied by each summation term in Equation 4 to influence the selection of
certain types of endmembers. For example, if more NPV-dominated endmembers
were desired for leaf-off species, then high weights could be applied to leaf-on
species in Equation 4, which would select against GV-dominated endmembers by
lowering their COBI values (Eq. 5). Another COBI improvement involves how
model strength is evaluated. Just because a candidate endmember successfully
models extraspecific spectra in the library, the endmember may still have stronger
model fits with its conspecific spectra. The COBI formula could be modified to
favor candidates with stronger conspecific over extraspecific fits, possibly by
applying a weighting scheme to RMSE, residuals and fractions from model outputs.
5.5. Conclusions
A major goal of this chapter was to assess airborne hyperspectral imagery for
the species-level classification of ITCs in a tropical rain forest. The MESMA
classifier explored in this chapter was expected to have optimal performance with
hyperspectral over multispectral imagery of equal spatial resolution. I found that
MESMA has potential as a classifier, providing 90% overall accuracy, but only
when presented with a library of 2100 model endmembers (300 endmembers per
species). Processing this quantity of models on a per-pixel basis greatly increases
processing time for large study extents.
The requirement for 300 endmembers per tree species suggests that their crowns
have many unique spectral signatures, probably due to spectral mixing. For
208
MESMA to be a more useful classifier, the number of candidate endmembers will
need to be constrained to those that encompass the unique spectral shapes that make
species distinct. The endmember selection technique presented in this chapter
provides one method for finding the optimal set of endmembers. For example,
MESMA with one optimal endmember from each species discriminated ITC species
with 49% and 50% overall accuracy, depending on pixel-majority or crown-scale
evaluation, respectively.
Although my count-based MESMA classification accuracy was not adequate for
species-level mapping, the technique may yield more physically-accurate fractions
over standard SMA with single GV and NPV endmembers. Such improvements
could greatly benefit applications involving land-use classification, estimation of
deciduous and evergreen components of forest canopy, or linking fractions to
biophysical parameters, such as aboveground biomass and percent cover. In this
context, the proper matching of endmembers to their respective image pixels would
yield optimal fraction estimates, a process that may benefit from another type of
classifier. I found LDA to be a stronger classifier for my study species (Chapter 3).
One approach to estimating endmember fractions would be to apply the LDA
classifier to pixels, perform MESMA using only endmembers for each pixel’s class,
and then select the best fractions based on model parameters (e.g., lowest RMSE).
209
Table 5.1. Summary statistics of optimal endmembers selected by the countbased index (COBI) and 2100 training image spectra (Section 5.2.4). Each
row is a selected candidate endmember (EM). Columns are the species code,
selection rank, crown identifier, reflectance at 800 nm, count of valid
conspecific models, percent of all conspecific models, COBI * 100 (Eq. 3),
and average RMSE*100 (EAR).
Species EM # Crown ID 800 nm Count
Percent COBI EAR
BAEL
1
106
66.0
124
41.3
25
20
BAEL
2
177
55.4
47
15.7
13
23
BAEL
3
166
64.1
70
23.3
18
21
BAEL
4
61
60.3
29
9.7
40
27
BAEL
5
35
63.5
12
4.0
30
23
BAEL
6
149
24.8
5
1.7
22
56
BAEL
7
126
73.6
6
2.0
31
30
BAEL
8
106
63.5
2
0.7
27
39
CEPE
1
116
50.5
155
51.7
31
17
CEPE
2
93
64.8
61
20.3
16
23
CEPE
3
204
74.5
22
7.3
26
35
CEPE
4
10
35.9
20
6.7
16
35
CEPE
5
173
45.6
19
6.3
21
22
CEPE
6
113
73.9
7
2.3
28
25
CEPE
7
10
34.6
2
0.7
10
57
CEPE
8
110
28.8
2
0.7
11
26
CEPE
9
116
44.4
3
1.0
12
26
DIPA
1
136
40.5
178
59.3
37
16
DIPA
2
107
33.7
38
12.7
26
29
DIPA
3
103
41.5
28
9.3
18
23
DIPA
4
154
44.8
11
3.7
12
30
DIPA
5
5
49.2
7
2.3
11
28
DIPA
6
97
40.7
6
2.0
13
46
DIPA
7
39
43.9
5
1.7
9
29
DIPA
8
14
54.3
10
3.3
8
27
DIPA
9
62
48.8
3
1.0
13
30
210
Table 5.1. (continued).
HYME
HYME
HYME
HYME
HYME
HYME
HYME
HYME
HYAL
HYAL
HYAL
HYAL
HYAL
HYAL
LEAM
LEAM
LEAM
LEAM
LEAM
LEAM
LEAM
LEAM
LEAM
TEOB
TEOB
TEOB
TEOB
TEOB
1
2
3
4
5
6
7
8
1
2
3
4
5
6
1
2
3
4
5
6
7
8
9
1
2
3
4
5
45
128
147
171
185
186
121
212
202
145
187
200
211
19
184
168
153
179
161
179
168
142
135
89
79
87
92
85
79.0
45.7
70.3
36.3
48.0
58.4
67.0
64.7
69.3
72.4
80.0
86.6
43.0
77.5
46.2
11.4
58.1
34.2
44.3
57.8
12.9
45.3
43.1
65.4
70.0
27.7
78.3
61.6
197
51
17
16
6
4
2
2
226
39
8
13
7
2
208
13
26
18
14
4
3
7
2
242
34
13
5
2
211
65.7
17.0
5.7
5.3
2.0
1.3
0.7
0.7
75.3
13.0
2.7
4.3
2.3
0.7
69.3
4.3
8.7
6.0
4.7
1.3
1.0
2.3
0.7
80.7
11.3
4.3
1.7
0.7
48
25
24
25
26
19
20
14
45
22
19
29
37
27
48
14
18
26
18
18
17
17
25
41
37
45
42
32
14
15
20
21
30
25
25
31
13
16
25
17
15
26
14
71
20
28
24
28
63
26
26
12
14
19
15
18
Table 5.2. Pixel-scale percent producer’s accuracy and overall accuracy
using Multiple Endmember Spectral Mixture Analysis (MESMA) and
models of 1 to 9 image endmembers per species and photometric shade
(1EM+Shade). Numbers are percentages except n, the total of non-training
pixels per class (total n = 35043).
EM /
specie
Overall
s
BAEL CEPE DIPA HYAL HYME LEAM TEOB Accuracy
3752 2691 16115 4859
2317
2561 2748
n
1
25.7
33.4
34.3
52.4
44.8
20.3
51.0
36.3
2
32.4
47.6
41.8
34.5
48.8
20.2
39.5
39.9
3
35.6
52.8
49.5
30.6
50.8
25.2
38.2
44.5
4
28.8
52.7
45.4
31.5
52.4
21.8
39.1
42.0
5-9
30.1
65.7
42.3
30.6
48.5
25.4
30.7
40.7
300
55.6
60.6
48.2
59.0
53.7
46.3
74.5
53.7
212
Table 5.3. Overall ITC classification accuracy using Multiple Endmember
Spectral Mixture Analysis (MESMA). 1EM+Shade models included
photometric shade and 1 image endmember. 2EM+Shade models included
photometric shade, 1 image endmember and 1 laboratory endmember (see
Section 5.2.4).
No.
No.
2EM+Shade
EMs per Models 1EM+Shade
EMs per
Model
Models
species
Models
species
s
EM1
EM1
EM2
Pixel-majority
1
7
48.6
1
1
7
46.7
2
14
53.7
1-2
1-2
19
55.1
3
21
59.8
1-3
1-3
34
53.3
4
28
54.7
1-4
1-4
41
53.3
5-9
54
51.9
n/a
n/a
n/a
n/a
300
2100
90.2
300
5-15
19800
91.6
Crown-scale
1
7
50.0
1
1
7
39.3
2
14
52.8
1-2
1-2
19
52.3
3
21
54.7
1-3
1-3
34
51.4
4
28
53.3
1-4
1-4
41
51.4
5-9
54
52.8
n/a
n/a
n/a
n/a
300
2100
67.3
300
5-15
19800
76.6
213
Classification
214
Table 5.4. Error matrix of pixel-majority classification using 3 COBI-selected image
endmembers per species in 1EM+Shade models (Kappa = 0.49).
Field Reference
Species
BAEL CEPE DIPA HYAL HYME LEAM TEOB
BAEL
16
12
1
4
1
CEPE
6
1
8
4
DIPA
5
1
60
2
10
1
HYAL
2
2
1
22
1
9
HYME
4
2
3
9
1
LEAM
2
1
6
TEOB
1
3
1
2
9
Total
29
10
81
34
14
21
25
Producer’s 55.2% 60.0% 74.1% 64.7% 64.3% 28.6% 36.0%
214
Total
34
19
79
37
19
9
16
214
User’s
47.1%
31.6%
76.0%
59.5%
47.4%
66.7%
56.3%
59.8%
Classification
215
Table 5.5. Error matrix of pixel-majority classification using 300 image endmembers per species in
1EM+Shade models (Kappa = 0.88).
Field Reference
Species
BAEL CEPE DIPA HYAL HYME LEAM TEOB
Total
User’s
BAEL
28
9
1
38
73.7%
CEPE
10
1
11
90.9%
DIPA
67
1
68
98.5%
HYAL
2
32
34
94.1%
HYME
1
12
13
92.3%
LEAM
2
19
21
90.5%
TEOB
2
1
1
25
29
86.2%
Total
29
10
81
34
14
21
25
214
Producer’s
96.6% 100.0% 82.7% 94.1% 85.7% 90.5% 100.0%
90.2%
215
.
BAEL (N =15)
Reflectance
80%
60%
60%
40%
40%
20%
20%
0%
0%
350
850
Reflectance
1350
1850
2350
DIPA (N = 10)
80%
350
850
60%
60%
40%
40%
20%
20%
1350
1850
2350
HYAL (N = 5)
80%
0%
0%
350
850
1350
1850
2350
HYME (N = 8)
80%
Reflectance
CEPE (N = 9)
80%
350
850
60%
40%
40%
20%
20%
0%
1850
2350
LEAM (N = 12)
80%
60%
1350
0%
350
850
1350
1850
2350
350
850
1350
1850
2350
Wavelength (nm)
Reflectance
80%
TEOB (N = 7)
60%
40%
20%
0%
350
850
1350
1850
2350
Wavelength (nm)
Figure 5.1. Bark mean (bold line) and standard deviation (±1 S.D., thin line)
of simulated HYDICE reflectance for each species.
216
Candidate Endmember Spectra
Modeled Spectra
Sp. A
Sp. C
Sp. B
4 3 2 1 4 3 2 1 4 3 2 1
Sp. C
Sp. B
Sp. A
1 2 3 4 1 2 3 4 1 2 3 4
COBI
0.0 -0.50 0.25 0.13 0.0 0.25 0.50 0.75 -1.0 -0.75-0.50-0.25
EXAMPLE:
Species A, candidate endmember #3
COBIcon = 2/4 = 0.50
COBIextra = (1/4 + 1/4)/2 = 0.25
COBI = COBIcon – COBIextra = 0.50 – 0.25 = 0.25
Figure 5.2. Hypothetical depiction of COBI endmember selection. Four candidate
endmember spectra (columns) from three species (A, B & C) are randomly selected
from the image pixels. Each candidate endmember is then used with photometric
shade to model every other spectrum (rows) in a spectral mixture analysis.
Successful models (defined in Section 5.2.4) are black (conspecific) and gray
(extraspecific) boxes with values of 1, while unsuccessful models are white boxes
with values of 0. The COBI is calculated using successful models in columns (see
formula in Section 5.2.4). The best species endmember has the highest COBI.
217
Candidate Endmember Spectra
TEOB LEAM HYAL HYME DIPA CEPE BAEL
Modeled Spectra
BAEL CEPE DIPA HYME HYAL LEAM TEOB
2100 spectra
Dark
Bright
300 pixels
Figure 5.3. Matrix of spectral mixture analysis counts used in count-based COBI
endmember selection (Section 5.2.4). A total of 2100 randomly-selected pixels (300
per species) were used to model the same spectra. Valid models were marked as
black points and blank otherwise. Within a species, spectra were sorted from dark
(left) to bright (right) according to their 800-nm reflectance; and thus, brighter
candidate endmembers successfully model more spectra.
218
90%
TEOB EM1
Reflectance (800 nm)
80%
DIPA EM3
70%
DIPA EM1
60%
50%
40%
30%
20%
10%
0%
0%
2%
4%
6%
8%
10%
12%
Reflectance (680 nm)
40%
DIPA EM1
Reflectance (1650 nm)
35%
30%
25%
20%
TEOB EM1
15%
10%
5%
0%
0%
10%
20%
30%
40%
50%
60%
70%
80%
Reflectance (800 nm)
Figure 5.4. A) Near-infrared vs. red scatter plot for deciduous Dipteryx (DIPA;
closed circles) and evergreen Terminalia (TEOB; open circles) pixels from the
spectral library. Three candidate endmembers selected by the COBI selection
scheme are plotted as vectors (DIPA – black, TEOB – gray) connecting to
photometric shade. B) A near-infrared vs. shortwave-infrared scatter plot for the
same spectra and endmembers.
219
90%
800 nm
Classified
800 nm
BAEL
HYME
CEPE
LEAM
DIPA
TEOB
HYAL
Classified
EM other species
EM1
EM2
EM3
Background or not modeled
Figure 5.5. Correctly-labeled crowns for the seven study species using image
endmember and shade (1EM+Shade) MESMA with 3 optimal endmembers per
species. The 800-nm reflectance is shown for each crown. The endmembers
selected by the MESMA classifier are shown for the classified images. Conspecific
endmembers are gray-scale (1 through 3), while white pixels were modeled by
extraspecific endmembers. The background and non-modeled pixels are black.
220
CHAPTER 6: Comparison of lidar and hyperspectral data for tree
classification
6.1. Introduction
Tree structure from leaf to crown scales can vary greatly among species, and
these differences are a vital tool for visual interpretation of aerial photography.
Through a complex visual and cognitive process, photo-interpreters discriminate
individual tree crown (ITC) species using crown color and structural properties, such
as branch architecture, canopy position, contour shape, size, foliage cover and
texture (Fournier et al., 1995; Herwitz et al., 1998; Myers & Benson, 1981; Trichon,
2001). Visual interpretation of fine spatial scale image objects, such as trees, is a
time consuming, costly and often inconsistent process when dealing with
photographs spanning large areas, and so computer-based automated techniques for
crown detection, delineation and species classification are needed (Gougeon &
Leckie, 2003; Leckie, Gougeon, Hill et al., 2003). However, training a computer to
discriminate tree species using remotely-sensed digital data is a challenging task.
Optical sensors provide digital color information, and in the case of full-range
hyperspectral sensors (i.e., imaging spectrometers, 400-2500 nm), the spectral
information recorded by the sensor surpasses that available from human vision.
With high spatial resolution imagery (< 5 m), spectral information alone may be
adequate for automated tree species classification (Carleer & Wolff, 2004; Gougeon,
221
1995; Leckie et al., 2005; Wang et al., 2004; Xiao et al., 2004), especially if species
have distinct phenology (i.e., leaf cover, flowering) at the time of image acquisition
(Chapter 3) or through multiple image dates (Key et al., 2001).
Digital optical imagery also encodes properties of crown structure. In essence,
crown structure refers to the size and three-dimensional arrangement of its
components (i.e., leaves, trunk and branches). Because these components have
distinct absorptive, transmittive and reflective properties (e.g., chlorophyll and water
content in leaves, lignin and cellulose in bark) and block light to create within-crown
shadows, there is spectral variation across the crown (Kimes, 1983; Sandmeier et al.,
1998). If pixel spatial resolution is finer than the scale of a crown, then the spatial
arrangement of pixel spectra (e.g., texture) may respond to aspects of crown
structure. For example, the shadows cast on one side of a crown due to illumination
geometry is related to the crown’s overall structure, and shadows and exposed bark
within the crown may create finer-scale spectral variation. Few studies have used
pixel spatial information for automated ITC species discrimination. Meyer et al.
(1996) reported that a standard deviation metric calculated from a near-infrared
(NIR) band improved classification accuracy of conifers and hardwoods over that
achieved with spectral data alone. At patch or stand scales, however, there is
inconclusive evidence that image texture greatly improves forest composition
classification (Franklin et al., 2000; Leckie, Gougeon, Walsworth et al., 2003, Wang
et al., 2004; although see Franklin et al., 2001; Zhang, et al., 2004).
Small-footprint lidar sensors record the three-dimensional height distribution of
surface materials (Lefsky et al., 2002), and so these sensors are ideal for quantifying
222
ITC structure. Several studies have shown that ITC location, shape and area can be
automatically described with algorithms applied to lidar-derived digital canopy
models, or DCMs (Brandtberg et al., 2003; Holmgren & Persson, 2004; Leckie,
Gougeon, Hill et al., 2003; Persson et al., 2002). Various crown structure metrics
can be calculated from the height distribution of within-crown DCM cells or xyz
points, as well as shape and area metrics from the delineated ITC polygons. It has
been shown that lidar-based metrics are strongly correlated to field-measured ITC
height and diameter (Brandtberg et al., 2003; Holmgren & Persson, 2004; Persson et
al., 2002; Popescu et al., 2003; and Chapter 2).
The properties of crown structure described by lidar metrics are also useful for
tree species discrimination.
Working with three hardwood species in leaf-off
conditions, Brandtberg et al. (2003) calculated mean, standard deviation, skewness,
and kurtosis from the histograms of lidar height (normalized to crown maximum
height) and NIR reflectance from automatically-segmented ITCs. All metrics had
highly significant differences among the study species due to species-level variation
in vertical structure and branch distribution. The study achieved an overall accuracy
of 60% when using all variables in a linear discriminant analysis (LDA)
classification scheme. Holmgren and Persson (2004) classified two conifer species
with 95% accuracy using a suite of lidar-derived crown shape and NIR reflectance
intensity measurements.
The proportion of first lidar returns and the standard
deviation of intensity within the crown were the most important variables in the
classification. Increased variance in laser-return intensity was thought to be caused
by gaps within the crown.
223
Chapters 3 through Chapter 5 focused on spectral-based discrimination of ITC
species with hyperspectral and multispectral data. The objective of this chapter is to
extend these analyses to include lidar-derived, crown structure metrics calculated
from the DCM (Chapter 1). Similar to methods in Chapters 3 and 4, I investigated
decision trees and linear discriminant analysis for species classification.
6.2. Methods
6.2.1. Canopy-emergent trees
Analyses focused on the classification of canopy emergent individuals of six
target species and 18 non-target species (Table 6.1). Emergent trees with large,
exposed crowns provided a large sample of pixels that were less influenced by
spectral shadowing or scattering by neighboring trees and they were relatively easy
to locate in the hyperspectral and lidar data. In my previous Chapters, the species
Ceiba pentandra (CEPE) was included as a target species. For this Chapter, I
excluded this species as a target species because there were only 6 individuals within
both the hyperspectral and lidar data sets.
As explained in Chapter 3, some overstory tree species are completely
deciduous, generally beginning in the first dry season, while others are evergreen
and continuously flush small amounts of leaves throughout the year. The
hyperspectral imagery was acquired in March, at the end of the first dry season, and
the target trees DIPA and LEAM were expected to be leaf-off (Table 6.1). The
species HYAL and TEOB had evergreen crowns with broad leaves, while the
species BAEL and HYME had fine, sparsely-distributed compound leaves that gave
224
them more intermediate leaf area. For the hyperspectral image, I thus expected leaf
area index (LAI) for my target species to be low for DIPA and LEAM, moderate for
BAEL and HYME, and high for HYAL and TEOB. Of the non-target species, I did
not expect any leaf-off species at the time of the hyperspectral image acquisition
(Table 6.1). The lidar data were acquired in mid-September, during the second and
smaller dry season (Section 2.1.1). O’Brien (2001) found that of the HYME
individuals studied, roughly 60% of the crowns had >75% mature leaf cover in
September, 1997 (Table 6.1). With visual verification using videography acquired at
the time of the lidar flight, in contrast, I found that all HYME individuals were leafoff (Fig. 6.2). Other leaf-off trees included INAL and PTOF (Table 6.1; Fig. 6.2INAL). Although BAEL individuals have leaves in September (O’Brien, 2001;
verified in videography), their leaves are sparse and finely compound giving them a
low-LAI crown structure (Fig. 6.2-BAEL; Chapter 3).
In this chapter, I was particularly interested in mapping large Dipteryx (DIPA)
trees (Fig. 6.2). The population of large individuals of this species is in decline
across the Sarapiquí region due to deforestation. An important ecological function
of large individuals of this species is to provide seeds and nesting cavities for the
endangered Great green macaw (Ara ambigua) (pers. comm., Powell 2001). Remote
sensing technology that can identify large Dipteryx crowns may contribute to macaw
conservation efforts by providing a rapid and cost-effective means to map their
habitat and migration corridors across the region.
The 2-dimensional area of the tree crowns were manually digitized over the
HYDICE imagery (see Chapter 3). As in previous studies, I refer to digitized crown
225
polygons as individual tree crowns (ITCs). The HYDICE runs were orthorectified
using prominent tree crowns in the DCM as spatial control points (Chapter 3). Some
mismatch in spatial location still existed between the lidar and hyperspectral data
sets after HYDICE orthorectification. To minimize this error, ITCs that were
digitized on HYDICE imagery were overlaid on the DCM and visually repositioned
to align with the ITC apparent in the lidar data, thus creating a second ITC layer.
There were also slight differences in crown shape between data sets due to
differences in spatial resolution, phenology and other architectural changes (e.g.
branch fall) between September, 1997 and March, 1998. For some ITCs in the lidar
layer, polygon shapes were modified to conform to the crown shape observed in the
DCM.
6.2.2. Lidar metrics
I used DCM cells from ITCs to calculate a suite of lidar metrics that characterize
crown structure. The first set of metrics quantified crown height and size (Table
6.2). Crown area was the total 2-dimensional area of each ITC (Crownarea). Crown
width was calculated as the maximum 2D distance between pairs of cells within the
ITC (Crownwidth). Maximum crown height (Maxheight) was the highest DCM cell
within an ITC (Table 6.2).
I determined crown base height using an automated method outlined in
Holmgren and Persson (2004). First, a height-profile histogram with 0.5-m bins was
calculated for DCM cells within the ITC (maximum bin was 54.5 to 55.0 m). Next,
those bins whose count was <1% of the total ITC cell count where coded as 0, and 1
226
otherwise. The influence of understory vegetation was minimized by passing a 9point median filter along the histogram. The crown base height (Baseheight; Table
6.2) was then set as the first height bin with a value of 1 when moving along the
histogram from the 0.0-0.5 height bin upward. Applied in a conifer-dominated
forest in Sweden, this technique estimated crown base height with a 0.84 linear
correlation (RMSE = 2.82 m), with an overestimation of 0.75 m. Crown height was
calculated as:
Crownheight = Maxheight – Baseheight
(1)
Another suite of metrics was calculated from the relative height of DCM cells
within each crown. Those ITC cells that had a height value less than the crown base
height were removed from the calculation, thereby filtering out cells that did not
belong to the crown’s 3-dimensional volume (i.e., heights from understory
vegetation). Relative height was calculated by dividing each of the remaining ITC
cells by the ITC’s maximum height (Maxheight). The median (Relmed), mean
(Relmean), mean 10th-percentile (Rel10perc), mean 90th-percentile (Rel90perc),
standard deviation (Relstdev), kurtosis (Relkurtosis), and skewness (Relskewness)
were calculated from the relative-height cells within the ITC (Table 6.2; sensu
Holmgren & Persson, 2004).
Finally, I explored omnidirectional variograms to quantify the spatial properties
of relative heights within ITCs (Goovaerts, 1997). Empirical variograms for each
ITC were calculated by treating the 33-cm DCM cells as xyz points, where x and y
227
were the 2D spatial location and z was the relative height. In calculated variograms,
cells that were lower than the relative base height (i.e., not in the crown volume)
were set to the relative base height value. Variograms were calculated with 0.5, 1.5,
3.0 and 4.5-m lags, with a 0.25-m distance tolerance around each lag. The
semivariance values at each lag were then used in ITC classification analyses (Table
6.2).
6.2.3. Hyperspectral metrics
A suite of hyperspectral metrics were calculated for each ITC from its crownscale spectrum (see Chapter 4 for details). Metrics targeted key photosynthetic
pigment, water or other biochemical absorption features and included narrowband
indices, derivative-based metrics, absorption-based metrics, and Spectral Mixture
Analysis (SMA) fractions (Table 4.4).
6.2.4. Tree species classification techniques
I investigated two classifiers for ITC species classification: linear discriminant
analysis (LDA: Chapter 3) and decision trees (DT: Chapter 4). In Chapter 3, I used
LDA to classify ITC species with optimally-selected HYDICE reflectance bands.
Bands were selected using forward-stepwise selection based on discriminant
analysis. This method was implemented using the SAS STEPDISC procedure (SAS
Institute Inc., Cary, NC, USA). I found that the first 30 significant bands (α = 0.05)
from throughout the whole VIS to SWIR spectrum produced the best classification
228
accuracy for seven target species. In Chapter 4, I used decision trees to classify
species using the 77 spectral metrics.
Similar to methods in Chapters 3 and 4, classifiers were applied to ITCs using a
leave-one-out approach. That is, each ITC was classified by holding it out for
testing, while all remaining crowns were used for classifier training.
Several
variable bundles were analyzed separately to assess the relative benefits of lidar and
hyperspectral information for species identification.
These variable bundles were:
1. Reflectance bands for crown-scale spectra (“reflectance bands”).
2. The 77 spectral metrics from Chapter 4 for crown-scale spectra (“spectral
metrics”; Table 4.4).
3. The 17 structural metrics derived from lidar data (“lidar metrics”; Table 6.2).
4. Combined set of variables from lidar (4) and reflectance bands (1) datasets
(“lidar + reflectance bands”).
5. Combined set of variables from lidar (4) and spectral metrics (2) datasets
(“lidar + spectral metrics”).
Three different levels of classification were analyzed: 6 target species (BAEL,
DIPA, HYAL, HYME, LEAM, TEOB), 6 target species and an “other” (Other)
species class, and DIPA and Other species. For LDA, each variable bundle (1
through 3) was first submitted to a STEPDISC procedure (Chapter 3) with a 95%
confidence criteria (α=0.05) for each classification level.
229
Selected variables in
bundles 1 and 2 where then combined with selected variables in bundle 3 (lidar) to
create variable bundles 4 and 5.
In contrast to LDA, all variables within each bundle were used in DT
classification.
This is because DTs have the ability to rank the importance of
variables. If a variable is not useful for discriminating species, it will be left out of
the decision rules.
Decision tree classification was implemented using methods detailed in Chapter
4. As with LDA, the DT classifier was applied with leave-one-out cross-validation
using code in the R statistical package (R Development Core Team, 2004) that
interfaced with the Tree package (Tree v1.0-18, R v2.0). Parameters for growing
trees were “mincut” of 5, “minsize” of 10, “mindev” of 0.001, and “deviance” as the
criteria for splitting data into homogenous sets. Each ITC was classified 50 times
using 50 samples per species (with replacement) from the training set, and the final
ITC label was chosen using the majority class from the 50 DTs (see Chapter 4).
6.3. Results & Discussion
6.3.1. Differences in lidar metrics among species
There were significant differences among means of the six target species for 15
out of 17 lidar metrics (Table 6.3, ANOVA). Maximum tree height (Maxheight),
crown base height (Baseheight) and crown height (Crownheight) were top-ranked
metrics (Table 6.3; Fig. 6.3). Average tree height was 42 m for the six target
species. Individuals of TEOB were the tallest of the six species while BAEL trees
were the shortest, and variance was relatively low within each species (Fig. 6.3).
230
The semivariance at lag distances 0.5 and 1.5 m had moderate utility for species
discrimination (Table 6.3). The species HYME had dramatically higher
semivariance than the other species for 0.5 lags (Fig. 6.3). I observed the same
relative pattern in the other lags. There was greater laser penetration into leaf-off
HYME crowns relative to crowns from the other leaf-on target species. This effect
can be observed in Figure 6.4 for the HYME example crown, where relatively high
branches (brighter pixels) are juxtaposed with low, within-crown gaps (dark pixels)
due to low leaf cover. These sharp contrasts in height over short distances created
relatively high semivariance at all four lags (Table 6.4). In contrast, leaf-on crowns
had relatively smooth surfaces (Figs. 2 & 4, DIPA and LEAM) due to less laser
penetration, and semivariance values were relatively low (Fig. 6.3, all species except
HYME; Table 6.4, DIPA and LEAM). The standard deviation metrics (Relstdev)
did not capture the structural differences between leaf-on and leaf-off ITCs as
clearly as semivariance (Fig. 6.3; Table 6.3).
Individuals of TEOB were generally the tallest of the target species, while DIPA
and HYME had the widest and largest crowns (Fig. 6.3, Maxheight, Crownwidth,
Crownarea). Individuals of HYME also had very deep crowns (Fig. 6.3, large
Crownheight values). This is because their base heights were relatively low while
their maximum heights were relatively high (see Eq. 1). As with semivariance
discussed above, the sharp contrast in HYME crown height is related to the leaf-off
state of the species during the lidar flight. High laser penetration into the crown
shifted more DCM cell counts into the lower bins of the crown-height histogram,
which tended to lower the estimated crown base height (Fig. 6.4). Providing that the
231
maximum crown height is unaffected by leaf phenology (i.e., the laser detects the
top-most branch), then lower leaf cover will tend to increase crown height as base
height goes down (i.e., deeper into crown). In contrast, height profiles from leaf-on
species such as DIPA and LEAM tended to have most cell counts in the upper-height
bins due to relatively low laser penetration (Fig. 6.4); and therefore, their crown base
heights were shallow relative to their maximum crown heights; and subsequently,
their calculated crown heights were relatively low (Fig. 6.4; Table 6.4). I do not
have field data to assess how well base or crown heights were predicted. However,
what is important for my analyses is that metrics detect species differences in crown
structure as mediated by leaf phenology. The Crownheight metric thus appears quite
useful for this purpose.
Leaf-off HYME crowns had the lowest (more negative) mean Relkurtosis and
highest (less negative) mean Relskewness values (Fig. 6.3). Negative kurtosis
indicates a platykurtic distribution of heights, where there are more values than
would be expected for a normal distribution (Zar, 1996). For HYME, kurtosis was
exaggerated by deeper laser penetration that resulted in more lower-crown versus
upper-crown heights (Fig. 6.4 & Table 6.4: HYME). Negative skewness values
indicate a distribution that is skewed to the left, with a longer tail toward the relative
base height. For the three example crowns in Figure 6.4, laser penetration created a
distribution that was less skewed for HYME and DIPA individuals relative to the
LEAM individual, which had a strong negative Relskewness (Table 6.4).
232
6.3.2. Decision-tree classification
When considering 6 target species, lidar metrics alone (bundle 3) produced a
41.1% overall accuracy. The lidar and spectral metrics (bundle 5) produced the
highest overall accuracy, at 69.6%. However, this was only 2.0% more accurate
than with spectral metrics alone (Table 6.5; Z = 0.51, not significant; Congalton,
1991). The full-spectrum reflectance bands with and without lidar (bundles 1 & 4)
had poorer classification accuracy than when using spectral metrics.
I next included an Other species class in the decision trees. This class
encompassed all 41 non-target ITCs (18 different species; Table 6.1) and was
expected to have more variation in predictor variables relative to target species.
With these 7 classes (6 target species + Other class), again the highest accuracy
achieved was with lidar metrics combined with spectral metrics (Table 6.5). This
overall accuracy, 62.9%, was a 3.6% increase over not using lidar metrics (Z = 0.76,
not significant). The Producer’s accuracy (omission error) and User’s accuracies
(commission error) of the Other class for all classifications were below 50%,
indicating a high level of confusion among non-target and target species.
The decision tree was next limited to only DIPA (n=86) and the Other-species
class, now comprising ITCs from all non-DIPA species (n=162; Table 6.1). The
classification that included only lidar metrics (bundle 3) produced an overall
accuracy of 66.9% (Table 6.5). Overall accuracy improved by 17.4% when the
analysis was limited to spectral metrics (bundle 2), and DIPA User’s accuracy was
72.8%. Combining lidar metrics with the spectral metrics (bundle 5) did not
improve the classification accuracy.
233
These results are similar to those in Chapter 4, in which spectral metrics
outperformed reflectance bands in DT classification. The analysis focusing on 6
target species is similar to our previous analysis of 7 species, which achieved an
overall accuracy of 70.1%. Overall accuracy with the 6 species was less (67.6%),
even though it did not include CEPE, which was poorly classified in the sevenspecies analysis. This discrepancy is likely due to the difference in ITC data
between the studies (this study had more DIPA and BAEL, and less LEAM and
TEOB). My analyses with the Other class and the addition of lidar metrics are new
for this study. In general, accuracies from decision trees are disappointing, and were
only acceptable for the DIPA vs. Other classification with spectral metrics.
6.3.3. Linear discriminant analysis classification
There were five selected lidar metrics for discriminating the 6 target species:
Maxheight, Relmean, Crownarea, Relperc90, and Relkurtosis. The Crownarea,
Relperc90 and Relkurtosis metrics had relatively low ranks in the ANOVA results
(Table 6.3). This is because the LDA selection finds the combination of variables
that best discriminate classes (Tabachnick & Fidell, 1989), while ANOVA assesses
whether groups have different means in a single response variable. The stepwiseselected lidar metrics (bundle 3) classified the 6 target species with only 45.9%
overall accuracy (Table 6.6). However, this was 4.8% more accurate than when
using all lidar metrics in decision trees (Table 6.5). The 34 selected reflectance
bands (bundle 1) produced the highest accuracy for the 6 target species (Table 6.6),
234
while classification accuracy was 0.5% less with the addition of the 5 lidar metrics
(Z = 0.01, not significant).
With the 34 reflectance bands, there was confusion between BAEL and DIPA,
and HYME was confused with LEAM, DIPA, and BAEL (Table 6.7). As explained
in Chapter 4, leaf-off DIPA and LEAM were confused with BAEL and HYME in the
hyperspectral imagery because BAEL and HYME had low LAI and exposed bark
due to their compound leaves, giving them spectral properties similar to each other
and to leaf-off trees (e.g., Fig. 6.2, BAEL). The addition of lidar metrics decreased
confusion between DIPA and BAEL, and between DIPA and HYME. The structural
properties of leaf-off (HYME) and low LAI (BAEL) crowns captured by the lidar
metrics thus aided in discriminating these species from DIPA, which had lidarstructure properties of a high-LAI crown (Figs. 2 & 4). For HYME, the addition of
lidar data improved Producer’s and User’s accuracies by 7.1% and 10.0%,
respectively. However, the confusion among species that had high-LAI in lidar
(DIPA, HYAL and LEAM) tended to increase because these species had similar
crown structure (Figs. 2 & 4). The addition of lidar may thus provide key structural
information for discriminating a particular species with distinct phenology (i.e., leafoff in this case), yet it may add to confusion among species with similar structural
properties.
When the Other species class was included in the LDA classifier, the best
accuracy was 81.9% with 41 reflectance bands and 3 lidar metrics (Table 6.6).
Overall accuracy was only 0.4% lower without the lidar metrics (Z=0.01, not
significant). Lidar metrics alone could only classify species with 36.7% overall
235
accuracy. In general, the User’s and Producer’s accuracy of the DIPA and Other
classes were higher with LDA relative to the DT classifier.
The final LDA analysis focused on classifying DIPA from the other species, as
was done with the DT analysis (Section 3.2). Two significant lidar metrics (bundle
3), Baseheight and Crownarea, were able to classify these two classes with 68.5%
overall accuracy (Table 6.6). However, DIPA Producer’s accuracy was very low. In
contrast to the other LDA classifications, spectral metrics and lidar metrics (bundle
5) produced the highest overall accuracy, at 90.7%, and DIPA Producer’s and User’s
accuracies were 86.0% and 87.1%, respectively. Removing the 2 lidar metrics only
lowered the accuracy by 0.04% (not significant) and caused DIPA Producer’s and
User’s accuracy to decrease 1.1% and 0.2%, respectively (Table 6.6). The 9
significant spectral metrics from LDA stepwise selection were (in order of
importance): Red-A2, ARVI, YE-DArea, NIR2-λ, SR, RE-λ, NE1-Mag, NPV, and
SWIR3-A2. There were 11 ITCs committed to the DIPA class (Table 6.8), which
decreased User’s accuracy. These crowns included 3 BAEL, 2 CEPE, 1 HYAL, 1
HYME, and 4 LEAM.
Another approach to map DIPA crowns was to classify the target species with
the Other species class, and then combine the classified non-DIPA target species into
one class with the other species. However, this approach did not substantially
improve the Producer’s and User’s accuracy of DIPA (Table 6.6). I therefore
concluded that the best method for discriminating DIPA from other species was to
focus the LDA classifier on two class—DIPA vs. all other species.
236
I also experimented with a probability threshold (see Chapter 3) on the LDA
classifier with spectral metrics (bundle 2). In this classification scheme, those ITCs
that had a DIPA class probability less than the threshold were left in the Other class.
Probability thresholds included 50, 60, 70, 80 and 90 percent. As I increased the
probability threshold, DIPA Producer’s accuracy declined because questionable
ITCs (i.e., low DIPA probability) were omitted from the DIPA class (Fig. 6.5).
However, the stricter probability thresholds tended to increase the DIPA User’s
accuracy because fewer non-DIPA crowns were committed to the DIPA class.
Overall accuracy remained relatively stable up to the 70% probability threshold.
This threshold gave high DIPA User’s accuracy (94.1%; Table 6.9), ensuring that
those DIPA in the map were likely to be DIPA in the field. The DIPA Producer’s
accuracy dropped to 74.4% with the seventy-percent threshold (Table 6.9),
indicating that not all DIPA were mapped (increased omission error). The
misclassified DIPA crowns included 2 CEPE, 1 HYME and 1 LEAM.
To show the potential of remote sensing for ecological applications, I made a
final map of DIPA crowns using the seventy-percent threshold (Fig. 6.6). When
compared to the DTM derived from the lidar data (Chapter 2), it can be seen that
DIPA were distributed on the edges of low-elevation landforms (old alluvial
terraces) and away from water drainages. Similar results are described in Clark et al.
(1998), which is not surprising since many of the DIPA crown locations came from
data in that study. A true depiction of DIPA crown distributions would be to
delineate every tree crown in the HYDICE image through automated means and then
classify the crowns as DIPA or Other. The non-random distribution of DIPA in
237
Figure 6.6 is intriguing from both an ecological and remote sensing perspective.
Many TRF species have non-random distributions relative to soil water and nutrient
gradients (Clark et al., 1998; Condit et al., 2002; Tuomisto, Ruokolainen et al.,
2003). If these factors can be mapped economically with remote sensing, then they
may provide additional information for species classification (e.g., Franklin, 1998).
For example, the lidar-derived DTM could be used to model soil moisture, which
could then be incorporated into the classification scheme. However, such an
analysis is beyond the scope of this study.
6.4. Conclusions
Species leaf phenology is a dominant factor in discriminating tropical rain forest
ITCs with either hyperspectral or lidar data. Chapters 3 through Chapters 5
described how ITC leaf cover affected the spectral response of my target species.
Here I focused my attention on the contribution of lidar data to species
discrimination. Lidar sensors measure the 3-dimensional height distribution of
materials within the crown, and thus lidar metrics can be used to quantify crown
structure. Deciduous leaf-off species had distinct architectural properties relative to
those species with fully-flushed crowns, and these differences were observed in
several lidar metrics, such as semivariance and crown height.
I found that lidar data alone was insufficient to adequately classify the study
species with neither the DT nor LDA classifier. This is likely because the lidar data
were acquired in the second dry season (September), when leaf phenological
variation among species was relatively low. Combined with the hyperspectral data
238
in the LDA classifier, lidar metrics were mostly useful for improving the accuracy of
Hymenolobium, a species that was leaf-on in the hyperspectral data, yet leaf-off in
the lidar data; however, the leaf-on species tended to have similar crown
architecture, as measured by lidar metrics, and their inter-species confusion
increased relative to that from hyperspectral data. In general, there is much more
variation in leaf phenology in the primary dry season, between January and April
(Table 6.1), when the hyperspectral data were acquired. It is unclear if lidar data
acquired in the primary dry season would have improved ITC species discrimination.
I conclude that hyperspectral data alone were generally adequate for classifying
the target species. In particular, Dipteryx could be classified with 84.9% Producer’s
and 86.9% User’s accuracy with LDA and optimal reflectance bands. However, my
classification methods require crown-scale spectra from segmented crowns (i.e.,
polygons). Operational ITC classification will require an automated crown
delineation algorithm (Chapter 1). Lidar data may be most useful for crown
delineation (Brandtberg et al., 2003; Leckie, Gougeon, Hill et al., 2003; Persson et
al., 2002), perhaps in combination with optical data to help distinguish near crowns
of similar height. However, lidar-based crown delineation will need to
accommodate the variable heights within leaf-off crowns (e.g., Brandtberg et al.,
2003).
Other types of ecological applications may benefit from the biochemical and
structural information offered by hyperspectral and lidar sensors, respectively. One
example is estimating aboveground biomass. Lidar data can provide relatively
accurate estimates of total aboveground biomass (Drake, Dubayah, Clark, et al.,
239
2002), or carbon stocks, across a tropical landscape. Hyperspectral data responds to
changes in canopy biochemistry associated with leaf phenology, and thus it may be
useful for estimating carbon flux, such as the percentage of canopy leaf turnover
through time.
240
Table 6.1. Individual tree crown species and counts. Bold indicates target species.
Tree phenology was unknown, evergreen or deciduous (months). For deciduous
species, the months indicate when a large percentage of individuals have low leaf
area (e.g., leaf-off). For some species, phenology was estimated using data on
another species of the same genus.
Species Name
Albizia sp.
Balizia elegans
Carapa nicaraguensis
Cecropia insignis
Cedrela odorata
Ceiba pentandra
Dipteryx panamensis
Dussia macroprophyllata
Hyeronima alchorneoides
Hymenolobium mesoamericanum
Inga alba
Lecythis ampla
Luehea seemannii
Minquartia guianensis
Ocotea hartshorniana
Pentaclethra macroloba
Pterocarpus officinalis
Sacoglottis trichogyna
Simarouba amara
Stryphnodendron microstachyium
Tachigali costaricensis
Terminalia oblonga
Vochysia ferruginea
Vochysia guatemalensis
a
O’Brien, 2001.
b
Frankie et al., 1974.
c
Visual verification from videography
Abbrevatio
n
ALSP
BAEL
CANI
CEIN
CEOD
CEPE
DIPA
DUMA
HYAL
HYME
INAL
LEAM
LUSE
MIGU
OCHA
PEMA
PTOF
SATR
SIAM
STMI
TACO
TEOB
VOFE
VOGU
241
Phenology
Count
1
41
5
1
1
6
86
2
26
14
3
17
1
1
1
7
1
2
3
3
1
23
1
1
Unknown
Novac
Evergreenbc
Evergreenbc
Jan-Febbc
Jan-Marbc
Mar-Mayac
Janbc
Evergreenbc
May-Febac
Septc
Mar-Juneac
Maybc
Evergreenbc
Evergreenb
Evergreenbc
Septc
Evergreenbc
Evergreenabc
Jan-Febbc
Unknownc
Evergreenbc
Evergreenbc
Evergreenbc
Table 6.2. Summary of lidar metrics organized by methods.
Height & Size
Crownarea (m2)
Crownwidth (m)
Maxheight (m)
Baseheight (m)
Crownheight (m)
Relbaseheight
Relmedian
Relmean
Relperc10
Relperc90
Relstdev
Relkurtosis
Relskewness
Semivariance
Lag_0.5
Lag_1.5
Lag_3.0
Lag_4.5
242
Table 6.3. Lidar metric ANOVA results for 6 target species (BAEL, DIPA,
HYAL, HYME, LEAM and TEOB), total individual tree crown n=207.
Metrics are ranked primarily by F statistic and secondarily by the number of
significant differences between species pairs (15 total comparisons).
Significance levels for ANOVA F statistic: ns = not significant, * = p≤0.05, **
= p≤0.01, *** = p≤0.001, **** = p≤0.0001. Bold indicates those metrics used
in the LDA classification.
Rank
Metric
Mean
Std. Dev.
Sig. Pairs
F
1
Maxheight
42.0
6.0
13.2 ****
7
2
Baseheight
27.4
5.3
10.1 ****
6
3
Crownheight
14.6
4.6
8.9 ****
6
4
Relmean
0.85
0.05
8.4 ****
5
5
Lag_0.5
0.005
0.004
8.2 ****
5
6
Relmedian
0.86
0.05
8.2 ****
5
7
Lag_1.5
0.007
0.006
7.8 ****
5
8
Relperc10
0.69
0.09
6.9 ****
5
9
Lag_3.0
0.008
0.006
6.6 ****
5
10
Relbaseheight
0.65
0.09
6.5 ****
5
11
Lag_4.5
0.010
0.007
5.9 ****
4
12
Relstdev
0.082
0.027
5.8 ****
5
13
Crownarea
372.0
181.6
3.6 **
1
14
Crownwidth
24.7
6.5
3.1 **
1
15
Relskewness
-0.47
0.34
3.1 **
2
16
Relperc90
0.97
0.03
2.3 ns
0
17
Relkurtosis
-0.44
1.00
0.7 ns
0
243
Table 6.4. Lidar metrics for example individual tree
crowns shown in Fig. 6.3. Species abbreviations in Table
6.1. Lag semivariance is multiplied by 100.
LEAM
DIPA
HYME
Metric
Tree #91
Tree #96
Tree #97
Crownarea
360.9
620.1
958.3
Crownwidth
24.4
29.8
38.0
Maxheight
42.7
48.6
50.9
Baseheight
28.5
32.5
24.0
Crownheight
14.2
16.1
26.9
Relbaseheight
0.67
0.67
0.47
Relmean
0.85
0.84
0.73
Relperc90
0.96
0.95
0.92
Relperc10
0.71
0.70
0.51
Relstdev
0.07
0.07
0.12
Relkurtosis
-0.52
-0.58
-0.70
Relskewness
-0.50
-0.33
-0.35
Relmedian
0.86
0.85
0.76
Lag_0.5
0.24
0.50
1.28
Lag_1.5
0.34
0.62
1.67
Lag_3.0
0.46
0.72
1.89
Lag_4.5
0.57
0.80
2.04
Number of
DCM pixels
3239
5570
8643
244
Table 6.5. Decision tree classification accuracy. Kappa variance multiplied by 100.
Sensor combination
reflectance bands
spectral metrics
lidar metrics
lidar + reflectance bands
lidar + spectral metrics
245
reflectance bands
spectral metrics
lidar metrics
lidar + reflectance bands
lidar + spectral metrics
reflectance bands
spectral metrics
lidar metrics
lidar + reflectance bands
lidar + spectral metrics
No. DIPA DIPA Other Other
Vars Prod User Prod User
6 Target Species
161 58.1 71.4
n/a
n/a
77
77.9 82.7
n/a
n/a
17
48.8 67.7
n/a
n/a
178 66.3 69.5
n/a
n/a
94
81.7 77.9
n/a
n/a
6 Target and Other Species
161 57.0 68.1 26.8
47.8
77
73.3 79.7 34.1
45.2
17
51.2 65.7 19.5
26.7
178 64.0 71.4 34.1
42.4
94
76.7 77.6 39.0
48.5
DIPA and Other
161 81.4 62.5 74.1
88.2
77
87.2 72.8 82.7
92.4
17
67.4 51.8 66.7
79.4
178 87.2 62.0 71.6
91.3
94
86.0 73.3 83.3
91.8
245
Overall
Kappa
Kappa
Var.
55.1
67.6
41.1
58.0
69.6
0.42
0.57
0.25
0.45
0.60
0.19
0.17
0.18
0.20
0.17
48.4
59.3
35.5
52.8
62.9
0.37
0.50
0.21
0.42
0.54
0.14
0.14
0.12
0.14
0.14
76.6
84.3
66.9
77.0
84.3
0.52
0.67
0.32
0.54
0.67
0.31
0.24
0.39
0.29
0.24
246
Table 6.6. Linear discriminant analysis classification accuracy. Bands were selected using
stepwise linear discriminant analysis with statistical significance set at α =0.05. Kappa
variance multiplied by 100.
No. DIPA DIPA Other Other
Kappa
Sensor combination
Vars Prod User Prod User Overall Kappa Var.
6 Target Species
reflectance bands
34 90.7 87.6
n/a
n/a
88.9
0.85
0.09
spectral metrics
24 86.0 91.4
n/a
n/a
83.6
0.78
0.12
lidar metrics
5 47.7 72.1
n/a
n/a
45.9
0.22
0.26
lidar + reflectance bands
39 89.5 90.6
n/a
n/a
88.4
0.85
0.09
lidar + spectral metrics
29 83.7 91.1
n/a
n/a
83.6
0.78
0.12
6 Target and Other Species
reflectance bands
41 86.0 84.1 61.0
67.6
81.5
0.77
0.10
spectral metrics
27 81.4 84.3 51.2
67.7
76.6
0.71
0.11
lidar metrics
3 76.7 39.5 17.1
23.3
36.7
0.12
0.21
lidar + reflectance bands
44 84.9 86.9 63.4
72.2
81.9
0.77
0.09
lidar + spectral metrics
30 83.7 85.7 53.7
64.7
79.8
0.75
0.10
DIPA and Other
reflectance bands
6 80.2 83.1 91.4
89.7
87.5
0.72
0.22
spectral metrics
9 84.9 86.9 93.2
92.1
90.3
0.79
0.17
lidar metrics
2 26.7 60.5 90.7
70.0
68.5
0.20
0.01
lidar + reflectance bands
8 77.9 82.7 91.4
88.6
86.7
0.70
0.23
lidar + spectral metrics
11 86.0 87.1 93.2
92.6
90.7
0.79
0.17
246
Classification
Table 6.7. Error matrix for LDA classification with 34 reflectance bands without and with 5
lidar metrics.
Classification
247
Species
BAEL
DIPA
HYAL
HYME
LEAM
TEOB
Total
Prod.
Species
BAEL
DIPA
HYAL
HYME
LEAM
TEOB
Total
Prod.
Reflectance Bands
Field Reference
HYAL HYME LEAM
TEOB
1
2
3
2
24
9
1
1
14
23
26
14
17
23
92.3% 64.3% 82.4% 100.0%
BAEL
36
4
1
41
87.8%
DIPA
5
78
1
2
86
90.7%
BAEL
36
3
2
41
87.8%
Reflectance Bands + Lidar Metrics
Field Reference
DIPA HYAL HYME LEAM
TEOB
4
1
2
77
2
3
2
23
10
3
2
14
23
86
26
14
17
23
89.5% 88.5% 71.4% 82.4% 100.0%
247
Total
42
89
26
10
17
23
207
87.8%
User
85.7%
87.6%
92.3%
90.0%
82.4%
100.0%
Total
43
85
27
10
19
23
207
User
83.7%
90.6%
85.2%
100.0%
73.7%
100.0%
88.9%
88.4%
Classification
Table 6.8. Error matrix for Dipteryx individual tree crown classification using 9
stepwise-selected spectral metrics and LDA with no threshold (Kappa = 0.79).
Reference
Species
DIPA Other
Total
User’s
DIPA
73
11
84
86.9%
Other
13
151
164
92.1%
Total
86
162
248
Producer’s
84.9% 93.2%
90.3%
Classification
Table 6.9. Error matrix for Dipteryx individual tree crown classification using 9
stepwise-selected spectral metrics and LDA with a 70-percent threshold (Kappa =
0.76).
Reference
Species
DIPA Other
Total
User’s
DIPA
64
4
68
94.1%
Other
22
158
180
87.8%
Total
86
162
248
Producer’s
74.4% 97.5%
90.5%
248
Figure 6.1. The La Selva Biological Station study site and extent of HYDICE
hyperspectral and FLI-MAP lidar datasets. The 248 study crowns are labeled with
black polygons.
249
BAEL
DIPA
Tree#108
Tree#58
HYAL
HYME
Tree#56
Tree#247
LEAM
TEOB
Tree#127
Tree#25
CEPE
INAL
Tree#162
Tree#197
Figure 6.2. Example individual tree crowns from color videography acquired
simultaneously with the lidar data. Video time codes are UTC. Species codes are
listed in Table 6.1.
250
40
50
Baseheight (m)
Maxheight (m)
60
40
30
20
10
0
30
20
10
0
BAEL DIPA HYAL HYME LEAM TEOB
BAEL DIPA HYAL HYME LEAM TEOB
Crownarea (m 2)
Crownheight (m)
30
25
20
15
10
5
0
700
600
500
400
300
200
100
0
35
30
25
20
15
10
5
0
BAEL DIPA HYAL HYME LEAM TEOB
1.0
0.8
Relmean
Crownwidth (m)
BAEL DIPA HYAL HYME LEAM TEOB
0.6
0.4
0.2
0.0
BAEL DIPA HYAL HYME LEAM TEOB
0.15
2.0
1.5
Relstdev
Lag_0.5 Semivariance
BAEL DIPA HYAL HYME LEAM TEOB
1.0
0.5
0.10
0.05
0.00
0.0
BAEL DIPA HYAL HYME LEAM TEOB
BAEL DIPA HYAL HYME LEAM TEOB
BAEL DIPA HYAL HYME LEAM TEOB
0.00
-0.5
-0.25
-1.0
-1.5
-2.0
Relskewness
Relkurtosis
BAEL DIPA HYAL HYME LEAM TEOB
0.0
-0.50
-0.75
-1.00
Figure. 6.3. Species mean (bars) and standard deviation (error bars) for selected lidar
metrics calculated from target-species tree crowns (n=207). Species codes are listed
in Table 6.1. Semivariance is multiplied by 100.
251
LEAM Tree #91
DIPA Tree #96
HYME Tree #97
Height (m)
DIPA – Tree # 96
Height (m)
50
48.6 m
40
16.1 m
50.0
25
Meters
0.0
32.5 m
30
20
Maxheight
HYME – Tree #97
10
50 100 150 200 250 300
Height (m)
Count
LEAM – Tree #91
50.9 m
40
26.9 m
30
24.0 m
20
50
42.7 m
Height (m)
40
Crownheight
0
50
Baseheight
10
14.2 m
30
28.5 m
0
50 100 150 200 250 300
Count
20
10
0
50 100 150 200 250 300
Count
Figure 6.4. A subset of the digital canopy model (DCM) with three example
individual tree crowns. Graphs depict the vertical height profile for each crown with
count of DCM pixels in each height 0.5-m height bin. The horizontal dashed line in
each graph is the maximum crown height and the solid line is the base height
(Baseheight), as calculated by the automated algorithm. The value between the bars
is the crown height (Eq. 1).
252
100
90
Percent
80
70
DIPA User's
60
DIPA Producer's
50
Other User's
40
Other Producer's
30
Overall
20
10
0
None
50%
60%
70%
80%
90%
LDA Probability Threshold
Figure 6.5. Overall accuracy and User’s and Producer’s accuracy for DIPA
(Dipteryx) and other species (Other) with change in linear discriminant analysis
probability threshold.
Figure 6.6. Map of DIPA (Dipteryx) crowns using the LDA, 70-percent threshold
classifier (Table 6.9). Four DIPA crowns were incorrectly classified.
253
CHAPTER 7: Conclusions
7.1. Summary of research
Overview
The primary goal of this research was to assess two types of emerging remote
sensing technology, hyperspectral and lidar sensors, for the automated, species-level
mapping of individual tropical rain forest trees. Such maps will be useful for broadscale forest inventory, prioritization for protection, long-term monitoring,
management and ecological analyses. The study was conducted in a species-rich
tropical wet forest in Costa Rica and focused on emergent individuals of 7 out of
more than 400 tree species. The findings presented here are therefore not
exhaustive, but rather represent a first attempt to conduct automated species
classification with these new types of remotely-sensed data.
The hyperspectral measurements respond to the biochemical and structural
properties of crowns, while the lidar measurements respond to crown structure. The
central question guiding this research was: can species could be discriminated based
on their spectral or structural properties? Does the combined biochemical and
structural information offered by both lidar and hyperspectral sensors improve
species classification accuracy? To answer these questions, I outlined several
objectives in Section 1.2:
254
1. Develop hyperspectral techniques for classifying tropical rain forest tree species
Five different hyperspectral classification techniques were investigated (Chapters
3-5). The linear discriminant analysis produced the best accuracy when presented
with an optimal set of 30 reflectance bands selected from across the full 400 to 2500
nm spectrum. This technique can be implemented using existing statistical software
packages.
2. Identify the optimal spectral regions and spatial scale for species discrimination
Initial species classification analyses using hyperspectral data spanned leaf, pixel
to crown scales. Although the leaf scale had accuracies reaching 100%, it was not a
realistic assessment of the technology in an operational setting because spectra
lacked atmospheric noise and spectral mixing. At operational scales, species were
generally best discriminated using crown-scale relative to pixel-scale spectra,
indicating that very high spatial resolution imagery is not needed for species
discrimination. Leaf phenology was a major factor that caused species-level
differences in spectral absorption features. The visible, near-infrared and shortwave
infrared regions were all useful for detecting this phenological spectral variation.
3. Evaluate the importance of lidar-derived crown structure information for species
discrimination
Many properties of crown structure measured from the lidar dataset were
influenced by tree leaf phenology. However, most species were leaf-on at the time
of the lidar acquisition, and there was relatively low variation in lidar-derived
255
metrics among leaf-on species for adequate species discrimination. The
hyperspectral data, which responded to biochemical and structural properties of
crowns at a time of high phenological contrast, were sufficient for distinguishing
species.
4. Assess lidar technology for ecological analyses of tropical rain forests
Although lidar-derived crown structure was not sufficient for species
discrimination, the technology did prove useful for other types of ecological
analysis. The terrain surface, generated as a by-product of vegetation analyses, had
remarkable detail and impressive accuracy, even below structurally-complex oldgrowth forest. The lidar sensor was also promising for scaling plot-scale forest
structure parameters (e.g., height) to landscape scales.
Chapter 2 summary
Chapter 2 involves the pre-processing of the small-footprint lidar dataset for later
analyses of individual tree crown (ITC) structure at LSBS (Chapter 6). A fullyautomated, local-minima algorithm was developed to separate lidar ground returns
from overlying vegetation returns in the original lidar height surface. The IDW and
OK geostatistical techniques were then used for interpolating a sub-canopy DTM.
OK was determined to be a superior interpolation scheme because it smoothed finescale variance created by spurious understory heights in the ground-point dataset.
The final DTM had a strong linear-correlation of 1.00 and a RMSE of 2.29 m when
compared against 3859 well-distributed ground-survey points. In old-growth forests,
256
RMSE on steep slopes was 0.67-m greater than on flat slopes. On flatter slopes,
variation in vegetation complexity associated with land-use caused highly significant
differences in DTM error distribution across the landscape. The highest DTM
accuracy observed in this study was on flat, open-canopy areas with relatively
smooth surfaces. Lidar ground-retrieval was complicated by dense, multi-layered
evergreen canopy in old-growth forests, causing DTM overestimation.
A DCM was calculated by subtracting the DTM from the original lidar surface.
Individual and plot-scale heights were estimated from DCM metrics and compared
to field data measured using similar spatial supports and metrics. For old-growth
forest emergent trees and isolated pasture trees, individual tree heights were
underestimated and had 3.67 and 2.33-m mean absolute error, respectively. Linearregression models explained 51% (4.15-m RMSE) and 95% (2.41-m RMSE) of the
variance, respectively. It was determined that improved elevation and field-height
estimation in pastures explained why individual pasture trees could be estimated
more accurately than old-growth trees. Mean height of tree stems in 32 young
plantation plots (0.38 to 18.53-m tall) was estimated with a mean absolute error of
0.90 m (r2=0.97; 1.08-m model RMSE) using the mean of lidar returns in the plot.
As in other small-footprint lidar studies, plot mean height was underestimated;
however, the plot-scale results from this analysis had stronger linear models than
previously-reported models for temperate-zone conifer and deciduous hardwoods.
257
Chapter 3 summary
Chapter 3 investigates the utility of high spectral and spatial resolution imagery
for the automated species-level classification of ITCs in the LSBS old-growth forest.
Laboratory spectrometer and airborne reflectance spectra (161 bands, 437-2434 nm)
were acquired from seven species of emergent trees. Analyses focused on leaf-,
pixel- and crown-scale spectra. The spectral regions and factors that most
influenced spectral separability among species were reviewed. Next, spectral-based
species classification was performed using traditional classifiers, spectral angle
mapper (SAM), maximum likelihood (ML) and linear discriminant analysis (LDA),
in combination with a stepwise band selection procedure. Optimal regions of the
spectrum for species discrimination varied with scale. However, near-infrared (7001327 nm) bands were consistently important regions across all scales. Bands in the
visible region (437-700 nm) and shortwave infrared (1994-2435 nm) were more
important at pixel and crown scales. Overall classification accuracy decreased from
leaf scales measured in the laboratory to pixel and crown scales measured from the
airborne sensor. The highest crown-scale ITC accuracy was 92% with LDA and 30
bands. Producer’s accuracies ranged from 70% to 100% and User’s accuracies
ranged from 81% to 100%. The SAM classifier performed poorly at all scales and
spectral regions of analysis.
ITCs were also classified using a pixel-majority approach in which crown
species labels were assigned according to the majority class of classified pixels
within a crown. An overall accuracy of 86% was achieved with a pixel-majority
LDA classifier applied to 30 bands of data. Pixel-majority and crown-scale ITC
258
classifications were significantly more accurate with 10 narrow-bands relative to
accuracies achieved with simulated multispectral, broadband data.
Chapter 4 summary
Chapter 4 approaches ITC classification with a less conventional technique.
Classification variables included hyperspectral metrics that responded to crown
structure and absorption features from photosynthetic pigments, water and other
biochemicals. The metrics included narrowband indices, derivative-based metrics,
absorption-based metrics and spectral mixture analysis fractions that were calculated
from spectra acquired at tissue (leaf, bark), pixel, and crown scales. Differences in
metrics among species were ranked using statistical tests. Leaf and pixel-scale
spectra were best discriminated by near-infrared water absorption features while
bark and crown-scale spectra were better distinguished by shortwave infrared
biochemical absorption features. Differences in spectral metrics among species at
pixel and crown scales were largely dependent on tree leaf phenology and structure,
which controlled the relative amounts of leaf and bark tissues within a crown. ITC
species were classified with a decision tree (DT) classifier using within-crown pixel
spectra and a pixel-majority classification rule or with crown-scale spectra. The best
classification scheme was with crown-scale metrics and had 70.1% overall accuracy.
Hyperspectral metrics and DT classifiers were instructive for identifying key
spectral reflectance properties for tropical tree species discrimination. In terms of
number of times appearing in DTs, important metrics characterized absorption
259
features across the whole 437-2435-nm spectrum, but particularly in the shortwave
infrared region (1467-2435 nm).
Chapter 5 summary
Chapter 5 explores multiple endmember spectral mixture analysis (MESMA) for
classifying ITCs. To increase computational efficiency of MESMA, an automated
technique was developed to select optimal endmembers for each species from image
and laboratory libraries. Candidate endmembers were used to model other spectra in
the library, and endmembers that maximized within-species modeling capability
where chosen over more generalist endmembers.
The species of ITCs were determined by applying the MESMA classifier to
crown-scale spectra or to pixel-scale spectra, followed by a pixel-majority
classification. MESMA classification with two- and three-endmember models was
explored. With a single endmember per species and a shade endmember in twoendmember models, MESMA could discriminate ITC species with 48.6% and 50.0%
with crown-scale and pixel-majority classifications, respectively. Increasing
classification accuracy required the addition of endmembers to accommodate
additional within-species spectral variability likely caused by spectral mixtures.
Pixel-majority overall accuracy reached 90.2% and 91.6% when including the full
spectral library as potential two- and three-endmember models, respectively.
However, the processing time needed to evaluate this large number of models was
determined to be too restrictive for operational use. The ability of MESMA to detect
260
unique spectral mixtures for each species is an encouraging result, and several
avenues for future research were identified.
Chapter 6 summary
Spectral and structural properties of tree crowns are important for visual
discrimination of species in aerial photography. It was expected that these properties
would be useful for computer-based species classification with remotely-sensed data.
Chapter 6 focuses on using hyperspectral and lidar data to classify the species of
ITCs in tropical rain forest. Crown-scale spectral data from the HYDICE sensor
combined with structural data from the FLI-MAP lidar sensor were used in
classification analyses. There were significant differences in the majority of lidarderived metrics among the study tree species, indicating that species have unique
structural properties. Crown leaf cover, especially in deciduous leaf-off trees, was
the primary factor controlling variation in lidar metrics. Following methods in
Chapters 3 and 4, the LDA and DT classifiers were used for discriminating tree
species, but at three levels of class aggregation. The best classifier was LDA at all
three levels of aggregation. Overall accuracy with spectral data was 88.9% when
classifying 6 target species, 81.5% for target species with an other-species class, and
90.3% when discriminating a species of conservation interest, Dipteryx, from other
species. In contrast, overall accuracy with lidar data alone was 45.9%, 36.7% and
68.1% for these same levels of classification, respectively. The addition of lidarderived structure information to the classifier did not improve classification
accuracies. However, lidar did help increase the accuracy of a particular species
261
with distinct leaf-off phenology at the cost of additional confusion among leaf-on
species.
7.2. Conclusions and recommendations
The main objective this research was to assess two types of emerging remote
sensing technology, hyperspectral and lidar sensors, for the discrimination of
tropical rain forest tree species. The hyperspectral data contained information on the
biochemical and structural properties of crowns, while the lidar data contained
structural information. I hypothesized that these two datasets combined would allow
greater species classification accuracy than either dataset alone.
I did not find that the structural information from lidar to be useful for species
classification. Considerable effort was required to prepare the lidar dataset for
analysis of individual tree crown structure (Chapter 2), and improvement in species
classification accuracy with lidar was negligible or absent (Chapter 6). However, a
byproduct of the lidar pre-processing, the DTM, had impressive detail and was quite
accurate, even under dense rain forest canopy (Chapter 2). For comparison, the
3800+ survey points took weeks to measure in the field and still did not provide the
same level of terrain detail as the FLI-MAP sensor, which was flown in just 2 days.
Lidar thus proved itself as a powerful new technology for generating DTMs in
tropical landscapes. Although large-footprint sensors can cover larger areas in the
same amount of time, my analyses showed that small-footprint sensors offer greater
DTM detail and accuracy.
There are no commercially-available large-footprint
sensors at this time, and small-footprint systems are still prohibitively expensive for
262
campaigns covering a whole region, such as a large watershed. However, costs are
expected to decline as the technology matures. As a reference, bids for a 2005 smallfootprint, multi-return lidar acquisition at LSBS and vicinity covering an area
approximately 6,000 ha were priced at US$95,000 or more. Roughly half of this
cost is for mobilizing the equipment and team from North America, and so costs
should drop as the technology spreads to other regions of the world.
The DCM has many uses in ecological applications that were not fully explored
in this research. I found that plot-scale estimates of forest height were more accurate
than ITC height estimates (Chapter 2), and since stand height has an allometric
relationship to aboveground biomass, it is expected that fine-scale, lidar-derived
DCMs can be used to predict carbon stocks and forest structure parameters (Lim et
al., 2003; Popescu et al., 2003). Such a strong relationship to plantation and oldgrowth forest structure has been found for a large-footprint system at LSBS (Drake,
Dubayah, Clark et al., 2002). However, as with the DTM, a small-footprint sensor
provides a more detailed perspective on forest structure that may be more
appropriate for certain kinds of ecological research, such as tracking tree-fall gap
dynamics, identifying wildlife habitat, or scaling point-based data (e.g., tower eddyflux carbon measurements) to a broader scale.
Hyperspectral data were found to be critical for remote classification of TRF tree
species.
Species were relatively easy to distinguish based on their leaf-scale
reflectance properties measured in the laboratory; however, accuracy decreased
when using the airborne hyperspectral data (Chapter 3). The importance of spectral
regions also varied with scale. Therefore, studies that classify laboratory reflectance
263
and infer crown-level species separability (e.g., Cochrane, 2000) should be
interpreted with caution. A more realistic test of species separability is to use data
from airborne or spaceborne imaging spectrometers, with all of the associated
variability introduced by poor radiometric calibration, atmospheric contamination
and illumination geometry.
There are two general findings from my research that are encouraging for
operational classification of TRF tree species. First, the elaborate methods and labor
involved in implementing the DT classifier with hyperspectral metrics or the
MESMA classifier with optimal endmembers did not translate into improved species
classification accuracy relative to using reflectance bands with LDA, a more
traditional classifier found in many statistical packages. Another important finding
was that crown-scale spectra provided more accurate ITC species classification, with
both LDA and DT classifiers, than the pixel-majority approach, and there was no
advantage in isolating sunlit crown spectra from the entire crown spectra. These
results suggest that image spatial resolution does not need to be extremely fine for
species discrimination, as long as pixels are not so coarse that they mix crown
radiance with background radiance (e.g., canopy gaps, other trees). This research
investigated hyperspectral imagery with 1.6-m pixels, and there may be an optimum
spatial resolution greater than 1.6-m for species discrimination.
However, one
advantage of this high resolution imagery was that it allowed the accurate
delineation of crowns, thereby minimizing spectral mixing with the background. As
mentioned, coarser spatial resolution imagery will not allow such detailed
delineation of crown boundaries.
264
Chapters 3 and 4 both identified the SWIR region as important for species
discrimination. Most analyses of tropical leaf spectra have covered only the VIS and
NIR regions (Cochrane, 2000; Fung et al., 1998; Poorter et al., 1995), likely because
hyperspectral sensors that include SWIR are generally more expensive. Results in
Chapter 3 indicate that a multispectral sensor with 6 SWIR bands (ASTER) tended
to increase accuracy over sensors with two broad bands (Landsat ETM+) or no
bands (IKONOS) in SWIR. In addition, Chapter 3 and 4 both showed that fullspectrum information—either with optimally-selected bands or in absorption-feature
metrics—provided optimal species discrimination. One advantage of hyperspectral
data is they are over-sampled, and information that does not optimize separation
among species can be discarded.
Leaf phenology was found to be an important consideration in mapping TRF tree
species. In the hyperspectral image, deciduous Dipteryx and Lecythis trees in near
leaf-off conditions had distinct volume-scattering and spectral mixing properties that
influenced the selection of reflectance bands (Ch. 3), spectral metrics (Ch. 4), and
endmembers (Ch. 5). It is unclear from my research whether classification
accuracies would have changed if the imagery had been flown in the wet season,
when Dipteryx and Lecythis have fully-flushed crowns. Also, phenological events
such as senescence before leaf drop or flowering produce changes in reflectance
spectra that may be amplified in leaf-on conditions due to volumetric scattering. If
these changes in spectral reflectance occur synchronously in all individuals of the
target population and do not overlap in time with other species, then image
acquisition may be timed to maximize classification accuracy for a particular target
265
species. For example, Dipteryx has pink flowers and is known to flower with a peak
between May and August when trees have leaves (Frankie et al., 1974; Newstrom et
al., 1994; O’Brien, 2001). However, TRF tree phenology is complex, little
understood, and difficult to generalize; at LSBS, Ceiba (CEPE) flower in the dry
season when trees are leafless, Terminalia (TEOB) populations have variation in
annual flowering intensity, and some overstory trees in the LSBS forest have
asynchronous flowering among individuals of the same species (Frankie et al.,
1974).
In the lidar data, Hymenolobium was leaf-off and had drastically different
structural properties relative to leaf-on species (Ch. 6). Many of the leaf-on species
in the lidar data had similar structural attributes due to low laser penetration.
Classification accuracy may have benefited from the addition of lidar if the data had
been acquired simultaneously with the hyperspectral data in the driest season, when
phenological differences were more pronounced and the laser could penetrate deeper
into the crowns of other leaf-off species.
7.3. Directions for future research
It is clear from Chapters 3-6 that studies seeking to discriminate TRF tree species
with remote sensing technology should focus efforts on hyperspectral data analysis.
Since hyperspectral imagery is typically acquired at spatial scales > 1.6 m, there is a
need for a sensitivity analysis to determine the trade-offs between pixel resolution,
ability to detect crown position and shape, spectral mixing with background
materials, and classification accuracy.
266
A limitation of my hyperspectral analyses is that they infer the phenological state
and associated biophysical properties of tree species based on field data and
observations not directly linked to the study ITCs. I recommend that future studies
acquire ITC-level estimates of crown structure (e.g., leaf and branch area) and other
biological information, such as liana cover, epiphyte cover, and flower density
through either field observation or interpretation of very high spatial resolution
imagery. Such data acquired at the time of image acquisition will permit a deeper
understanding of the conspecific and interspecific spectral variation among ITCs that
is expressed in canopy-level reflectance spectra or spectral metrics.
Lidar analyses suggest that TRF tree crowns are not easily distinguished based
on crown structure. However, besides measuring crowns when phenological
variation was relatively low, my analyses were also constrained by first-return lidar
data. The FLI-MAP sensor was state-of-the-art in 1997, but since then there has
been a steady improvement in lidar technology that has permitted higher post density
and multiple-return recording. For example, an ALTM 3100 lidar sensor (Optech
Inc., Toronto, Canada) scheduled to fly over LSBS in July, 2005 has the capability
to record 4 returns, including a last return. As discussed in Chapter 2, I expect that
ground retrieval will be more accurate with this last-return data; and consequently,
the DCM should also have improved vertical accuracy. In my analyses, I did not
have access to the original xyz point dataset from the FLI-MAP lidar acquisition.
Instead, the DSM data were interpolated from the xyz points, and some detail of
crown structure was lost in the interpolation process. The ALTM 3100 dataset
includes xyz data as part of the delivered products. With multiple returns and point
267
data, it is likely that the ALTM 3100 data will contain a richer representation of ITC
internal structure. This improved detail may translate into greater species
separability, especially if metrics were calculated directly from the point data, rather
than through an interpolated surface (for example, see Brandtberg et al., 2003;
Holmgren & Persson, 2003).
I found that videography, even with its poor image quality, provides a useful
aerial perspective for assessing ITC phenological state in the lidar data. The lidar
campaign at LSBS will also acquire simultaneous, 18-cm orthorectified color
imagery (ALTM 4k x 4k Digital Camera, Optech Inc., Toronto, Canada). ITC
properties that are difficult to measure from the ground, such as leaf phenology,
liana coverage, and epiphyte load, should be visually-interpretable in the aerial view
afforded by 18-cm imagery, providing an unprecedented opportunity to investigate
how these factors affect variability in lidar metrics. Since the digital data will
include 3 multispectral bands (blue, green, red), there are also opportunities to
investigate how crown structure, as measured by the lidar, affects spectral response
and texture in imagery from optical sensors.
Although hyperspectral sensors have immense potential for tree species
discrimination, one major limitation encountered in this research was a lack of
representative trees for each species; I only analyzed 7 out of 400+ tree species at
LSBS. Given natural variability within populations and individuals of TRF tree
species, I do not expect that hyperspectral and lidar technology are capable of
discriminating all tree species in the forest canopy. Auxiliary variables may be
needed to constrain classification rules, such as with the inclusion of soil moisture
268
variables from a lidar-derived DTM (Chapter 6). However, I have demonstrated that
with hyperspectral imagery, certain species of interest may be discernable from the
canopy-matrix of non-target species, especially if the target species have sharp
phenological contrast. The capability of hyperspectral imagery to map target species
thus permits remote and systematic monitoring of long-term changes in key-stone,
endemic, rare or commercial tree species caused by factors, such as selective logging
and climatic change, which are otherwise undetectable by coarse-resolution
multispectral sensors.
Higher classification accuracy may be found by grouping ITCs into functional
types rather than individual species. Tree functional types may be based on common
growth form, metabolism, water balance or disturbance properties (Box, 1996;
Nobel & Gitay, 1996), and not all functional types have a physical manifestation
(e.g. structure, chemistry) which can be detected by remote sensing. Deciduousness
is an important tropical tree functional type that responds to climate, such as changes
in regional precipitation patterns (Condit et al., 1996; Condit et al., 2000).
Knowledge of the proportion of deciduous crowns in a canopy can be used to
calibrate remote sensing estimates of forest parameters, such as aboveground
biomass, productivity, or chlorophyll content (Condit et al., 2000). Leaf phenology
of deciduous and evergreen species was a factor driving species separability in this
research, and I expect that high spatial resolution hyperspectral and lidar data would
be useful for mapping crowns into deciduous and evergreen classes. One potential
application is to use a detailed map of deciduous and evergreen ITCs to calculate the
269
proportion of canopy occupied by deciduous trees. This information could then be
used as reference data for studying how canopy-level deciduousness affects the
spectral response of coarser spatial resolution imagery, such as from the Moderate
Resolution Imaging Spectroradiometer (MODIS) with 250-m to 1000-m resolution
pixels.
Hyperspectral and lidar technology also have great potential for the remote
sensing of aboveground biomass (AGB), a variable vital for assessing carbon stocks
and flux at broad scales and for input into biogeochemical models (Hall et al., 1995).
Tropical forests contain a large proportion of terrestrial carbon, and consequently
have the greatest potential to increase atmospheric carbon dioxide from deforestation
(Dixon et al, 1994), yet remote sensing AGB estimates from optical and synthetic
aperture radar sensors tend to saturate in older secondary and old-growth TRF
forests (Huete et al., 2002; Imhoff, 1995; Luckman et al., 1998; Steininger, 1996;
Steininger, 2000). Drake, Dubayah, Clark et al. (2002) showed that lidar metrics
applied to large-footprint waveforms have immense potential for estimating AGB
over a wide range of tropical forest conditions without saturation. Small-footprint
lidar has also been used to assess AGB, volume and tree height at stand scales for
commercial forestry applications (Holmgren et al., 2003; Lim et al., 2003; Nilsson,
1996), and the plot-scale analysis of tree height in Chapter 2 could be extended to
estimate biomass of plantation, secondary or old-growth forests.
In tropical forests, a large proportion of plot-level AGB is explained by the
largest trees (Brown & Lugo, 1992; Clark & Clark, 2000; Nascimento and Laurance,
270
2002). One advantage of small-footprint lidar data is that large ITCs can be detected
in the canopy, and then AGB may be estimated for each tree from crown-level lidar
metrics. These ITC estimates of AGB can then be aggregated to calculate plot- or
stand-scale AGB. For example, Popescu et al. (2003) found that ITC diameter
information from small-footprint lidar improved stand-scale AGB estimates for
mixed conifer forests. With the ITC approach to AGB estimation, high spatial
multispectral or hyperspectral imagery may be useful to calibrate AGB regression
models to account for species differences in LAI, such as through the incorporation
of the GV or NPV fractions from spectral mixture analysis. Furthermore, species
composition explains a broad gradient of variation in AGB across Amazonian
tropical forests due to taxonomic differences in wood specific gravity (Baker et al.,
2004), yet the variable is generally not used when calculating AGB from allometric
equations. Perhaps hyperspectral imagery can be most beneficial for biomass
estimation if it is used to classify ITCs by wood specific gravity functional types.
Such a classification could be developed from a reclassification of species-level (or
genera-level) ITC maps, such as those found in this research. However, if wood
specific gravity is related to deciduousness (Borchert, 1994; Choat et al., 2005), then
it may be best mapped indirectly from phenology functional classes.
This research also provides a foundation for applications which seek to estimate
species richness across broad spatial scales. Several studies have shown that the
spatial and temporal properties of NDVI may useful for predicting plant species
richness (Bawa et al., 2002; Fairbanks & McGwire, 2004; Gould, 2000; Oindo &
271
Skidmore, 2002). NDVI was used as a predictor variable because it is linked to
ecosystem net primary productivity and biomass as well as because red-well and
NIR bands needed to compute NDVI are available in multispectral satellite sensors.
Other spectral metrics may be more useful; for example, a metric that incorporates
SWIR information should respond to species richness by detecting canopy
variability in leaf phenology. The lidar DCM may also provide information on
canopy structure, such as size and distribution of canopy gaps, which could be
related to species richness (Denslow, 1995).
This research encompasses an initial exploration of yet immature hyperspectral
and lidar technology for future ecological applications in tropical forests.
Considerable additional research is needed with different species, temporal variation,
and sensors to confirm and extend my results. Despite the global importance of
tropical rain forests, there is a major lack of hyperspectral and lidar data from
tropical landscapes, mainly due to cost constraints and large distance between
tropical study sites and major research institutions. It is my hope that this series of
ITC classification analyses based at La Selva will stimulate further lidar and
hyperspectral remote sensing research in tropical forests of the world.
272
REFERENCES
Achard, F., Eva, H.D., Stibig, H.J., Mayaux, P., Gallego, J., Richards, T., &
Malingreau, J.P. (2002). Determination of deforestation rates of the world's
humid tropical forests, Science, 297(5583), 999-1002.
Anderson, M.J. (2004). DISTLM v.5: a FORTRAN computer program to calculate a
distance-based multivariate analysis for a linear model. Department of Statistics,
University of Auckland, New Zealand.
Anderson, M.J. (2001). A new method for non-parametric multivariate analysis of
variance. Austral Ecology, 26(1), 32-46.
Asner, G.P. (1998). Biophysical and biochemical sources of variability in canopy
reflectance. Remote Sensing of Environment, 64(3), 234-253.
Asner, G.P. & Heidebrecht, K.B. (2002). Spectral unmixing of vegetation, soil and
dry carbon cover in arid regions: comparing multispectral and hyperspectral
observations. International Journal of Remote Sensing, 23(19), 3939-3958.
Asner, G.P., Keller, M., Pereira, R., & Zweede, J.C. (2002). Remote sensing of
selective logging in Amazonia - Assessing limitations based on detailed field
observations, Landsat ETM+, and textural analysis. Remote Sensing of
Environment, 80(3), 483-496.
Baltsavias, E.P. (1999a). Airborne laser scanning: basic relations and formulas.
ISPRS Journal of Photogrammetry and Remote Sensing, 54, 199-214.
273
Baltsavias, E.P. (1999b). Airborne laser scanning: existing systems and firms and
other resources. ISPRS Journal of Photogrammetry and Remote Sensing, 54,
164-198.
Baker, T. R., Phillips, O. L., Malhi, Y., Almeida, S., Arroyo, L., Di Fiore, A.,Erwin,
T., Killeen, T. J., Laurance, S. G., Laurance, W. F., Lewis, S. L., Lloyd, J.,
Monteagudo, A., Neill, D. A., Patino, S., Pitman, N. C. A., Silva, J. N. M.,
Martinez, R. V. (2004). Variation in wood density determines spatial patterns in
Amazonian forest biomass. Global Change Biology, 10(5), 545-562.
Bartier, P. M., & Keller, C. P. (1996). Multivariate interpolation to incorporate
thematic surface data using inverse distance weighting (IDW). Computers and
Geosciences, 22(7), 795-799.
Basedow, R.W., Carmer, D.C. & Anderson, M.E. (1995). HYDICE system:
implementation and performance. SPIE Proceedings. Vol. 2480. Feb. 17-24,
Orlando, FL, USA.
Bates, P.D., & De Ro, A.P.J. (2000). A simple raster-based model for flood
inundation simulation. Journal of Hydrology, 236(1-2), 54-77.
Bateson, C. A., Asner, G. P., & Wessman, C. A. (2000). Endmember bundles: A new
approach to incorporating endmember variability into spectral mixture analysis.
IEEE Transactions on Geoscience and Remote Sensing, 38(2), 1083–1094.
Bawa, K., Rose, J., Ganeshaiah, K.N., Barve, N., Kiran, M.C., & Umashaanker, R.
(2002). Assessing biodiversity from space: an example from the Western Ghats,
India. Conservation Ecology, 6(2), 7.
274
Birnbaum, P. (2001). Canopy surface topography in a French Guiana forest and the
folded forest theory. Plant Ecology, 153(1-2), 293-300.
Blackburn, G.A. (1998). Quantifying chlorophylls and caroteniods at leaf and
canopy scales: An evaluation of some hyperspectral approaches. Remote Sensing
of Environment, 66(3), 273-285.
Blackburn, G. A. & Milton, E. J. (1995). Seasonal variations in the spectral
reflectance of deciduous tree canopies. International Journal of Remote Sensing,
16(4), 709-720.
Blair, J.B., & Hofton, M.A. (1999). Modeling laser altimeter return waveforms over
complex vegetation using high-resolution elevation data. Geophysical Research
Letters, 26(16), 2509-2512.
Bohlman, S. & Lashlee, D. (in press). High Spatial and Spectral Resolution Remote
Sensing of Panama Canal Zone Forests: An Applied Example Mapping Tropical
Tree Species. The Rio Chagres: A Multidisciplinary Profile of a Tropical
Watershed. Harmon, R. S. ed. Kluwer Press, Dordrecht.
Borchert, R. (1994). Soil and stem water storage determine phenology and
distribution of tropical dry forest trees. Ecology, 75(5), 1437-1449.
Box, E.O. (1996). Plant functional types and climate at the global scale. Journal of
Vegetation Science, 7, 309-320.
Brandtberg, T. & Walter, F. (1998). Automated delineation of individual tree crowns
in high spatial resolution aerial images by multiple-scale analysis. Machine
Vision and Applications, 11, 64–73.
275
Brandtberg, T., Warner, T.A., Landenberger, R.E., & McGraw, J.B. (2003).
Detection and analysis of individual leaf-off tree crowns in small footprint, high
sampling density lidar data from the eastern deciduous forest in North America.
Remote Sensing of Environment, 85(3), 290-303.
Breiman, L., Freedman, J., Olshen, R., Stone, C. (1984). Classification and
Regression Trees. Wadsworth, Belmont, California.
Broge, N.H. & Leblanc, E. (2000). Comparing prediction power and stability of
broadband and hyperspectral vegetation indices for estimation of green leaf area
index and canopy chlorophyll density. Remote Sensing of Environment, 76(2),
156-172.
Brown, I.F., Martinelli, L.A., Thomas, W.W., Moreira, M.Z., Ferreira, C.A.C., &
Victoria, R.A. (1995). Uncertainty in the biomass of Amazonian forests - an
example from Rondonia, Brazil, Forest Ecology and Management, 75(1-3), 175189.
Brown, S. & Lugo, A.E. (1992). Aboveground biomass estimates for tropical moist
forests of the Brazilian Amazon. Interciencia, 17(1), 8-18.
Carleer, A. & Wolff, E. (2004). Exploitation of very high resolution satellite data for
tree species identification. Photogrammetric Engineering and Remote Sensing,
70(1), 135-140.
Castro-Esau, K.L., Sánchez-Azofeifa, G.A., Caelli, T. (2004). Discrimination of
lianas and trees with leaf-level hyperspectral data. Remote Sensing of
Environment, 90, 353–372.
276
Choat, B., Ball, M.C., Luly, J.G., & Holtum, J.A.M. (2005). Hydraulic architecture
of deciduous and evergreen dry rainforest tree species from north-eastern
Australia. Trees-Structure and Function, 19(3), 305-311.
Clark, D.A., Piper, S.C., Keeling, C.D. & Clark, D.B. (2003). Tropical rain forest
tree growth and atmospheric carbon dynamics linked to interannual temperature
variation during 1984-2000. Proceedings of the National Academy of Sciences of
the United States of America. 100(10), 5852-5857.
Clark, D.B., Read, J.M., Clark, M.L., Murillo Cruz, A., Fallas Dotti, M., & Clark,
D.A. (2004). Application of 1-m and 4-m resolution satellite data to studies of
tree demography, stand structure and land-use classification in tropical rain
forest landscapes. Ecological Applications, 14(1), 61–74.
Clark, D.B., Castro, C.S., Alvarado, L.D.A., & Read, J.M. (2004). Quantifying
mortality of tropical rain forest trees using high-spatial-resolution satellite data.
Ecology Letters, 7(1), 52-59.
Clark, D.B., Palmer, M. & Clark, D.A. (1999). Edaphic factors and the landscapescale distributions of tropical rain forest trees. Ecology, 80, 2662-2675.
Clark, D.B., Clark, D.A., & Read, J.M. (1998). Edaphic variation and the mesoscale
distribution of tree species in a neotropical rain forest. Journal of Ecology, 86(1),
101-112.
Clark, D.B. & D.A. Clark. (2000). Landscape-scale variation in forest structure and
biomass in a tropical rain forest. Forest Ecology and Management, 137(1-3),
185-198.
277
Clark, D.B., & Clark, D.A. (1996). Abundance, growth and mortality of very large
trees in neotropical lowland rain forest. Forest Ecology and Management, 80(13), 235-244.
Clark, D.B., Clark, D.A., Rich, P.M., Weiss, S. & Oberbauer, S.F. (1996).
Landscape scale evaluation of understory light and canopy structure: Methods
and application in a neotropical lowland rain forest. Canadian Journal of Forest
Research, 26(5), 747-757.
Clark, M.L., Roberts, D.A., & Clark, D.B. (2005). Hyperspectral discrimination of
tropical rain forest tree species at leaf to crown scales. Remote Sensing of
Environment, 96(3-4), 375-398.
Clark, M.L., Clark, D.B., & Roberts, D.A. (2004). Small-footprint lidar estimation of
sub-canopy elevation and tree height in a tropical rain forest landscape. Remote
Sensing of Environment, 91(1), 68-89.
Clark, R.N., Swayze, G.A., Livo, K.E., Kokaly, R.F., Sutley, S.J., Dalton, J.B.,
McDougal, R.P., & Gent, C.A. (2003). Imaging spectroscopy: Earth and
planetary remote sensing with the USGS Tetracorder and expert systems.
Journal of Geophysical Research-Planets, 108, E12.
Clément, J. & Guellec, J. (1974). Utilization des photographies aériennes au 1/5.000
en coleur pour la détection de l’Okoumé dans la forêt dense du Gabon. Bois et
Forêts de Tropiques, 153, 3-22.
Cobby, D.M., Mason, D.C., & Davenport, I.J. (2001). Image processing of airborne
scanning laser altimetry data for improved river flood modelling. ISPRS Journal
of Photogrammetry and Remote Sensing, 56, 121-138.
278
Cochrane, M.A. (2000). Using vegetation reflectance variability for species level
classification of hyperspectral data. International Journal of Remote Sensing,
21(10), 2075-2087.
Cochrane, M.A., Alencar, A., Schulze, M.D., Souza, C.M., Nepstad, D.C., Lefebvre,
P., & Davidson, E.A. (1999). Positive feedbacks in the fire dynamic of closed
canopy tropical forests. Science, 284(5421), 1832-1835.
Condit, R., Pitman, N., Leigh, E. G., Chave, J., Terborgh, J., Foster, R. B., Nunez,
P., Aguilar, S., Valencia, R., Villa, G., Muller-Landau, H. C., Losos, E., Hubbell,
S. P. (2002). Beta-diversity in tropical forest trees. Science, 295(5555), 666-669.
Condit, R., Watts, K., Bohlman, S.A., Perez, R., Foster, R.B., Hubbell, S.P. (2000).
Quantifying the deciduousness of tropical forest canopies under varying
climates. Journal of Vegetation Science, 11(5), 649-658.
Condit, R., Hubbel, S.P., & Foster, R.B. (1996). Assessing the response of plant
functional types to climatic change in tropical forests. Journal of Vegetation
Science, 7, 405-416.
Congalton, R. (1991). A review of assessment the accuracy of classifications of
remotely sensed data. Remote Sensing of Environment, 37, 35-46.
Congalton, R.G. & Mead, R.A. (1983). A quantitative method to test for consistency
and correctness in photointerpretation. Photogrammetric Engineering and
Remote Sensing, 49(1), 69-74.
Curran, P.J., Dungan, J.L., Peterson, D.L. (2001). Estimating the foliar biochemical
concentration of leaves with reflectance spectrometry: Testing the Kokaly and
Clark methodologies. Remote Sensing of Environment, 76(3), 349-359.
279
Culvenor, D.S. (2002). TIDA: an algorithm for the delineation of tree crowns in high
spatial resolution remotely sensed imagery. Computers & Geosciences, 28(1), 3344.
Curran, P.J. (1989). Remote sensing of foliar chemistry. Remote Sensing of
Environment, 30(3), 271-278.
Demetriades-Shah, T.H., Steven, M.D. & Clark, J.A. (1990). High-resolution
derivative spectra in remote-sensing. Remote Sensing of Environment, 33(1), 5564.
Dennison, P.E., Roberts, D.A., Thorgusen, S.R., Regelbrugge, J.C., Weise, D., &
Lee, C. (2003). Modeling seasonal changes in live fuel moisture and equivalent
water thickness using a cumulative water balance index. Remote Sensing of
Environment. 88(4), 442-452.
Dennison, P.E. & Roberts, D.A. (2003a). Endmember selection for multiple
endmember spectral mixture analysis using endmember average RMSE. Remote
Sensing of Environment, 87(2-3), 123-135.
Dennison, P.E. & Roberts, D.A. (2003b). The effects of vegetation phenology on
endmember selection and species mapping in southern California chaparral.
Remote Sensing of Environment, 87(2-3), 295-309.
Denslow, J.S. (1995). Disturbance and diversity in tropical rain-forests - the density
effect. Ecological Applications, 5(4), 962-968.
Deutsch, C.V., & Journel, A.G. (1992). GSLIB: Geostatistical Software Library and
User's Guide. New York: Oxford University Press.
280
Dixon, R.K., Brown, S., Houghton, R.A., Solomon, A.M., Trexler, M.C., &
Wisniewski, J. (1994). Carbon pools and flux of global forest eco-systems.
Science, 263 (5144), 185–190.
Drake, J.B., Dubayah, R.O., Clark, D.B., Knox, R.G., Blair, J.B., Hofton, M.A.,
Chazdon, R.L., Weishampel, J.F., & Prince, S.D. (2002). Estimation of tropical
forest structural characteristics using large-footprint lidar. Remote Sensing of
Environment, 79, 305-319.
Drake, J.B., Dubayah, R.O., Knox, R.G., Clark, D.B., & Blair, J.B. (2002).
Sensitivity of large-footprint lidar to canopy structure and biomass in a
neotropical rainforest. Remote Sensing of Environment, 83, 378-392.
Duda, R.O. & Hart, P.E. (1973). Pattern Classification and Scene Analysis. John
Wiley & Sons, Inc, New York.
Dymond, C.C., & Johnson, E.A. (2002). Mapping vegetation spatial patterns from
modeled water, temperature and solar radiation gradients. ISPRS Journal of
Photogrammetry and Remote Sensing, 57(1-2), 69-85.
Elvidge, C.D. & Chen, Z.K. (1995). Comparison of broadband and narrowband red
and near-Infrared vegetation indexes. Remote Sensing of Environment, 54(1), 3848.
Elvidge, C. D. (1990). Visible and near-Infrared reflectance characteristics of dry
plant materials. International Journal of Remote Sensing, 11(10), 1775-1795.
Fairbanks, D.H.K. & McGwire, K.C. (2004). Patterns of floristic richness in
vegetation communities of California: regional scale analysis with multitemporal NDVI. Global Ecology and Biogeography, 13(3), 221-235.
281
Fearnside, P.M. (1999). Biodiversity as an environmental service in Brazil's
Amazonian forests: risks, value and conservation. Environmental Conservation,
26(4), 305-321.
Ferreira, L.G., Yoshioka, H., Huete, A., & Sano, E.E. (2003). Seasonal landscape
and spectral vegetation index dynamics in the Brazilian Cerrado: An analysis
within the Large-Scale Biosphere-Atmosphere Experiment in Amazonia (LBA).
Remote Sensing of Environment, 87(4), 534-550.
Foody, G.M. (2003). Remote sensing of tropical forest environments: towards the
monitoring of environmental resources for sustainable development.
International Journal of Remote Sensing, 24(20), 4035-4046.
Foody, G.M., Boyd, D.S., & Cutler, M.E.J. (2003). Predictive relations of tropical
forest biomass from Landsat TM data and their transferability between regions.
Remote Sensing of Environment, 85(4), 463-474.
Foody, G.M., Cutler, M.E., McMorrow, J., Pelz, D., Tangki, H., Boyd, D.S. &
Douglas, I. (2001). Mapping the biomass of Bornean tropical rain forest from
remotely sensed data. Global Ecology and Biogeography, 10(4), 379-387.
Fournier, R.A., Edwards, G. & Eldridge, N.R. (1995). A catalogue of potential
spatial discriminators for high spatial resolution digital images of individual
crowns. Canadian Journal of Remote Sensing, 21, 285-298.
Frankie, G.W., Baker, H.G., Opler, P.A. (1974). Comparative phenological studies
of trees in tropical wet and dry forests in the lowlands of Costa Rica. Journal of
Ecology, 62(3), 881-919.
282
Franklin, J. (1998). Predicting the distribution of shrub species in southern
California from climate and terrain-derived variables. Journal of Vegetation
Science, 9(5), 733-748.
Franklin, J., Prince, S.D., Strahler, A. H., Hanan, N. P. & Simonett, D.S. (1991).
Reflectance and transmission properties of West African savanna trees from
ground radiometer measurements. International Journal of Remote Sensing,
12(6), 1369-1385.
Franklin, S.E., Maudie, A.J., Lavigne, M.B. (2001). Using spatial co-occurrence
texture to increase forest structure and species composition classification
accuracy. Photogrammetric Engineering and Remote Sensing, 67(7), 849-855.
Franklin, S.E., Hall, R.J., Moskal, L.M., Maudie, A.J., Lavigne, M.B. (2000).
Incorporating texture into classification of forest species composition from
airborne multispectral images. International Journal of Remote Sensing, 21(1),
61-79.
Friedl, M.A. & Brodley, C.E. (1997). Decision tree classification of land cover from
remotely sensed data. Remote Sensing of Environment, 61(3), 399-409.
Fuentes, D.A., Gamon, J.A., Qiu, H.L., Sims, D.A., & Roberts, D.A. (2001).
Mapping Canadian boreal forest vegetation using pigment and water absorption
features derived from the AVIRIS sensor. Journal of Geophysical ResearchAtmospheres, 106(D24), 33565-33577.
Fung, T., Ma, F.Y. & Siu, W.L. (1998). Hyperspectral data analysis for subtropical
tree species recognition. Symposium Proceedings, IGARSS '98, Sensing and
Managing the Environment, Seattle, WA, USA, July 6-10, 1998 (pp. 1298-1300,
283
Vol. 3). IEEE International Geoscience and Remote Sensing, New York, NY,
USA.
Gamon, J.A., Serrano, L., & Surfus, J.S. (1997). The photochemical reflectance
index: an optical indicator of photosynthetic radiation use efficiency across
species, functional types, and nutrient levels. Oecologia, 112(4), 492-501.
Gao, B.C. (1996). NDWI - A normalized difference water index for remote sensing
of vegetation liquid water from space. Remote Sensing of Environment, 58(3),
257-266.
Gao, B.C. & Goetz, A.F.H. (1990). Column atmospheric water-vapor and vegetation
liquid water retrievals from airborne imaging spectrometer data. Journal of
Geophysical Research-Atmospheres, 95(D), 3549-3564.
Gates, D. M., Keegan, H. J., Schleter, J. C., Weidner, V. R. (1965). Spectral
properties of plants. Applied Optics, 4(1), 11-20.
Gausman, H.W. (1985). Plant Leaf Optical Properties in Visible and Near-infrared
Light. Texas Tech Press, Lubbock, Texas.
Gaveau, D.L.A. & Hill, R.A. (2003). Quantifying canopy height underestimation by
laser pulse penetration in small-footprint airborne laser scanning data. Canadian
Journal of Remote Sensing, 29(5), 650-657.
Gessler, P. E., Chadwick, O. A., Chamran, F., Althouse, L., & Holmes, K. (2000).
Modeling soil-landscape and ecosystem properties using terrain attributes. Soil
Science Society of America Journal, 64(6), 2046-2056.
284
Gong, P., Pu, R., & Heald, R.C. (2002). Analysis of in situ hyperspectral data for
nutrient estimation of giant sequoia. International Journal of Remote Sensing,
23(9), 1827-1850.
Gong, P., Pu, R.L., & Yu, B. (1997). Conifer species recognition: An exploratory
analysis of in situ hyperspectral data. Remote Sensing of Environment, 62(2),
189-200.
Goovaerts, P. (1997). Geostatistics for Natural Resources Evaluation. New York:
Oxford University Press.
Gotway, C.A., & Cressie, N.A.C. (1990). A spatial analysis of variance applied to
soil-water infiltration. Water Resources Research, 26(11), 2695-2703.
Gougeon, F.A. (1999). Automatic individual tree crown delineation using a valleyfollowing algorithm and a rule-based system. Proceedings of the international
forum: automated interpretation of high spatial resolution digital imagery for
forestry, Victoria, British Columbia, February 10–12, 1998 (pp. 11-24). Victoria,
BC: Canadian Forest Service, Pacific Forestry Centre.
Gougeon, F.A. (1995). Comparison of possible multispectral classification schemes
for tree crowns individually delineated on high spatial resolution MEIS images.
Canadian Journal of Remote Sensing, 21, 1-9.
Gougeon, F.A. & Leckie, D.G. (2003). Forest information extraction from high
spatial resolution images using an individual tree crown approach. Information
Report BC-X-396, Pacific Forestry Centre, Canadian Forest Service, Victoria,
British Columbia, Canada.
285
Gould, W. (2000). Remote sensing of vegetation, plant species richness, and regional
biodiversity hotspots. Ecological Applications, 10(6), 1861-1870.
Grant, L. (1987). Diffuse and specular characteristics of leaf reflectance. Remote
Sensing of Environment, 22, 309–322.
Haertel, V. & Landgrebe, D.A. (1999). On the classification of classes with nearly
equal spectral response in remote sensing hyperspectral image data. IEEE
Transactions on Geoscience and Remote Sensing, 37(5), 2374-2386.
Halligan, K.Q. (2002). Multiple Endmember Spectral Mixture Analysis of Vegetation
in the Northeast Corner of Yellowstone National Park. Master Thesis. University
of California, Santa Barbara, California. 64 p.
Hall, F.G., Shimabukuro, Y.E., & Huemmrich, K.F. (1995). Remote-sensing of
forest biophysical structure using mixture decomposition and geometric
reflectance models. Ecological Applications, 5(4), 993-1013.
Harding, D.J., Lefsky, M.A., Parker, G.G., & Blair, J.B. (2001). Laser altimeter
canopy height profiles: Methods and validation for closed-canopy, broadleaf
forests. Remote Sensing of Environment, 76, 283-297.
Hartshorn, G.S. & Hammel, B.E. (1994). Vegetation types and floristic patterns. In
McDade L.A., Bawa, K.S., Hespenheide, H.A., & Hartshorn, G.S. (Eds.), La
Selva: ecology and natural history of a neotropical rain forest (pp. 73-89).
Chicago: University of Chicago Press.
Herwitz, S.R., Slye, R.E. & Turton, S.M. (1998). Redefining the ecological niche of
a tropical rain forest canopy tree species using airborne imagery: long-term
crown dynamics of Toona ciliata. Journal of Tropical Ecology, 14, 683-703.
286
Herwitz, S.R., & Slye, R.E. (1995). 3-dimensional modeling of canopy tree
interception of wind-driven rainfall. Journal of Hydrology, 168(1-4), 205-226.
Hill, D.A. & Leckie, D.G. (1999). Proceedings of the international forum:
automated interpretation of high spatial resolution digital imagery for forestry,
Victoria, British Columbia, February 10–12, 1998 (402 pages). Victoria, BC:
Canadian Forest Service, Pacific Forestry Centre.
Hinsley, S.A., Hill, R.A., Gaveau, D.L.A., & Bellamy, P.E. (2002). Quantifying
woodland structure and habitat quality for birds using airborne laser scanning.
Functional Ecology, 16(6), 851-857.
Hodgson, M.E., Jensen, J.R., Schmidt, L., Schill, S., & Davis, B. (2003). An
evaluation of LIDAR- and IFSAR-derived digital elevation models in leaf-on
conditions with USGS Level 1 and Level 2 DEMs. Remote Sensing of
Environment, 84, 295-308.
Hofton, M.A., Rocchio, L.E., Blair, J.B., & Dubayah, R. (2002). Validation of
vegetation canopy lidar sub-canopy topography measurements for a dense
tropical forest. Journal of Geodynamics, 34, 491-502.
Holdridge, L.R. (1971). Forest environments in tropical life zones: a pilot study.
New York: Pergamon Press.
Holmgren, J. & Persson, A. (2004). Identifying species of individual trees using
airborne laser scanner. Remote Sensing of Environment, 90(4), 415-423.
Holmgren, J., Nilsson, M. & Olsson, H. (2003). Estimation of tree height and stem
volume on plots-using airborne laser scanning. Forest Science, 49(3), 419-428.
287
Horler, D.N.H., Dockray M., Barber J. (1983). The red edge of plant leaf reflectance.
International Journal of Remote Sensing, 4(2), 273-288.
Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., & Ferreira, L.G. (2002).
Overview of the radiometric and biophysical performance of the MODIS
vegetation indices. Remote Sensing of Environment, 83(1-2), 195-213.
Huete, A.R. (1988). A Soil-Adjusted Vegetation Index (SAVI). Remote Sensing of
Environment, 25(3), 295-309.
Huete, A.R., Jackson, R.D., Post, D.F. (1985). Spectral response of a plant canopy
with different soil backgrounds. Remote Sensing of Environment, 17(1), 37-53.
Huising, E.J., & Gomes Pereira, L.M. (1998). Errors and accuracy estimates of laser
data acquired by various laser scanning systems for topographic applications.
ISPRS Journal of Photogrammetry and Remote Sensing, 53, 245-261.
Hurtt, G., Xiao, X., Keller, M., Palace, M., Asner, G.P., Braswell, R., Brondizio,
E.S., Cardoso, M., Carvalho, C.J.R., Fearon, M.G. (2003). IKONOS imagery for
the Large Scale Biosphere-Atmosphere Experiment in Amazonia (LBA). Remote
Sensing of Environment, 88(1-2), 111-127.
Imhoff, M.L. (1995). Radar backscatter and biomass saturation - ramifications for
global biomass inventory. IEEE Transactions on Geoscience and Remote
Sensing, 33(2), 511-518.
Isaaks, E.H., & Srivastava, R.M. (1989). An Introduction to Applied Geostatistics.
New York: Oxford University Press.
288
Jackson, Q. & Landgrebe, D.A. (2001). An adaptive classifier design for highdimensional data analysis with a limited training data set. IEEE Transactions on
Geoscience and Remote Sensing, 39(12), 2664-2679.
Jordan, C.F. (1969). Derivation of Leaf-Area Index from quality of light on the
forest floor. Ecology, 50(4), 663-666.
Kaufman, Y.J. & Tanré, D. (1992). Atmospherically Resistant Vegetation Index
(ARVI) for EOS-MODIS. IEEE Transactions on Geoscience and Remote
Sensing, 30(2), 261-270.
Keshava, N. & Mustard, J.F. (2002). Spectral unmixing. IEEE Signal Processing
Magazine, 19(1), 44-57.
Key, T., Warner, T.A., McGraw, J.B., & Fajvan, M.A. (2001). A comparison of
multispectral and multitemporal information in high spatial resolution imagery
for classification of individual tree species in a temperate hardwood forest.
Remote Sensing of Environment. 75(1), 100-112.
Kimes, D.S. (1983). Dynamics of directional reflectance factor distributions for
vegetation canopies. Applied Optics, 22(9), 1364–1372.
Kokaly, R.F., Despain, D.G., Clark, R.N. & Livo, K.E. (2003). Mapping vegetation
in Yellowstone National Park using spectral feature analysis of AVIRIS data.
Remote Sensing of Environment, 84(3), 437-456
Kokaly, R.F. & Clark, R.N. (1999). Spectroscopic determination of leaf
biochemistry using band- depth analysis of absorption features and stepwise
multiple linear regression. Remote Sensing of Environment, 67(3), 267-287.
289
Kraus, K., & Pfeifer, N. (1998). Determination of terrain models in wooded areas
with airborne laser scanner data. ISPRS Journal of Photogrammetry and Remote
Sensing, 53, 193-203.
Kruse, F.A., Lefkoff, A.B., Boardman, J.W., Heidebrecht, K.B., Shapiro, A.T.,
Barloon, P.J., & Goetz, A.F.H. (1993). The Spectral Image-Processing System
(SIPS) - Interactive Visualization and Analysis of Imaging Spectrometer Data.
Remote Sensing of Environment, 44(2-3), 145-163.
Krzanowski, W.J. (2001). Data-based interval estimation of classification error.
Journal of Applied Statistics, 28(5), 585-595.
Lamar, W.R., McGraw, J.B., Warner, T.A. (2005). Multitemporal censusing of a
population of eastern hemlock (Tsuga canadensis L.) from remotely sensed
imagery using an automated segmentation and reconciliation procedure. Remote
Sensing of Environment, 94(1), 133-143.
Laurance, W.F. (1999). Reflections on the tropical deforestation crisis. Biological
Conservation, 91, 109-117.
Laurance, W.F., Oliveira, A.A., Laurance, S.G., Condit, R., Nascimento, H.E.M.,
Sanchez-Thorin, A.C., Lovejoy, T.E., Andrade, A., D’Angelo, S., Ribeiro, J.E.,
& Dick, C.W. (2004). Pervasive alteration of tree communities in undisturbed
Amazonian forests. Nature, 438, 171-175.
Leckie, D.G., & Gougeon, F.A. (1999). An assessment of both visual and automated
tree counting and species identification with high spatial resolution multispectral
imagery. Proceedings of the international forum: automated interpretation of
high spatial resolution digital imagery for forestry, Victoria, British Columbia,
290
February 10–12, 1998 (pp. 141–154). Victoria, BC: Canadian Forest Service,
Pacific Forestry Centre.
Leckie, D.G., Gougeon, F.A., Tinis, S., Nelson, T., Burnett, C.N., & Paradine, D.
(2005). Automated tree recognition in old growth conifer stands with high
resolution digital imagery. Remote Sensing of Environment, 94(3), 311-326.
Leckie, D., Gougeon, F., Hill, D., Quinn, R., Armstrong, L., Shreenan, R. (2003).
Combined high-density lidar and multispectral imagery for individual tree crown
analysis. Canadian Journal of Remote Sensing, 29(5), 633-649.
Leckie, D.G., Gougeon, F.A., Walsworth, N. & Paradine, D. (2003). Stand
delineation and composition estimation using semi-automated individual tree
crown analysis. Remote Sensing of Environment, 85(3), 355-369.
Lefsky, M.A., Cohen, W.B., Parker, G.G., & Harding, D.J. (2002). Lidar remote
sensing for ecosystem studies. Bioscience, 52(1), 19-30.
Lim, K., Treitz, P., Baldwin, K., Morrison, I., & Green, J. (2003). Lidar remote
sensing of biophysical properties of tolerant northern hardwood forests.
Canadian Journal of Remote Sensing, 29(5), 658-678.
Lohmann, P., & Koch, A. (1999). Quality assessment of laser-scanner-data.
Proceedings of the International Society of Photogrammetry and Remote
Sensing. (ISPRS) workshop "Sensors and Mapping from Space 1999”,
Hannover, Germany, September 27-30, 1999.
Lloyd, C.D., & Atkinson, P.M. (2002a). Non-stationary approaches for mapping
terrain and assessing prediction uncertainty, Transactions in GIS, 6(1), 17-30.
291
Lloyd, C.D., & Atkinson, P.M. (2002b). Deriving DSMs from LiDAR data with
kriging. International Journal of Remote Sensing, 23(12), 2519-2524.
Luckman, A., Baker, J., Honzák, M. & Lucas, R. (1998). Tropical forest biomass
density estimation using JERS-1 SAR: Seasonal variation, confidence limits,
and application to image mosaics. Remote Sensing of Environment, 63, 126-139.
Magnussen, S., & Boudewyn, P. (1998). Derivations of stand heights from airborne
laser scanner data with canopy-based quantile estimators. Canadian Journal of
Forest Research, 28, 1016-1031.
Martin, M.E., Newman, S.D., Aber, J.D., & Congalton, R.G. (1998). Determining
forest species composition using high spectral resolution remote sensing data.
Remote Sensing of Environment, 65(3), 249-254.
Martin, M.E. & Aber, J.D. (1997). High spectral resolution remote sensing of forest
canopy lignin, nitrogen, and ecosystem processes. Ecological Applications, 7(2),
431-443.
McArdle, B.H. & Anderson, M.J. (2001). Fitting multivariate models to community
data: A comment on distance-based redundancy analysis, Ecology, 82(1), 290297.
McGraw, J.B., Warner, T.A., Key, T.L. & Lamar, W.R. (1998). High spatial
resolution remote sensing of forest trees. Trends in Ecology and Evolution,
13(8), 300-301.
McGwire, K., Minor, T., & Fenstermaker, L. (2000). Hyperspectral mixture
modeling for quantifying sparse vegetation cover in arid environments. Remote
Sensing of Environment, 72(3), 360-374.
292
Menalled, F.D., Kelty, M.J., & Ewel, J.J. (1998). Canopy development in tropical
tree plantations: a comparison of species mixtures and monocultures. Forest
Ecology and Management, 104, 249-263.
Merton, R.N. (1998). Monitoring community hysteresis using spectral shift analysis
and the Red-edge Vegetation Stress Index. Summaries of the Seventh JPL
Airborne Earth Science Workshop, Pasadena, CA: Jet Propulsion Laboratory.
Meyer, P., Staenz, K. & Itten, K.I. (1996). Semi-automated procedures for tree
species identification in high spatial resolution data from digitized colour
infrared-aerial photography. ISPRS Journal of Photogrammetry and Remote
Sensing, 51(1), 5-16.
Montgomery, R. A. & Chazdon, R. L. (2001). Forest structure, canopy architecture,
and light transmittance in tropical wet forests. Ecology, 82(10), 2707-2718.
Moore, G.E. (1998). Cramming more components onto integrated circuits (Reprinted
from Electronics, p. 114-117, April 19, 1965), Proceedings of the IEEE, 86(1),
82-85.
Myers, B.J. and Benson, M.L. (1981). Rainforest species on large-scale colour
photos. Photogrammetric Engineering and Remote Sensing, 47, 505-513.
Næsset, E. (2002). Predicting forest stand characteristics with airborne scanning
laser using a practical two-stage procedure and field data. Remote Sensing of
Environment, 80(1), 88-99.
Næsset, E. (1997). Determination of mean tree height of forest stands using airborne
laser scanner data. ISPRS Journal of Photogrammetry and Remote Sensing, 52,
49-56.
293
Næsset, E. & Økland, T. (2002). Estimating tree height and tree crown properties
using airborne scanning laser in a boreal nature reserve. Remote Sensing of
Environment, 79(1), 105-115.
Nagler, P.L., Glenn, E.P., Thompson, T.L. & Huete, A. (2004). Leaf area index and
normalized difference vegetation index as predictors of canopy characteristics
and light interception by riparian species on the Lower Colorado River.
Agricultural and Forest Meteorology, 125(1-2), 1-17.
Nascimento, H.E.M. & Laurance, W. F. (2002). Total aboveground biomass in
central Amazonian rainforests: a landscape-scale study. Forest Ecology and
Management, 168(1-3), 311-321.
Nepstad, D.C., P.R.Moutinho, C. Uhl, I.C.Vieira, & J.M. Cardosa da Silva. (1996).
The ecological importance of forest remnants in an eastern Amazonian frontier
landscape. in Forest Patches in Tropical Landscapes. J.Schelhas and R. Greenberg
(eds.), Island Press, Washington, D.C.
Newstrom, L.E., Frankie, G.W., Baker, H.G., & Colwell, R.K. (1994). Diversity of
long-term flowering patterns. In McDade L.A., Bawa, K.S., Hespenheide, H.A.,
& Hartshorn, G.S. (Eds.), La Selva: ecology and natural history of a neotropical
rain forest (pp. 142-160). Chicago: University of Chicago Press.
Nilsson, M. (1996). Estimation of tree heights and stand volume using an airborne
lidar system. Remote Sensing of Environment, 56(1), 1-7.
Nobel, I.R. & Gitay, H. (1996). A functional classification for predicting the
dynamics of landscapes. Journal of Vegetation Science, 7, 329-336.
294
O'Brien, J.J. (2001). The effects of climate on the growth and physiology of tropical
rainforest trees. Ph.D. Dissertation. Florida International University, Miami,
Florida.
Oindo, B.O. & Skidmore, A.K. (2002). Interannual variability of NDVI and species
richness in Kenya. International Journal of Remote Sensing, 23(2), 285-298.
Okin, G.S., Roberts, D.A., Murray, B., & Okin, W.J. (2001). Practical limits on
hyperspectral vegetation discrimination in arid and semiarid environments.
Remote Sensing of Environment, 77(2), 212-225.
Osborne, B.G. & Fearn, T. (1986). Near Infrared Spectroscopy in Food Analysis.
Longman Scientific and Technical, Essex, England.
Painter, T. H., Roberts, D. A., Green, R. O., & Dozier, J. (1998). The effect of grain
size on spectral mixture analysis of snow-covered area from AVIRIS data.
Remote Sensing of Environment, 65, 320– 332.
Pal, M. & Mather, P.M. (2003). An assessment of the effectiveness of decision tree
methods for land cover classification. Remote Sensing of Environment, 86(4),
554-565.
Palmer, K.F. & Williams, D. (1974). Optical properties of water in the near infrared.
Journal of the Optical Society of America, 64, 1107-1110.
Pannatier, Y. (1996). VARIOWIN: Software for spatial data analysis in 2D. New
York: Springer-Verlag.
Peñuelas, , J., Llusia, J., Pinol, J., & Filella, I. (1997a). Photochemical reflectance
index and leaf photosynthetic radiation-use-efficiency assessment in
295
Mediterranean trees. International Journal of Remote Sensing, 18(13), 28632868.
Peñuelas, J., Pinol, J., Ogaya, R., & Filella, I. (1997b). Estimation of plant water
concentration by the reflectance water index WI (R900/R970). International
Journal of Remote Sensing, 18(13), 2869-2875.
Persson, A., Holmgren, J., & Söderman, U. (2002). Detecting and measuring
individual trees using an airborne laser scanner. Photogrammetric Engineering
and Remote Sensing, 68(9), 925-932.
Petzold, B., Reiss, P., & Stössel, W. (1999). Laser scanning—surveying and
mapping agencies are using a new technique for the derivation of digital terrain
models. ISPRS Journal of Photogrammetry and Remote Sensing, 54, 95-104.
Phillips, O.L., Vargas, P.N. Monteagudo, A.L., Cruz, A.P., Zans, M.E.C., Sanchez,
W.G., Yli-Halla, M., & Rose, S. (2003). Habitat association among Amazonian
tree species: a landscape-scale approach, Journal of Ecology, 91, 757–775.
Phillips, O.L. (1997). The changing ecology of tropical forests, Biodiversity and
Conservation, 6, 291-311.
Poorter, L., Oberbauer, S.F. & Clark, D.B. (1995). Leaf optical-properties along a
vertical gradient in a tropical rain-forest canopy in Costa-Rica. American Journal
of Botany, 82(10), 1257-1263.
Popescu, S.C., Wynne, R.H., Nelson, R.F. (2003). Measuring individual tree crown
diameter with lidar and assessing its influence on estimating forest volume and
biomass. Canadian Journal of Remote Sensing, 29(5), 564-577.
296
Powell, G, Dr. (2001). Senior Conservation Scientist, World Wildlife Fund,
Monteverde, Costa Rica. Personal communication.
Powell, R.L., Matzke, N., de Souza, C. Jr., Clark, M., Numata, I., Hess, L.L., &
Roberts, D.A. (2004). Sources of error in accuracy assessment of thematic landcover maps in the Brazilian Amazon. Remote Sensing of Environment, 90, 221–
234.
Preston, R.A., Culvenor, D.S. & N.C. Coops, N.C. (1999). Modeling of tree species
and structural attributes from high resolution multi-spectral imagery using decision
tree analysis for east coast Eucalypt forests of Australia. Proceedings of the
international forum: automated interpretation of high spatial resolution digital
imagery for forestry, Victoria, British Columbia, February 10–12, 1998 (pp. 225242). Victoria, BC: Canadian Forest Service, Pacific Forestry Centre.
Price, J.C. (1994). How unique are spectral signatures? Remote Sensing of
Environment, 48, 181-186.
Price, J.C. (1992). Variability of high-resolution crop reflectance spectra.
International Journal of Remote Sensing, 13(14), 2593-2610.
Pu, R., Ge, S., Kelly, N.M., & Gong, P. (2003). Spectral absorption features as
indicators of water status in coast live oak (Quercus agrifolia) leaves.
International Journal of Remote Sensing, 24(9), 1799-1810.
Pu, R.L., Gong, P., Biging, G.S. & Larrieu, M.R. (2003). Extraction of red edge
optical parameters from Hyperion data for estimation of forest leaf area index.
IEEE Transactions on Geoscience and Remote Sensing, 41(4), 916-921.
297
R Development Core Team (2004). R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria.
ISBN 3-900051-07-0, http://www.R-project.org.
Raber, G.T., Jensen, J.R., Schill, S.R., & Schuckman, L. (2002). Creation of digital
terrain models using an adaptive lidar vegetation point removal process.
Photogrammetric Engineering and Remote Sensing, 68(12), 1307-1315.
Rabus, B., Eineder, M., Roth, A., & Bamler, R. (2003). The shuttle radar topography
mission—a new class of digital elevation models acquired by spaceborne radar.
ISPRS Journal of Photogrammetry and Remote Sensing, 57(4), 241-262.
Ranson, K.J., Sun, G., Weishampel, J.F. and Knox, R.G. (1997). Forest biomass
from combined ecosystem and radar backscatter modeling. Remote Sensing of
Environment, 59, 118-133.
Raven, P.H., Evert, R.F., & Eichhorn, S.E. (1992). Biology of Plants, 5th Edition.
Worth Publishers, Inc., New York, New York.
Raven, P.H. (1988). Our diminishing tropical forests. in Biodiversity. E.O. Wilson,
ed. pp.119-122. National Academy Press, Washington, D.C.
Raupach, M.R. (1994). Simplified expressions for vegetation roughness length and
zero-plane displacement as functions of canopy height and area index.
Boundary-Layer Meteorology, 71(1-2), 211-216.
Read, J.M., Clark, D.B., Venticinque, E.M. & Moreira, M.P. (2003). Application of
merged 1-m and 4-m resolution satellite data to research and management in
tropical forests. Journal of Applied Ecology, 20, 592–600.
298
Reutebuch, S.E., McGaughey, R.J., Anderson, H.-E., & Carson, W.W. (2003).
Accuracy of a high-resolution lidar terrain model under a conifer forest canopy.
Canadian Journal of Remote Sensing, 29(5), 527-535.
Ritchie, J.C., Evans, D.L., Jacobs, D., Everitt, J.H., & Weltz, M.A. (1993).
Measuring canopy structure with an airborne laser altimeter. Transactions of the
ASAE, 36(4), 1235-1238.
Roberts, D.A., Ustin, S.L., Ogunjemiyo, S., Greenberg, J., Dobrowski, S.Z., Chen,
J., & Hinckley, T.M. (2004). Spectral and structural measures of Northwest
forest vegetation at leaf to landscape scales. Ecosystems, 7(5), 545-562.
Roberts, D.A., Dennison, P.E., Hetzel, Y., Gardner, M., Ustin, S.L., Lee, C. (2003).
Evaluation of the potential of Hyperion for fire danger assessment by
comparison to the airborne visible infrared imaging spectrometer. IEEE
Transactions on Geoscience and Remote Sensing, 41(6), 1297-1310.
Roberts, D.A., Numata, I., Holmes, K., Batista, G., Krug, T., Moteiro, A., Powell,
B., and Chadwick, O. A. (2002). Large area mapping of land-cover change in
Rondônia using multitemporal spectral mixture analysis and decision tree
classifiers. Journal of Geophysical Research-Atmospheres, 107(D20), 8073.
Roberts, D.A., Brown, K., Green, R., Ustin, S., & Hinckley, T. (1998). Investigating
the relationship between liquid water and leaf area in clonal Populus. Summaries
of the Seventh JPL Airborne Earth Science Workshop, Pasadena, CA: Jet
Propulsion Laboratory.
Roberts, D.A., Batista, G. Pereira, J., Waller, E, and Nelson, B. (1998). Change
identification using multitemporal spectral mixture analysis: Applications in
299
eastern Amazonia, in Remote Sensing Change Detection: Environmental
Monitoring Applications and Methods, eds. C. Elvidge and R. Lunetta, pp. 137–
161, Ann Arbor Press, Chelsea, Mich.
Roberts, D.A., Gardner, M., Church, R., Ustin, S., Scheer, G. & Green, R.O. (1998).
Mapping chaparral in the Santa Monica Mountains using multiple endmember
spectral mixture models. Remote Sensing of Environment, 65(3), 267-279.
Roberts, D. A., Nelson, B. W., Adams, J. B., & Palmer, F. (1998). Spectral changes
with leaf aging in Amazon caatinga. Trees-Structure and Function, 12(6), 315325.
Roberts, D.A., Green, R.O, & Adams, J.B. (1997). Temporal and spatial patterns in
vegetation and atmospheric properties from AVIRIS. Remote Sensing of
Environment, 62(3), 223-240.
Roberts, D.A., Smith, M.O., & Adams, J.B. (1993). Green vegetation,
nonphotosynthetic vegetation, and soils in AVIRIS Data. Remote Sensing of
Environment, 44(2-3), 255-269.
Rocchio, L. E. (2000). Lidar remote sensing of sub-canopy topography. Master's
Thesis, University of Maryland, College Park, Maryland, USA.
Rouse, J.W., Haas, R.H., Schell, J.A., & Deering, D.W. (1973). Monitoring
vegetation systems in the great plains with ERTS. Third ERTS Symposium,
NASA SP-351, volume 1, 309-317.
Sader, S.A., Waide, R.B., Lawrence, W.T., & Joyce, A.T. (1989). Tropical forest
biomass and successional age class relationships to a vegetation index derived
from Landsat TM data. Remote Sensing of Environment, 28, 143-156.
300
Sandmeier, S., Muller, C., Hosgood, B., Andreoli, G. (1998). Sensitivity analysis
and quality assessment of laboratory BRDF data. Remote Sensing of
Environment, 64(2), 176-191.
Sanford, R.L., Jr., Pabby, P., Luvall, J.C., & Phillips, E. (1994). Climate,
geomorphology, and aquatic systems. In McDade L.A., Bawa, K.S.,
Hespenheide, H.A., & Hartshorn, G.S. (Eds.), La Selva: ecology and natural
history of a neotropical rain forest (pp. 19-33). Chicago: University of Chicago
Press.
Serrano, L., Ustin, S.L., Roberts, D.A., Gamon, J.A., & Peñuelas, J. (2000). Deriving
water content of chaparral vegetation from AVIRIS data. Remote Sensing of
Environment, 74(3), 570-581.
Sims, D.A. & Gamon, J.A. (2002). Relationships between leaf pigment content and
spectral reflectance across a wide range of species, leaf structures and
developmental stages. Remote Sensing of Environment, 81(2-3), 337-354.
Skole, D. & Tucker, C. (1993). Tropical deforestation and habitat fragmentation in
the Amazon - satellite data from 1978 to 1988. Science, 260(5116), 1905-1910.
Smith, M.L., Ollinger, S.V., Martin, M.E., Aber, J.D., Hallett, R.A., & Goodale,
C.L. (2002). Direct estimation of aboveground forest productivity through
hyperspectral remote sensing of canopy nitrogen. Ecological Applications, 12(5),
1286-1302.
Souza, C. & Barreto, P. (2000). An alternative approach for detecting and
monitoring selectively logged forests in the Amazon. International Journal of
Remote Sensing, 21(1), 173-179.
301
Souza, C. & Roberts, D.A. (2005). Mapping forest degradation in the Amazon
region with Ikonos images. International Journal of Remote Sensing, 26(3), 425429.
Spanner, M., Johnson, L., Miller, J., McCreight, R., Freemantle, J., Runyon, J. &
Gong, P. (1994). Remote-sensing of seasonal Leaf-Area Index across the Oregon
transect. Ecological Applications, 4(2), 258-271.
Steininger, M.K., Tucker, C.J., Townshend, J.R.G., Killeen, T.J., Desch, A., Bell, V.
& Ersts, P. (2001). Tropical deforestation in the Bolivian Amazon.
Environmental Conservation, 28(2),127-134.
Steininger, M.K. (2000). Satellite estimation of tropical secondary forest
aboveground biomass: data from Brazil and Bolivia. International Journal of
Remote Sensing, 21(6), 1139-1157.
Steininger, M. K. (1996). Tropical secondary forest regrowth in the Amazon: Age,
area and change estimation with Thematic Mapper data. International Journal of
Remote Sensing, 17(1), 9-27.
Tabachnick, B.G. & Fidell, L.S. (1989). Using multivariate statistics, Second
Edition. Harper Collins Publishers, Inc. New York, USA.
Thenkabail, P.S., Enclona, E.A., Ashton, M.S., Legg, C., De Dieu, M.J. (2004).
Hyperion, IKONOS, ALI, and ETM+ sensors in the study of African rainforests.
Remote Sensing of Environment, 90(1), 23-43.
Thomas, C.D., Cameron, A., Green, R.E., Bakkenes, M., Beaumont, L.J.,
Collingham, Y. C., Erasmus, B.F.N., de Siqueira, M.F., Grainger, A., Hannah,
L., Hughes, L., Huntley, B., van Jaarsveld, A.S., Midgley, G.F., Miles, L.,
302
Ortega-Huerta, M.A., Peterson, A.T., Phillips, O.L., & Williams, S.E. (2004).
Extinction risk from climate change. Nature, 427(6970), 145-148.
Trichon, V. (2001). Crown typology and identification of rain forest trees on largescale aerial photographs. Plant Ecology, 153, 301-312.
Tucker, C.J. (1979). Red and photographic infrared linear combinations for
monitoring vegetation. Remote Sensing of Environment, 8(2), 127-150.
Tuomisto, H., Poulsen, A. D., Ruokolainen, K., Moran, R. C., Quintana, C., Celi, J.,
Canas, G. (2003). Linking floristic patterns with soil heterogeneity and satellite
imagery in Ecuadorian Amazonia. Ecological Applications, 13(2), 352-371.
Tuomisto, H., Ruokolainen, K., &Yli-Halla, M. (2003). Dispersal, environment, and
floristic variation of western Amazonian forests. Science, 299(5604), 241-244.
Uhl, C., Kauffman, J. (1990). Deforestation, fire susceptibility, and potential tree
responses to fire in the eastern Amazon. Ecology, 71(2), 437-449.
Ustin, S.L. & Xiao, Q.F. (2001). Mapping successional boreal forests in interior
central Alaska. International Journal of Remote Sensing, 22(9), 1779-1797.
Ustin, S. L., Roberts, D. A., Pinzón, J., Jacquemoud, S., Gardner, M., Scheer, G.,
Castañeda, C. M., & Palacios-Orueta, A. (1998). Estimating canopy water
content of chaparral shrubs using optical methods. Remote Sensing of
Environment, 65(3), 280-291.
van Aardt, J.A.N. & Wynne, R.H. (2001). Spectral separability among six Southern
tree species. Photogrammetric Engineering and Remote Sensing, 67(12), 13671375.
303
Wang, L., Sousa, W.P., & Gong, P. (2004). Comparison of IKONOS and QuickBird
images for mapping mangrove species on the Caribbean coast of Panama.
Remote Sensing of Environment, 91, 432-440.
Warner, T.A., Lee, J.Y. & McGraw, J.B. (1999). Delineation and identification of
individual trees in the Eastern deciduous forest. Proceedings of the international
forum: automated interpretation of high spatial resolution digital imagery for
forestry, Victoria, British Columbia, February 10–12, 1998 (pp. 81-92). Victoria,
BC: Canadian Forest Service, Pacific Forestry Centre.
Weishampel, J.F., Blair, J.B., Knox, R.G., Dubayah, R. & Clark, D.B. (2000).
Volumetric lidar return patterns from an old-growth tropical rainforest canopy.
International Journal of Remote Sensing. 21(2), 409-415.
Whitmore, T.C. (1990). An introduction to tropical rain forests. Clarendon Press,
Oxford.
Williams, D.L. (1991). A comparison of spectral reflectance properties at the needle,
branch, and canopy level for selected conifer species. Remote Sensing of
Environment, 35(2-3), 79-93.
Woolley, J.T. (1971). Reflectance and transmittance of light by leaves. Plant
Physiology, 47, 656-662.
Xiao, Q., Ustin, S.L., & McPherson, E.G. (2004). Using AVIRIS data and multiplemasking techniques to map urban forest tree species. International Journal of
Remote Sensing, 25(24), 5637-5654.
304
Yin, Z. Y., & Wang, X. H. (1999). A cross-scale comparison of drainage basin
characteristics derived from digital elevation models. Earth Surface Processes
and Landforms. 24(6), 557-562.
Zar, J.H. (1996). Biostatistical analysis, 3rd Edition. Prentice Hall, New Jersey.
Zarco-Tejada, P.J., Miller, J.R., Mohammed, G.H., & Noland, T.L. (2000).
Chlorophyll fluorescence effects on vegetation apparent reflectance: I. Leaf-level
measurements and model simulation. Remote Sensing of Environment, 74(3),
582-595.
Zhang, C., Franklin, S.E., & Wulder, M.A. (2004). Geostatistical and texture
analysis of airborne-acquired images used in forest classification. International
Journal of Remote Sensing, 25(4), 859-865.
305
APPENDIX I: List of Acronyms
AGB
Above-Ground Biomass
ARVI
Atmospherically Resistant Vegetation Index
ASTER
Advanced Spaceborne Thermal Emission & Reflection
Radiometer
BE
Blue Edge
COBI
COunt-Based Index
DCM
Digital Canopy Model
DSM
Digital Surface Model
DT
Decision Tree
DTM
Digital Terrain Model
EAR
Endmember Average Root mean square error
ETM
Enhanced Landsat Thematic Mapper
EVI
Enhanced Vegetation Index
EWT
Equivalent Water Thickness
FOV
Field of View
GP
Green Peak
GV
Green Vegetation
HYDICE
HYperspectral Digital Imagery Collection Experiment
IDW
Inverse Distance Weighted
ITC
Individual Tree Crown
306
LAI
Leaf Area Index
LDA
Linear Discriminant Analysis
LIDAR
LIght Detection And Ranging
LSBS
La Selva Biological Station
MAE
Mean Absolute Error
MESMA
Multiple Endmember Spectral Mixture Analysis
ML
Maximum Likelihood
MODIS
MODerate Resolution Imaging Spectroradiometer
NASA
National Aeronautics and Space Administration
NDVI
Normalized Difference Vegetation Index
NDWI
Normalized Difference Water Index
NE
Near-infrared water Edge
NIR
Near InfraRed
NPMANOVA Non-Parametric Multivariate Analysis of Variance
NPV
Non-photosynthetic Vegetation
PRI
Photochemical Reflectance Index
OK
Ordinary Kriging
RE
Red Edge
RMSE
Root Mean Square Error
RVSI
Red-edge Vegetation Stress Index
RW
Red Well
SAM
Spectral Angle Mapper
307
SAR
Synthetic Aperature Radar
SAVI
Soil-Adjusted Vegetation Index
SE
Shortware-infrared Edge
SMA
Spectral Mixture Analysis
SqrtMAE
Square-Root Transformation of the Mean Absolute Error
SR
Simple Ratio
SWIR
ShortWave InfraRed
TRF
Tropical Rain Forest
VIS
Visible
WBI
Water Band Index
308
APPENDIX II: Summary of spectral metrics (Chapter 4)
309
Appendix 2.1. Global mean and standard deviation (in parentheses) of narrow-band, ratio-based indices for
the seven study species at different scales of measurement. Significant differences among species means
were tested with an ANOVA test. Pair-wise multiple comparisons between species means tested with the
Tukey’s Honestly Significant Difference method (p≤0.05). Significance levels for ANOVA F statistic: ns =
not significant, * = p≤0.05, ** = p≤0.01, *** = p≤0.001, **** = p≤0.0001. Bark n=66; Leaf n=152; Pixel
n=2100; Crown n=214.
Mean
Mean
Mean
Mean
F
F
F
F
Metric (S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
Vegetation Indices
SR
2.37
1.6 ns
9.59
8.1 **** 14.43
244.8 **** 12.87
64.3 ****
(0.88) 0
(3.96) 9
(6.02) 19
(5.19) 15
NDVI 0.38
2.3 *
0.78
4.7 ***
0.84
173.4 **** 0.83
36.9 ****
(0.12) 1
(0.09) 5
(0.08) 18
(0.06) 13
SAVI
0.28
6.1 **** 0.57
7.4 **** 0.54
64.1 ****
0.54
39.1 ****
(0.10) 6
(0.08) 6
(0.15) 16
(0.09) 12
PRI
-0.38
2.3 *
-0.78
4.7 ***
-0.84
173.4 **** -0.83
36.9 ****
(0.12) 1
(0.09) 5
(0.08) 18
(0.06) 13
EVI
0.05
4.8 ***
0.10
7.3 **** 0.09
64.5 ****
0.09
38.9 ****
(0.02) 4
(0.02) 7
(0.03) 15
(0.02) 12
309
310
Appendix 2.1. (continued)
Mean
F
Metric (S.D.)
Pairs
ARVI 0.20
1.5 ns
(0.14) 0
RVSI
0.006
9.1 ****
(0.00) 10
Mean
Mean
F
F
(S.D.)
Pairs
(S.D.)
Pairs
0.74
2.9 **
0.77
197.1 ****
(0.10) 0
(0.10)
18
0.001 8.1 ****
0.015
78.8 ****
(0.01) 9
(0.00)
18
Liquid Water Content Indices
WBI
0.94
14.0 **** 1.01
16.0 **** 1.09
155.5 ****
(0.09) 10
(0.02) 9
(0.09)
18
NDWI -0.14
14.8 **** 0.02
14.5 **** 0.03
280.0 ****
(0.12) 10
(0.03) 9
(0.05)
18
Feature abbreviations listed in Table 4.4.
310
Mean
(S.D.)
0.76
(0.09)
0.014
(0.00)
F
Pairs
42.8 ****
12
31.0 ****
14
1.09
(0.07)
0.03
(0.04)
19.1 ****
12
48.2 ****
14
311
Appendix 2.2. Global mean and standard deviation (in parentheses) of derivative-based metrics for
the seven study species at different scales of measurement. Tests are the same as in Appendix 2.1.
Bark
Leaves
Pixels
Crowns
Mean
Mean
Mean
Mean
F
F
F
F
Metric
(S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
Derivative Inflection Wavelength Position (nm)
BE-λ
518.7
2.3 *
521.8
4.9 ***
519.9
16.3 **** 519.9
9.3 ****
(3.3)
1
(1.7)
4
(2.4)
8
(1.2)
5
GP-λ
552.4
5.9 ****
555.8
39.0 ***
554.9
10.3 ****
n/d
n/d
(4.1)
4
(12.3)
14
(11.0)
5
YE-λ
572.8
4.9 ***
572.0
1.1 ns
570.5
3.0 **
570.4
0.6 ns
(4.1)
6
(2.4)
0
(6.5)
2
(5.0)
0
RW-λ
663.7
1.4 ns
669.5
5.5 ****
665.9
44.2 **** 666.0
11.2 ****
(2.1)
0
(6.5)
7
(1.8)
14
(1.0)
7
RE-λ
692.7
3.3 **
712.1
28.7 **** 725.2
58.4 **** 724.7
9.0 ****
(8.9)
1
(7.5)
10
(4.4)
15
(3.2)
6
NE1-λ
1028.2 2.7 *
1030.6 3.1 **
996.0
13.9 **** 992.9
3.2 **
(25.9) 1
(25.9)
2
(17.0)
10
(5.4)
1
NE2-λ
1144.3 2.9 *
1143.1 2.1 ns
1130.2 5.0 ****
1127.6 0.2 ns
(5.8)
1
(3.4)
0
(8.5)
3
(8.4)
0
SE-λ
1498.0 2.8 *
1496.1 7.2 ****
1522.0 36.6 **** 1512.8 12.1 ****
(9.8)
3
(3.8)
6
(21.1)
14
(16.7)
9
311
312
Appendix 2.2. (continued)
Derivative Inflection Magnitude (x 1000) or Percent Reflectance
BE-Mag
0.65
8.1 ****
1.72
5.0 ***
0.88
31.0 **** 0.84
(0.27) 6
(0.88)
3
(0.38)
12
(0.21)
GP-Refl
93.79
3.0 **
42.85
13.3 **** 42.81
n/d
n/d
(35.1
2
(18.6)
10
(8.9)
YE-Mag
0.27
4.7 ***
-1.27
1.8 ns
-0.49
41.2 **** -0.37
(0.20) 5
(1.18)
0
(0.31)
15
(0.18)
RW-Refl
138.2
3.7 **
51.67
2.9 *
25.40
51.5 **** 26.51
(47.7) 1
(27.5)
2
(13.3)
16
(9.4)
RE-Mag
2.92
11.5 **** 7.72
8.5 ****
5.34
70.7 **** 5.03
(1.63) 8
(1.48)
6
(2.49)
15
(1.50)
NE1-Mag
0.93
6.3 ****
0.24
9.8 ****
1.32
86.6 **** 1.27
(0.41) 6
(0.14)
6
(0.64)
17
(0.42)
NE2-Mag
-1.00
2.9 *
-0.62
32.5 **** -2.91
56.1 **** -2.38
(1.30) 2
(0.22)
15
(1.62)
14
(1.01)
SE-Mag
1.09
2.9 *
1.22
9.3 ****
0.91
15.3 **** 0.93
(0.25) 1
(0.20)
7
(0.43)
12
(0.21)
312
18.0 ****
8
4.6 ***
5
26.6 ****
10
19.1 ****
11
40.0 ****
12
17.6 ****
12
15.5 ****
10
5.9 ****
5
Appendix 2.2. (continued)
313
Derivative-based Area (x 100)
BE2.30
8.1 ****
4.31
4.9 ***
2.35
22.4 **** 2.28
12.0 ****
DArea
(0.96) 6
(2.21)
3
(0.99)
12
(0.53)
7
YE1.22
6.0 ****
-1.95
2.9 *
-0.70
116.2 **** -0.63
49.7 ****
DArea
(0.61) 7
(1.25)
1
(0.51)
16
(0.37)
13
RWE21.12
13.3 **** 36.08
6.6 ****
33.52
56.1 **** 32.19
32.8 ****
DArea
(9.06) 11
(6.95)
5
(14.3)
15
(7.99)
12
RWE13.76
4.1 **
38.88
6.9 ****
35.22
53.7 **** 33.77
31.5 ****
DNArea
(9.57) 4
(7.98)
6
(15.4)
15
(8.61)
12
RWE0.22
8.3 ****
0.06
8.0 ****
0.05
41.5 **** 0.05
9.4 ****
2DNArea
(0.14) 7
(0.10)
6
(0.05)
14
(0.03)
7
Feature abbreviations listed in Table 4.4. λ = wavelength, Mag = derivative magnitude, Refl =
percent reflectance, DArea = area under 1st derivative, DNArea = area under normalized 1st
derivative, 2DNArea = area under 2nd derivative
313
314
Appendix 2.3. Global mean and standard deviation (in parentheses) of absorption-based metrics for
the seven study species at different scales of measurement. Tests are the same as in Appendix 2.1.
Bark
Leaves
Pixels
Crowns
Mean
Mean
Mean
Mean
F
F
F
F
Metrics
(S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
(S.D.)
Pairs
Equivalent Water Thickness
EWT
0.07
3.7 **
0.04
14.8 ****
0.31
160.2 **** 0.30
15.8 ****
(0.11)
3
(0.03)
12
(0.17) 16
(0.13) 12
Maximum Depth (% Reflectance)
Blue-D
0.06
1.6 ns
0.32
11.4 ****
0.41
73.0 ****
0.36
35.6 ****
(0.03)
0
(0.14)
10
(0.14) 16
(0.10) 11
Red-D
0.29
2.4 *
0.80
4.7 ***
0.83
162.3 **** 0.82
36.5 ****
(0.13)
1
(0.08)
4
(0.09) 18
(0.07) 13
NIR1-D
0.03
2.9 *
0.02
20.0 ****
0.13
165.9 **** 0.12
14.7 ****
(0.04)
2
(0.01)
11
(0.05) 16
(0.05) 14
NIR2-D
0.05
2.7 *
0.04
47.3 ****
0.17
145.0 **** 0.17
15.1 ****
(0.05)
2
(0.01)
15
(0.07) 16
(0.06) 12
SWIR1-D 0.01
3.3 *
0.06
35.9 ****
0.06
9.1 ****
(0.01)
1
n/d
n/d
(0.02) 15
(0.02) 4
SWIR2-D
0.07
36.0 ****
0.06
45.1 ****
n/d
n/d
n/d
n/d
(0.03) 14
(0.02) 13
SWIR3-D 0.05
20.9 ****
0.03
15.9 ****
0.10
10.5 ****
0.07
22.3 ****
(0.03)
11
(0.02)
011
(0.06) 10
(0.01) 12
314
Appendix 2.3. (continued).
Blue-λ
Red-λ
NIR1-λ
315
NIR2-λ
SWIR1-λ
495.8
(2.9)
672.3
(4.1)
990.0
(9.9)
1177.3
(12.1)
1727.5
(10.3)
1.1 ns
0
3.6 **
1
3.1 **
3
4.0 **
3
4.5 **
4
n/d
2293.7
(18.8)
n/d
22.2 ****
11
SWIR2-λ
SWIR3-λ
Wavelength (nm)
498.1
12.3 ****
492.8
(5.6)
10
(7.0)
678.6
4.1 ***
670.0
(2.4)
5
(0.9)
994.1
2.9 *
958.0
(14.5)
1
(12.1)
1174.5 3.1 **
1168.6
(9.1)
0
(12.9)
1749.3
n/d
n/d
(3.8)
2103.8
n/d
n/d
(27.6)
2307.9 2.4 *
2313.9
(24.0)
1
(23.6)
315
5.1 ****
6
0.6 ns
0
16.0 ****
10
8.3 ****
6
16.9 ****
11
26.3 ****
14
10.5 ****
11
495.8
(2.5)
669.9
(0.0)
957.7
(8.6)
1169.4
(9.6)
1748.4
(1.1)
2097.9
(21.1)
2299.2
(7.4)
2.7 *
1
3.7 **
4
2.8 *
1
4.6 ***
4
1.8 ns
1
9.3 ****
9
1.5 ns
0
Appendix 2.3. (continued).
Blue-W
316
44.1
(2.9)
Red-W
73.6
(11.0)
NIR1-W
49.9
(17.5)
NIR2-W
74.0
(4.2)
SWIR1-W 26.8
(12.0)
SWIR2-W
n/d
SWIR3-W 57.1
(20.5)
Blue-A1
Blue-A2
Red-A1
Red-A2
NIR1-A1
2.5
(1.1)
2.5
(1.1)
22.3
(13.0)
22.5
(12.2)
2.1
(3.1)
1.2 ns
0
3.2 **
3
6.9 ****
6
3.8 **
3
3.0 *
1
45.8
(3.8)
101.6
(9.1)
61.9
(15.8)
73.0
(3.1)
Width (nm)
6.6 ****
7
5.7 ****
5
4.7 ***
3
3.9 **
3
n/d
n/d
n/d
14.8 ****
9
n/d
51.7
(28.1)
1.9 ns
1
1.9 ns
1
2.2 ns
1
2.3 *
1
3.3 **
2
14.8
(7.3)
14.2
(6.9)
81.4
(13.2)
76.9
(12.1)
1.2
(0.6)
n/d
11.5 ****
8
Area
11.1 ****
10
11.4 ****
10
4.5 ***
4
4.3 ***
4
25.8 ****
11
316
46.4
(3.4)
109.2
(5.8)
63.6
(10.0)
74.2
(12.1)
44.1
(9.4)
69.1
(29.8)
50.3
(25.0)
20.7 ****
12
179.3 ****
18
31.6 ****
15
5.7 ****
6
196.4 ****
18
158.4 ****
17
19.5 ****
13
47.2
(1.4)
108.1
(4.7)
65.5
(7.0)
77.5
(9.2)
45.7
(7.4)
78.5
(22.7)
65.1
(12.5)
23.4 ****
9
46.9 ****
12
3.1 **
2
0.4 ns
0
59.4 ****
14
38.4 ****
13
0.3 ns
0
19.0
(6.5)
18.1
(6.1)
91.3
(13.3)
85.5
(12.4)
8.2
(3.5)
90.9 ****
17
75.2 ****
17
199.6 ****
18
206.9 ****
18
246.8 ****
18
17.3
(5.1)
16.7
(4.9)
89.0
(11.4)
83.3
(10.7)
8.1
(3.1)
36.3 ****
11
36.6 ****
11
46.8 ****
13
49.0 ****
13
18.8 ****
14
317
Appendix 2.3. (continued).
NIR1-A2 1.9
3.6 **
(3.3)
3
NIR2-A1 3.8
2.7 *
(3.6)
2
NIR2-A2 3.7
2.8 *
(3.5)
2
SWIR10.4
2.4 *
A1
(0.4)
0
SWIR10.2
3.3 *
A2
(0.7)
1
SWIR2A1
n/d
n/d
SWIR2A2
n/d
n/d
SWIR32.7
5.5 ***
A1
(1.4)
3
SWIR33.1
10.7 ****
A2
(1.5)
9
1.2
(0.7)
2.6
(0.9)
2.5
(0.8)
20.8 ****
11
58.4 ****
15
56.8 ****
15
n/d
n/d
n/d
n/d
n/d
n/d
n/d
1.9
(1.6)
1.6
(1.7)
n/d
20.6 ****
11
17.3 ****
11
317
8.3
(3.9)
13.0
(5.5)
12.9
(5.5)
2.9
(1.3)
2.7
(1.6)
4.9
(3.2)
3.7
(3.6)
4.8
(3.4)
4.1
(5.0)
221.8 ****
16
160.7 ****
16
160.3 ****
16
79.0 ****
19
72.9 ****
19
102.5 ****
14
144.2 ****
16
8.9 ****
9
4.7 ****
4
8.2
(3.4)
12.8
(4.4)
12.7
(4.4)
3.0
(1.0)
2.9
(1.2)
4.7
(2.3)
4.0
(2.6)
4.3
(1.3)
4.1
(1.1)
18.7 ****
14
16.4 ****
13
16.2 ****
13
21.6 ****
7
21.3 ****
7
62.6 ****
13
67.4 ****
13
8.4 ****
5
20.2 ****
11
Appendix 2.3. (continued).
Asymmetry (x 100)
6.6 ****
91.0
20.4 ****
90.9
22.6 ****
7
(0.6)
12
(0.3)
9
Red-As
5.1 ****
84.7
184.1 **** 84.8
48.3 ****
3
(0.8)
18
(0.6)
12
NIR1- As
4.9 ***
93.6
28.1 ****
93.4
2.7 *
3
(1.0)
14
(0.7)
2
NIR2- As
3.7 **
93.9
6.0 ****
93.6
0.5 ns
3
(1.0)
6
(0.8)
0
SWIR197.5
196.6 **** 97.4
59.5 ****
As
n/d
n/d
(0.5)
18
(0.4)
14
SWIR296.8
159.5 **** 96.4
38.9 ****
As
n/d
n/d
n/d
n/d
(1.4)
17
(1.0)
13
SWIR397.5
14.8 ****
97.8
11.6 ****
97.9
19.9 ****
97.2
0.4 ns
As
(0.9)
9
(1.2)
8
(1.1)
13
(0.5)
0
Feature abbreviations listed in Table 4.4. n/d = not detected. D = depth, λ = wavelength, W=width,
A1 = area calculated using width and depth, A2 = area calculated using tabulation, As = Asymmetry
Blue-As
318
91.5
(0.5)
89.5
(1.5)
95.1
(1.7)
93.9
(0.3)
98.5
(0.7)
1.2 ns
0
3.3 **
3
6.8 ****
6
3.6 **
2
3.0 *
1
91.2
(0.7)
85.7
(1.2)
94.0
(1.5)
94.0
(0.2)
318
Appendix 2.4. Global mean and standard deviation (in parentheses) of Spectral
Mixture Analysis (SMA) fractions and model error among the study species (3
endmember model) at pixel and crown scales of measurement. Tests are the
same as in Appendix 2.1.
Pixels
Crowns
Fraction
Mean
No.
Mean
No.
(S.D.)
Pairs
(S.D.)
Pairs
F
F
41.3
76.7
39.3
37.5
GV
(18.4)
****
17
(11.4) ****
12
15.1
189.4
17.2
43.8
NPV
(14.9)
****
18
(11.6) ****
14
43.6
19.4
43.5
8.1
Shade
(20.9)
****
11
(9.2)
****
7
1.2
23.3
1.0
5.2
RMSE
(0.5)
****
13
(0.3)
****
3
GV = % green vegetation, NPV= % non-photosynthetic vegetation, Shade = %
photometric shade, RMSE=root mean square error, units percent reflectance.
319