Spectral Separability among
Six Southern Tree Species
Jan A.N. van Aardt and Randolph H. Wynne
Abstract
Spectroradiometer data (350 to 2500 nm) were acquired in late
summer 1999 over various forest sites in Appomattox Buckingham State Forest, Virginia, to assess the spectral differentiability among six major forestry tree species: loblolly pine
(Pinus taeda), Virginia pine (Pinus virginiana), shortleaf pine
(Pinus echinata), scarlet oak (Quercus coccinea), white oak
(Quercus alba), and yellow poplar (Liriodendron tulipifera).
Data were smoothed and curve shape was determined using
first- and second-difference operators. Stepwise discriminant
analysis was used to decrease the number of independent
variables, after which a canonical discriminant analysis and
a normal discriminant analysis were performed. Cross-validation accuracies varied from 99 percent to 100 percent (hardwood versus pine groups), 62 percent to 84 percent (within
pine group), and 78 percent to 93 percent (within hardwood
group). The second difference of a nine-point weighted average proved most accurate overall, with cross-validation
accuracies of 84 percent (within pine separability), 93 percent
(within hardwood separability), and 100 percent (between
group separability). Landsat simulation data had lower
accuracies, varying from 93 percent to 96 percent (hardwood
versus pine groups), 45 percent to 60 percent (within pine
group), and 54 percent to 70 percent (within hardwood group).
The relatively low accuracies for Landsat simulation data
indicate the need for high spectral resolution data for within
group separability. The variables significant in defining
spectral separability within and between groups were largely
located in the visible (350- to 700-nm) and shortwave infrared
1(700- to 1850-nm) regions of the spectrum, with markedly
less representation in the shortwave infrared I1 (1 700- to 2500nm) region. Some wavelengths related to nitrogen concentration and 0 - H bond regions were evident, but not dominant.
An accurate classification of any given forested area is
important for forest management, aiding in forest inventory
(yield per species or group), pest and environmental stress
management (dyingtdecayingtdrought-stressedtrees), and
assessing habitat ranges or managing human impact on a forest
environment. To be able to effectively manage forests and assess
forest conditions, it is therefore imperative not only to locate
forest species (or forest taxonomic groups such as hardwoods
and softwoods),but also to assess various biophysical parameters subsequent to identifying the speciestgroup location. One
of the most important techniques for regional forest assessment
and inventory is remote sensing. Airborne and spaceborne sensors are now available in a wide variety of spatial, spectral, and
temporal resolutions, with forest classification having evolved
through time as new sensors became available.
The use of hyperspectral sensors often makes discernment
of an area's composition through spectral response discrimination more effective than is possible with the broader band
multispectral sensors (Birk and McCord, 1994). Sensors such
as the Airborne Visible and Infrared Imaging Spectrometer
(AVIRIS) (range:400 nm to 2500 nm; resolution: 10 nm), Hymap
(range: 400 nm to 2500 nm; resolution: 16 nm), and Hydice
(range:400 nm to 2500 nm; resolution: 10.2 nm) have prepared
the way for a new set of applications to be explored by providing data of sufficient spectral resolution to resolve the natural
variability in features such as minerals, vegetation, and atmospheric gases (Birk and McCord, 1994).
The identification of bands correlated with leaf chemical
compounds using airborne data (Curran, 1989;Wessman et al.,
1989; Yoder and Pettigrew-Cosby, 1995; Kokaly and Clark,
1999;Kokaly, 2001) created new possibilities for forest classification, because leaf chemistry characteristics might vary
between forest taxonomic groups or even species. A more comprehensive approach that involves the use of hyperspectral
data and their analysis has been taken of late (Lawrence et al.,
1993; Gong et al., 1997; Martin et al., 1998; Fung et al., 1999)
rather than the traditional "broader wave range and fewer
classes" approach (Nelson et al., 1984; Shen et al., 1985; Franklin, 1994;White et al., 1995)that has been utilized for so long
in natural resources research.
The importance of the visible and near-infrared wavelength regions, as well as derivative spectral analysis, has been
stressed in almost all of the vegetation studies utilizing hyperspectral data. Results obtained by Gong et al. (1997) using a
field spectroradiometer (250- to 1050-nm range; 2.6 nm spectral
resolution; less than 2 m height sampling) were promising,
showing very good spectral differentiation among the six coniferous species tested: sugar pine (Pinus lambertiana), ponderosa pine (Pinus ponderosa), white fir (Abies concolor),
Douglas-fir (Pseudotsuga menziesii), incense cedar (Calocedrus decurrens), giant sequoia (Sequoiadendron giganteum),
and one hardwood species, California black oak (Quercus keloggii). In some cases, accuracies of greater than 91 percent were
obtained using sunlit samples alone. Their study also found
that the visible bands had higher discriminating power than
near-infrared bands (the best being blue-green followed by the
red-edge), although this contrasted with other findings (Gong et
al., 1997).Fung et al. (1999)used a spectroradiometer (210- to
1050-nmrange; controlled laboratory conditions) to construct a
hyperspectral database for the species studied. These species
included slash pine (Pinus elliottii), baldcypress (Taxodium
distichum), tallowtree (Sapium sebiferum), punktree (Mela-
Photogrammetric Engineering & Remote Sensing
Vol. 67, No. 12, December 2001, pp. 1367-1375.
Department of Forestry, College of Natural Resources, Virginia
Polytechnic Institute and State University, 319 Cheatham Hall
(0324),Blacksburg, VA 24061 (jaardtW.edu; wynneW.edu).
PHOTOGRAMMETRIC ENGINEERING 81REMOTE SENSING
0099-1112/0ll6712-1367$3.00/0
O 2001 American Society for Photogrammetry
and Remote Sensing
December 2001
1367
leuca quinquenervia), and bottletree (Firmiana simplex). The
first and second derivatives of the spectra were used in a linear
discriminant analysis, and the classification accuracies varied
from 56 percent to 91 percent. The original spectra tended to
produce better results than the first and second differences
(Fung et al., 1999).A classification of AVWS data into 11different forest-cover types, including red maple (Acerrubrum),red
oak (Quercus rubra), white pine (Pinus strobus), red pine
(Pinus resinosa), Norway spruce (Picea abies), and pure hemlock (Tsuga canadensis), as well as mixtures thereof, was
attempted and yielded an overall classification accuracy of 75
percent (Martin et al., 1998). This approach implemented a
maximum-likelihood classifier and was based on 11AVNS
bands previously used to derive relationships between foliar
chemistry (nitrogen and lignin concentration) and hyperspectral data. Avila et al. (2000) found differences in spectral
reflectance over the late growing season for ponderosa pine, but
not for Douglas-fir samples. Although averaged at canopy
level, preliminary results showed differences in reflectance
between new and old growth for these two species (Avila et al.,
2000).Even though in-fieldlairborne data are noisier than laboratory data due to atmospheric interferences, field data more
closely resemble operational conditions, hence the use of field
data, as opposed to laboratory data, in our study.
Given past contributions to the field of spectral classification of forestry species, it was felt that canopy-level spectral
separability using a ground-based spectroradiometer was still
lacking, especially for the six species chosen. The knowledge
gaps we address are as follows:
the quantification of inherent spectral separability among the
six species in question;
the determination of whether hyperspectral data are necessary
for separability among the species considered;
the identification of the most amenable pre-processing level;
and
given that these species are indeed separable, the determination
of the most effective wavelength ranges.
Materials and Methods
Sampling Methods
Spectral data were collected in the Appomattox Buckingham
State Forest in Appomattox County, Virginia on 16-18 August,
28 August, and 08 September 1999. The data were collected
using a Dodge bucket truck with a 15-mboom length. Because
canopy spectral data were used for this study, the spectral samples were taken from either above tree canopies or at least in
the top third of the crown of larger (taller than 15 m) trees. Care
was also taken to sample only sunlit portions of the crown so
as to avoid the inclusion of shadow in the samples. The data
were only collected between 1000 and 1500 (EasternDaylight
Time) each day to ensure high enough sun angles for adequate
lighting of each sample (reduction of shadow effects).During
the five sampling days, 291 samples were collected, this total
being limited by accessibility and weather conditions. An
effort was made to only collect a spectral sample when the sun
was not obscured by clouds or excessive haziness. Eventually
280 samples were used in the statistical analysis, with eleven
samples eliminated due to abnormal leaf discoloration or suspicious reflectance values. Leaf discoloration was especially
evident in some yellow poplar samples, with seasonal changes
(yellowing) and brown leaf spots occurring. The samples that
were identified as having suspicious reflectance values might
have been over-saturated at sampling time and had zero values
for most of the visible portion (350 nm to 700 nm) of the spectrum. The division of the samples per taxonomic group and
species is given in Table 1.Sequence and time of day of collection, as well as the foreoptic angle, were varied. Spatial randomization was sometimes difficult, because access with the
truck and heights of the trees that could be reached using a
15-mboom proved to be limiting. All these attempts at randomization, as well as controlling the viewing and illumination
geometry, were crucial because they allow for changes in canopy reflectance attributable to the canopy itself to be detected
(Curtiss and Goetz, 1994).
Equipment Used
The spectroradiometer and digital camera were used in the
bucket for taking the spectral reflectance reading and associThis study explored the possibility of distinguishing (at canopy ated recordings. A battery-operated fan and two umbrellas were
level) among six important forestry species in the southern
used to prevent overheating of the instrument in the summer
Appalachian region of Virginia on a spectral basis using
temperatures of up to 35 OC (95 OF).
overstory field spectra in the 350- to 2500-nm range (3- to 12Field spectra were colIected with the Analytical Spectral
nm spectral resolution). The four specific objectives were as
Devices, Incorporated FR (Full Range) Fieldspec spectroradiofollows:
meter. It uses a fiber optic bundle for light collection and covers
the range from 350 nm to 2500 nm. The light is projected from
to assess taxonomic group and species level separability of field the fiber optics onto a holographic diffraction grating where the
different wavelength components are separated and reflected
canopy spectra derived from the pine taxonomic group, reprefor collection by three different detectors. The visiblelnearsented by loblolly (Pinus taeda), Virginia (Pinus virginiana),
and shortleaf (Pinus echinata) pine, and the hardwood group,
infrared (WR) portion of the spectrum (350 to 1050 nm) is measrepresented by scarlet oak (Quercus coccinea), white oak (Quer- ured by a 512-channel silicon photodiode array and has a speccus alba), and yellow poplar (Liriodendron tulipifera);
tral resolution of approximately 3 nm at a 700-nm wavelength.
to evaluate the separability within and between these groups
The shortwave infrared (SWIR)region is measured by two detecas compared to separability using simulated Landsat data;
to identify the wavelength regions that define the inherent spec- tors, SWIRl(900to 1850 nrn) and SwR2 (1700 to 2500 nm). The
Objectives
tral separability between these taxonomic groups and species; and
to identify the data preprocessing techniques most suited as
preparation for species separability tests.
The species chosen are not well represented in literature,
but are very important to forestry operations in the southern
United States. All six species are significant contributors to
pineloaklmixed hardwood timber sales (Universityof Georgia,
2000), define certain ecosystems such as mixed, uneven-aged
hardwoods and pine plantations, and are fairly widespread in
the southern United States (Eyre, 1980).
1368
I
1%
$1
,
L
-11
11
TABLE1. NUMBER OF SAMPLES
PER TAXONOMIC
GROUP/SPECIES
USED FOR
STATISTICAL
ANALYSIS
Taxonomic group
Hardwoodsldeciduous
Softwoods/coniferous (Pines)
Species
Samples
Total
Scarlet oak
White oak
Yellow poplar
Loblolly pine
Shortleaf pine
Virginia pine
43
47
134
42
54
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
spectral resolution varies between 10 and 1 2 nm and the sampling interval is approximately 2 nm. The detectors convert the
incoming photons into electrons that are stored or integrated
until the detector reading is finished. This photoelectric current for each detector is then converted to a voltage, which is in
turn digitized by a 16-bit analog to digital converter. The digital
data are then transferred to the controlling computer resting on
top of the spectroradiometer (Beal, 1998).
For this study, an 8" foreoptic was attached to the bare fiber
bundle, enabling the user to point the instrument more accurately. This allowed for a field-of-view of approximately 1.4
units diameter at a sampling distance of ten units, where the
relatively small view angle facilitated the collection of pure
canopy spectra, mimicking species endmembers in an image.
Each reading consisted of a ten-sample average for that particular field-of-view.
As the instrument measures the intensity of the light
source through a given point in space, reflectance and transmittance are calculated using measurements from both the
unknown material (target) and a reference material with
known spectral (reflectanceltransmittance) properties. A white
reference material must have approximately 100 percent
reflectance across the entire spectrum. Therefore, the characteristics (change)of the light source are taken into account by
taking a "white reference" and using this reading as the denominator in the ratio (relative reflectance). This is especially crucial if light conditions vary even slightly in the field. It also
makes the comparison of samples taken at different times of day
and between days possible (Beal, 1998).A Spectralon surface of
5 by 5 centimeters was used to collect the white reference data
before each sample reading. Absolute reflectance can also be
obtained by multiplying the relative spectra with the calibration "spectrum" (specific to the white reference used) during
postprocessing (Beal,1998).
A crude device was attached to the pistol grip for measuring the angle of the reading. A Kodak DC260 Zoom digital camera was used to collect a digital
picture
of each tree sampled for
species verification purposes.
Analysis
SAS Version 7.00 TS Level OOPl software was used for the statistical analysis of the spectral data. The data sets used for statistical analysis varied by the type of smoothing and the
derivative function applied. A total of 1 5 data sets were evaluated, which included the raw relative reflectance data, a 9point weighted moving average filter, a 9-point moving median
filter, a 10-point static average filter (simulation of general
AVWS spectral resolution), a Landsat simulation data set, and
each of these data sets' first and second derivatives, as shown in
Figure 1.
In all of these data sets, the water absorption regions, identified through literature (Palacios-Oruetaa i d ~stinr1996;
Price,
1998)and visual inspection, were removed. They included the
spectral ranges 1350 to 1416 nm, 1796 to 1970 nm, and 2470 to
2500 nm.
The 9-point weighted averaged data sets had the following
weights:
where Y, is the weighted reflectance of the target for a particular channel, R is the reflectance measured by the instrument for
a particular channel, and n is the channel number (Rock et al.,
1994).The calculation of the first difference of reflectance data
used a two-cell (Ax = 1nm), zero-sum convolution filter that
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
approximates the first derivative of a function f(x)as follows
(Jain et al., 1995):
Applying the same template on the first derivative values
results in the second difference of the function f (x).
Given the availability of a limited number of samples versus the large number of variables present in the hyperspectral
data set, the reduction of data dimensionality is critical. Data
reduction is also necessary for determining correlation
between adjacent wavelengths in a sample (Hoffbeck and Landgrebe, 1996). Reduction of data dimensionality was done by
using a stepwise discriminant procedure (PROC STEPDISC in
SAS),which selects the variables that minimize within statistical group variance while maximizing the between group variance for a given a-level. An a-level = 0.0025 was used, as this
resulted in the selection of 5 to 20 variables from the 2150 original wavelengths. The variables chosen by the stepwise discriminant procedure were used as input for discriminant and
canonical discriminant procedures (PROC DISCRIM and CANDISC in SAS) to test between and within taxonomic group species separability. The results were verified using each data set's
own observations by running a cross-validation routine within
the discriminant procedure. Plots of the canonical variables
resulting from the canonical discriminant procedure were used
to view and validate results visually. The variance of the
canonical variables for each taxonomic group or species was
evaluated as a measure of the spectral tightness (similarity) for
that group. A data analysis flow chart is shown in Figure 1.
Results and Discussion
Spectral Data Smoothing
The effects of the three smoothing filters (moving 9-point
weighted average, moving %point median, static 10-point
average) for part of the short-wave infrared I1 region are shown
in Figure 2. The smoothing showed significant visual improvement only at wavelengths longer than 1400 nm. The smoothing
of raw reflectance curves plays an important role in further
derivative analysis, especially because derivative techniques
are very sensitive to noise (Tsai and Philpot, 1998).Timelspace
filters were chosen for this study, mainly because of their simplicity and ease of interpretation.
Statistical Analysis
The discriminant function showed all 45 spectral separability
tests (15 data sets, three tests per data set) significant at an alevel = 0.05 withP-values of 0.0001, using Wilks' Lambda test.
Spectral Separability between Deciduous and Coniferous Trees
Spectral discrimination between the deciduous and coniferous
groups was very successful, but this is not surprising, considering the distinct difference between these two groups in the
near-infrared reflectance region. The lowest cross-validation
result was 99 percent (9-point average data), with four of the
data sets having accuracies of 100 percent (raw relative
reflectance, second difference of raw relative reflectance, second difference of 9-point averaged data, second difference of 9point median data). Hardwood discriminant accuracy ranged
from 98 percent to 100 percent. Except in the case of the Landsat simulation data, pines were never misclassified, with at
most two hardwood samples being misclassified as belonging
to the pine group. This is probably due to the distinct spectral
characteristics of coniferous trees as a group, while, at least in
this case, there was more spectral variability in the deciduous
group. This is attributable in part to visible phenological
December 2001
1369
Orl$nd R d d w
ReUechuceDPt.
(Noisy absarptlon
r d o m removed)
I
I
I
Mo-
Avs*
(~-POIII~:
mmtbed &a)
Raw Data
~elatlw
~~n-ce
raw data)
I
Movlng Medlan
opdnt:
moothed dntn)
I
I
I
S d cA v s q
(10-POW:
10 nm rerolntlon
moothed data)
Landsat Data
I
I
I
(d~~~~pted:
Bmds 1-5, 7)
Each r ~ ~ ~ e c t l v e p r e - p r o cdata
~~m
s ed
t and fist and seeond
dlllermcer thssof - b a w w n m d wltbln group mpprnblllty bdr
(45 t& ln total: 15 data sets;
3 t a t r p e r data &)
I
1
StspalseMserlmlnant Andyds
@ a w n groups m d wltbln each group)
Alpha-level = 0.0025
I
Ms&hmt
I
m d Canonlcd DlmLmlnmt Andydc
Results rhowlns betaem groups m d alW each group lor d forty-five
p e c t r d separability bdr (15 dntn rdr; 3 tern p s data s&)
Cross-vrlldPtlon r e d t s and canodcd plots
Evaluntlon of corrdatlonr
Figure 1. Data analysis flow chart.
10-polat Bhdc Average
&I
!
f
9-point M o h g Avenge
e 21
9-potot Mo*
-a
fi
Medba
Its" Pehtlve ReUesbme
2000
2050
2100
2150
2200
2250
2300
Wavelength (m)
Figure 2. Smoothing of the spectral data-a visual comparison of the three methods used (a white oak spectral sample; short-wave infrared II portion).
changes, particularly in yellow poplar (slight yellowing in
some leaves). Even the "coarser" AVIRIS simulation data with a
10-nm resolution never had accuracies lower than 99 percent
for hardwoods and 100 percent for pines. Accuracies for the
Landsat simulation data ranged from 93 percent (second difference of the Landsat simulation data) to 96 percent (raw Landsat
simulation data), with at most one pine sample and 19 hardwood samples being misclassified. These results indicate the
existence of inherent spectral differences (leaf-on)between the
deciduous and coniferous species studied, confirming the
1370
December 2001
spectral separability of these two groups at an operational level
(Nelson et al., 1984;Frank, 1988;White et al., 1995).
Spectral Separability among the Three Deciduous Species
Spectral discrimination among the three hardwood species
proved less accurate, but was still higher than 80 percent for
most of the 45 tests perfomed, with the discriminant classification accuracy ranging between 78 percent and 93 percent. The
best results were those of the second difference of the 10-nm
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
simulation data, the second difference of the 9-point
average data (all at 93 percent), and the first difference of the 9point median data (92 percent). Although the overall discrimination accuracy was the same for all three of the best performing data sets, the species-specific accuracies varied
among them. In the case of the second difference of the 10-nm
AWUS simulation data, scarlet oak, white oak, and yellow poplar showed accuracies of 91 percent, 89 percent, and 98 percent,
respectively (the first difference of the 9-point median relative
reflectance had a lower percentage only for white oak at 87
percent).
The second difference of the 9-point average data set
showed accuracies of 88 percent, 91 percent, and 98 percent,
for scarlet oak, white oak, and yellow poplar, respectively.
These results are very encouraging, especially in the case of
the 10-nm resolution AVIRIS simulation data, which mimics
the resolution available in most operational airborne
sensors.
The worst performer was the second difference of the raw
relative reflectance (scarlet oak: 65 percent, white oak: 74 percent, yellow poplar: 95 percent), with a discriminant accuracy
of 78 percent. This is likely due to the sensitivity of the difference operator to noise in unsmoothed data sets.
Accuracies for the Landsat simulation data ranged from 52
percent (second difference Landsat simulation data) to 70 percent (raw Landsat simulation data). The Landsat sensor's spectral resolution does not seem to be adequate for spectral
discrimination among these three hardwood species in midsummer.
There was considerable confusion between scarlet and
white oak, which might be expected given that they belong to
the same genus. This confusion was especially evident at the
non-derivative data set level (i.e., there was very little spectral
magnitude separation in the genus Quercus). The discrimination accuracy for white oak ranged between 74 percent (second
difference of raw relative reflectance, 9-point median data, second difference of the 9-point median data) and 91 percent (second difference of the 9-point averaged data). For scarlet oak the
range was 65 percent (second difference of raw relative
reflectance)to 91 percent (first difference of the 9-point median
S
data). In comparidata, second difference of A ~ simulation
son, accuracies for the Landsat simulation data ranged from 43
percent (second difference data) to 66 percent (raw data) and
47 percent (second difference data) to 65 percent (raw data), for
white oak and scarlet oak, respectively. Yellow poplar's high
discriminant accuracy can, in turn, probably be attributed to its
higher reflectance (lighter colored leaves).
The importance of the difference operators is also underlined, because these "curve shape descriptors" seem to be amenable to within species spectral discrimination. Even the first
difference of the raw relative reflectance data set had high
cross-validationresults, with scarlet oak, white oak, and yellow
poplar results being 86 percent, 89 percent, and 95 percent,
respectively.
Spectral discrimination among these three deciduous species at an operational level is likely, but one has to consider (1)
the unavailability of high spectral resolution airborne and
spaceborne sensors such as the one used in this study and (2)
that the AVIRIS simulation data give only a rough indication of
spectral data with a 10-nm resolution and are not true to A m S
operational conditions (e.g., differences in wavelength sensitivities and probable atmospheric effects were not taken into
account). Even given these caveats, the very promising crossvalidation result from the second difference of the A m s simulation data is an indicator of possible operational implementation. The best cross-validation results (second difference of the
10-nmresolution AWS simulation data selected as an example) for within deciduous group spectral separability are shown
in Table 2.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Spectral Separability among the Three Coniferous Species
The cross-validation accuracy percentages for the three pine
species were generally lower than those for the hardwoods.
This can probably be attributed to the Pinus genus having less
spectral variation among species, rendering pine species spectrally less separable. The results for the three pine species
ranged between 62 percent (second difference of raw relative
reflectance) and 84 percent (second difference of the 9-point
averaged).The second difference of the 9-point averaged data
set showed discrimination accuracies of 90 percent, 83 percent,
and 80 percent for loblolly, shortleaf, and Virginia pine, respectively. The first and second differences of the smoothed data
sets were better for spectral separation among the species than
were the differenced data sets of the raw relative reflectance.
There does not seem to be one pine species in particular that is
more separable than any other. The second difference of the raw
relative reflectance (62 percent), as one of the least accurate
data sets, again illustrates the second difference's sensitivity to
noisy data. Accuracies for the Landsat simulation data ranged
from 45 percent (second difference of the Landsat simulation
data) to 60 percent (raw Landsat simulation data).
The most confusion was present between Virginia pine and
the other two species, followed by shortleaf pine and lastly loblolly pine. Loblolly pine's cross-validation results ranged
between 54 percent (second difference of the raw relative
reflectance) and 90 percent (second difference of the 9-point
averaged data). The cross-validation ranges for shortleaf and
Virginia pine ranged from 62 percent (9-point median data) to
88 percent (first difference of the 9-point averaged data) and 56
percent (second difference of the raw relative reflectance) to 87
percent (first difference of the 9-point median data), respectively. The associated simulated Landsat data accuracies
ranged between 58 percent (second difference data) and 70
percent (raw data) for loblolly pine, 14 percent (second difference data) and 55 percent (raw data) for shortleaf pine, and 50
percent (first difference data) and 56 percent (second difference
data) for Virginia pine. Loblolly pine tends to have a denser
canopy leaf structure compared to Virginia and shortleaf pine,
with this perhaps being the reason for the lowest confusion
and generally better cross-validation results. The best crossvalidation results (second difference of the 9-point averaged
data) for within coniferous group spectral separability are
shown in Table 3. The lack of good results for the within group
discrimination using Landsat simulation data, likely due to its
inability to resolve fine spectral differences, shows clearly that
one-point-in-time Landsat data are not useful for spectral discrimination among these species.
V a a i a k SlgnlRcant for Spectral Dlscrlmination
The wavelength variables (nm) indicated by the stepwise discriminant procedure to be significant at (Y = 0.0025 for between
group and among species discrimination using the second difference of the %point averaged data set, one of the best performing data sets, were distributed across the entire 350-nm to
2500-nm range (shown in Figure 3). When considering the
variables that were shown to be significant for each data set and
separability test, there are some notable trends:
The shortwave infrared I1 (smz: 1700 to 2500 nm) region is
poorly represented for almost all the separability tests. This is
especially true when relative reflectance data, and not the difference data, are considered. The s m z region is better represented in the difference data. This phenomenon might be due
to spectral differences that exist in curve shape, and not curve
magnitude, between and within taxonomic groups. This is substantiated by Martin et al.'s (1998) work in which AVINs data
(coarser 10-nm resolution] were used, and longer wavelengths
also proved beneficial. Derivative noise, where it occurs in
abundance, might also be significant because of chance contribution alone.
December 2001
1371
RESULTS FOR WITHINHARDWOOD~ E P A R A T(SECOND
~ B ~ L ~DIFFERENCE
N
1 0 - N RESOLLITION
~
AVlRlS S~MULAT~ON
DATA)
TABLE2. BESTCROSS-TABULATION
Scarlet oak
FromlTo
n
~~~~~
Scarlet oak
White oak
Yellow poplar
White oak
Yellow poplar
Total
%
n
%
n
%
n
%
Discrimination %
91%
11%
0%
4
42
9%
89%
2%
0
0
43
0%
0%
98%
43
47
44
100%
100%
100%
93% (134 samples)
-
39
5
0
1
TABLE3. BEST CROSS-TABULATION
RESULTS FOR WITHINPINE
SEPARABILIM
(SECOND
DIFFERENCE%POINT
AVERAGEDATA)
Loblolly pine
Shortleaf pine
Virginia pine
FromlTo
n
90
n
%
n
Loblolly pine
Shortleaf pine
Virginia pine
45
3
7
90%
7%
7%
3
35
6%
83%
13%
4
43
7
w-
a
..
I-
I ,"',li
a
r#
M; -
0
0.
500
loo0
1500
2000
.
.. .
.
2500
Wavelength (nm)
Figure 3. Significant variables (stepwise procedure; a =
0.0025) using the second difference of the Spoint averaged
data set for the between group (BG), within hardwoods (WH),
and within pines (WP) separability tests.
2
Total
Yo
n
'%o
Discrimination %
4%
10%
80%
50
42
54
100%
100%
100%
84% (146 samples)
reduction from the initial 2150 variables to as few as five reduces
the likelihood of overfitting.
The higher order second difference is not useful when the data
have been smoothed by a relatively large (10-point),static filter.
This can be attributed to the suppression of local curve shape
information by the smoothing algorithm, i.e., the shape is generalized instead of well defined.
Although specific wavelengths associated with stretching
in 0-H bonds (970 nm, 1200 nm, 1400 nm, and 1940 nm) do not
seem to stand out from the data as a whole, these wavelengths
are represented in some of the data sets and tests, but not to such
an extent that it might be deemed significant. Wavelengths in
the 2036- to 2180-nm range related to nitrogen concentration
(McClellanet al., 1991; Bolster et al., 1996; Kokaly and Clark,
1999),as well as 1555nrn (Wessman et al., 1989),seemto be represented as well. The representative distribution across the
instrument's spectral range indicates the usefulness of the
whole wavelength range for discriminating between taxonomic
groups and among species.
Canonical Dlscrimlnant Analysls
The canonical discriminant analysis plots shown serve as
visual indicators of spectral separability among species and
the spectral similarity (tightness) within each species. The variances of the canonical variables further supplement the visual
interpretations with quantitative measures of the within taxoA strong presence of the visible wavelengths in all the separabil- nomic group and species spectral similarities.
For the between taxonomic group tests, the variance for the
ity test cases is striking. This highlights the usefulness of the
visible spectrum for species and taxonomicgroup discrimination canonical variable was much higher for the hardwood than for
the pine group in each case, with examples being the first differutilizing the subtle differences not detectable by the human
eye, and corresponds with the results foundby Gong et al. (1997). ence of the raw relative reflectance (pines: 0.606, hardwoods:
The near-infrared and shortwave infrared region I (SWIRI:
900 1.429),the raw 9-point averaged data (pines: 0.593, hardwoods:
to 1785 nm) are also very well represented in all the test
1.443),the first difference of the 9-point averaged data (pines:
cases. This is not an unexpected occurrence, because the near- 0.569, hardwoods: 1.47),the raw 9-point median data (pines:
infrared region is well known for its information content in
0.589, hardwoods: 1.448),and the raw AVIRI~simulation data
vegetation studies. The s m l region is especially important
for between group (coniferous versus deciduous) discrimina- (pines: 0.543, hardwoods: 1.50). This is very indicative of the
tion. Even though AVIRIS data were used, Martin et al. (1998) higher spectral variation present in the deciduous or hardalso found these regions of importance in species discrimina- wood group as opposed to the coniferous or pine group.
For the within hardwood separability tests, yellow poplar
tion, but a strong presence of bands greater than 2000 nm was
also found. The presence of wavelengths in the 700- to 750seemed to form a very distinct grouping compared to the two
nm range is encouraging, because Luther and Carroll (1999)
oak species (scarlet and white oak). Yellow poplar had, howfound high correlations [RZ2 0.75) between the 711-nm band ever, the highest pooled variance for the two canonical variables
and chlorophyll a. This is confirmed through findings by
in most of the test cases, one example being the second differKupiec and Curran (1993), who found an R2 = 0.96 when
ence of the 9-point averaged data (yellow poplar: 1.313,white
regressing bands 723 nm, 1552 nm,and 2371 nm with chlorooak: 0.82, scarlet oak: 0.877). Only scarlet oak had a higher
phyll concentration in slash pine (Pinus elliottii).
pooled variance for the second difference of the 9-point median
The wavelengths that are useful for between taxonomic group
(yellow poplar: 1.036, white oak: 0.926, scarlet oak: 1.044) and
discrimination at cr = 0.0025 are limited to between 5 (within
the raw A w S simulation data (yellowpoplar: 1.066,white oak:
pines, 9-point average relative reflectance] and 20 (between
groups, second difference of raw relative reflectance]. This
0.853, scarlet oak: 1.093).The two oak species generally had
1372
December 2001
much lower variances than yellow poplar, but there was clearly
a higher degree of confusion between the oaks.
As far as the pine species are concerned, Virginia pine had
the highest pooled variance in most cases with loblolly (second difference of the raw relative reflectance) and shortleaf
pine (first difference of the 9-point median data) having a
higher pooled variance in only one test case each. Examples are
the first difference of the 9-point averaged data (loblolly pine:
0.877, shortleaf pine: 0.584, Virginia pine: 1.435) and the first
difference of the A W S simulation data (loblolly pine: 0.899,
shortleaf pine: 0.588, Virginia pine: 1.412). There generally
seemed to be a higher degree of confusion among the pine species as opposed to the hardwood species, with many
"confusedlmisclassified" samples present on the periphery of
each of the pine groups. As examples, the canonical plots for
the data sets that yielded the best and worst discriminant accuracy for the within deciduous species separability, namely the
second difference of the simulated 10-nm AVIRIS data set and
the second difference of the raw relative reflectance, are shown
in Figures 4a and 4b, respectively.
4-
2 -
¤
8
v
r
v
.
v
va
VZWV
8@ $
rn
C
.
-2-
+a
l
l
a
-
5
-7
,
-6
,
-5
,
i
-3
-4
,
,
-2
-1
1
,
1
0
,
2
,
3
,
4
,
,
5
6
7
Canonical vdable 1
(a)
-
4
Conclusions
The use of hyperspectral data for between and within forest taxonomic groups (deciduous versus coniferous) classification
has great potential, but is not without pitfalls. From an analytical perspective the greatest challenges are data smoothing and
reduction to those variables that truly delineate taxonomic
group or species differences.From an operational perspective,
the inherent spectral separability of some forest groups and
species is not conclusive, with much confusion found among
especially the coniferous (pine)species.
.
3-
a
32
N
-
l
+
18
v
-1 -2
.
l
.via..'
.
-
Spectral Separability between Groups and among Species
As expected, the between group separation (deciduous versus
coniferous species) had very high (99 percent to 100 percent)
cross-validation accuracies. The highest accuracy for within
hardwood classification was in the low 90th percentile. The
three coniferous species studied had lower cross-validation
results than did the deciduous species for within species discrimination, with the best accuracy being in the mid-80th percentile. Operational conditions will likely lower these
accuracies.
SignlRcant Variables
Variables shown by the stepwise discriminant procedure to be
significant for discriminating between taxonomic groups and
among species were distributed across the entire 350- to 2500nrn range. Variables for the different data sets were largely limited to the visible and near-infrared regions, with very few variables in the short-wave infrared region at wavelengths greater
than 2000 nm. This confirms the findings by Gong et al. (1997),
who identified the visible portion of the spectrum as being
important for species discrimination, and Martin et al. (1998),
who identified wavelengths in the near-infrared regions as
being important for spectral separation between species. For
between taxonomic group tests using raw reflectance,the use of
near-infrared variables was especially striking, with the significantly higher reflectance of deciduous species relative to coniferous species again being confirmed. Specific wavelengths
shown to be significant by literature for characterizingnitrogen
and chlorophyll content were among the variables chosen for
their discriminative power (Kupiec and Curran, 1993; Luther
and Carroll, 1999).Variables greater than 1800 nm were better
represented in the second difference data sets than in the raw
data, which might indicate that this spectral region has greater
value for separation studies when the shape (slope of the slope,
curvature),and not the magnitude of the curves, is considered.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
-5
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
Canonical variable 1
(b)
Figure 4. (a) Canonical variable plot: Second difference of
10-nm data set for hardwood spectral separability (best performing data set-within deciduous species). (b) Canonical
variable plot: Second difference of the raw relative
reflectance data for hardwood spectral separability (worst
performing data set-within deciduous species).
PreProcesslngand Analysls Techniques
Canonical discriminant analysis was an effective means of
assessing specieslgroup spectral separability in an analytical
context. Analogous use in more operational contexts is also
possible. For example, Zhoa and Maclean (2000) have shown
canonical discriminant analysis to yield more accurate classifications than principal component analysis when used as a spectral transformation technique for forest type delineation using
Landsat TM data. The canonical discriminant analysis,
although not without pitfalls (e.g., species specificity), maximizes between group variation when a priori knowledge about
classes exists, which in turn increases classification accuracy
(Zhoa and Maclean, 2000).
The use of derivative techniques to enhance spectral differences between and within taxonomic groups proved to be very
effective, contrary to findings by Fung et al. (1999). Given the
sensitivity of derivative operators to noisy data, spectral data
smoothing does appear to be a prerequisite for the use of derivatives. The second difference of the 9-point averaged data, for
December 2001
1373
example, was one of the data sets with the highest cross-validation accuracies (84 percent to 100 percent) for all three separability tests. A moving weighted average or median filter does
seem preferable over a static filter, because the latter may suppress localized curve shape (difference operators) because of its
more general smoothing effect.
Urnitations
Among the limitations of this study were the following:
the difficulty of randomizing sampling points;
the angle variation (0" to 15') of the spectroradiometer's
foreoptic, which allowed for relatively large variations in
reflectance magnitude;
the noisy SWIRZ (1700- to 2500-nm) region; and
the possibly limited applicability to AWS (or similar hyperspectral) data, given that (a) the AVIRI~sensor's spectral sensitivity to different wavelengths was not addressed in the static
average, (b) the AWS sensor has a lower signal-to-noise ratio
in the visible portion of the spectrum when compared to the
Fieldspec FR spectroradiometer used, and (c) the requisite
smoothing may additionally reduce the effective AVIRIS spectral resolution.
Future Research
It is recommended that further research focus on identifying
and testing the spectral differentiation among different tree
types (e.g.,white oaks versus red oaks) within a genus of interest. This might be a very enlightening research avenue,
because the natural variation present in many species might not
allow for accurate species differentiation from a spaceborne
platform. As was seen in the case of the distinct yellow poplar
and oak groupings, spectral separation at the genus level will
probably have to be ascertained prior to attempts to separate
species within a genus. The use of other techniques (such as
continuum removal, cross-entropy as a spectral dissimilarity
measure, and the use of vegetation spectral features such as the
red-edge inflection point or chlorophyll absorption wel1)'can
also be evaluated. The effect of seasonal variation on spectral
differentiation of taxonomic groups and species and its quantification also deserves attention, as do the effects of varying soil
and growing conditions, crown closure, and tree age, which are
all closely related to seasonal variation. The success of spectral
unmixing techniques given these variable factors also needs to
be established.
Following the determination of spectral separability of different genera, the identification of distinct types within each
genus of interest (e.g.,the white oaks versus the red oaks) might
be the next step toward discrimination among species on a
spectral basis. The accurate, categorically specific type maps
thus produced would be beneficial to both forest monitoring
and management.
Acknowledgments
This research was supported by the J. William Fulbright Scholarship Program (Institute of International Education), the Center
for Earth Observing and Space Research at George Mason University, the National Council for Air and Stream Improvement,
the McIntire-Stennis Research Program (VA-136589), the Virginia Department of Forestry, and the Department of Forestry
at Virginia Polytechnic Institute and State University.
The authors would like to thank John Scrivani (Virginia
Department of Forestry) and the staff at the Appomattox Buckingham State Forest for their willing support, as well as Jared
Wayman (Westvaco),Glenn Fields, Marleen van Aardt (Virginia Tech), Sorin Popescu (Virginia Tech), Rebecca Musy (Virginia Tech), and Johnny Harris (Virginia Department of Forestry) for spending many hours on various aspects of this study.
Their support and help were invaluable to the success of this
effort.
1374
December 2001
Our sincere thanks go to Drs. Richard G. Oderwald and
James B. Campbell for advice in their respective areas of expertise, and to Michele Marini and Ilya Lipkovich at the Statistical
Consulting Center at Virginia Tech for their guidance in the statistical analysis of this study. We would like to thank Cheryl
Albig for editorial comments and help in the preparation of the
final draft of this manuscript.
References
Avila, R.A., P.E. Gessler, and M. Bienkowski, 2000. Monitoring Douglasfir and ponderosa pine spectra1 reflectance in northern Idaho,
Proceedings of ASPRS Annual Conference 2000, 22-26 May,
Washington D.C., 11 p.
Beal, D., 1998. Fieldspec" User's Guide, Analytical Spectral Devices
Incorporated, Boulder, Colorado, 85 p.
Birk,
T.B. McCord, 1994. ~ i r b o i hyperspectral
e
sensor systems, Proceedings of the Institute of Electrical and Electronics
Engineers, Inc. OEEE) National ~ e r o s ~ a and
c e Electronics Conference 1994, 23-27 May, Dayton, Ohio, pp. 309-318.
Bolster, K.L., M.E. Martin, and J.D. Aber, 1996. Determination of carbon
fraction and nitrogen concentration in tree foliage by near infrared
reflectance: A comparison of statistical methods, Canadian Journal
of Forest Research, 26:590-600.
Curran, P.J., 1989. Remote sensing of foliar chemistry, Remote Sensing
of Environment, 30271-278.
Curtiss, B., and A.F.H. Goetz, 1994. Field spectrometry: Techniques
and instrumentation, Proceedings of ISSSR 1994, 10-15 July, San
Diego, California, 1:195-203.
Eyre, EH., (editor), 1980. Forest Cover lSIpes of the United States and
Canada, Society of American Foresters, Washington, D.C., 148 p.
Frank, T.D., 1988. Mapping dominant vegetation communities in the
Colorado Rocky Mountain front range with Landsat Thematic
Mapper and digital terrain data, Photogrammetric Engineering &
Remote Sensing, 54(12):1727-1734.
Franklin, S.E., 1994. Discrimination of subalpine forest species and
canopy density using digital CASI, SPOT PLA and Landsat TM
data, Photogmmmetric Engineering 6 Remote Sensing,
60(10):1233-1241.
Fung, T., F.Y. Ma, and W.L. Sui, 1999. Hyperspectral data analysis for
subtropical tree species identification, Proceedings of ASPRS
Annual Conference 1999, 17-21 May, Portland, Oregon, 10 p.
Gong, P., R. Pu, and B. Yu, 1997. Conifer species recognition: An
exploratory analysis of in situ hyperspectral data, Remote Sensing
of Environment, 62:189-200.
Hoffbeck, J.P., and D.A. Landgrebe, 1996. Classification of remote sensing images having high spectral resolution, Remote Sensing of
Environment, 57:119-126.
Jain, R., K. Rangachar, and B.G. Schunck, 1995. Machine Vision, MIT
Press and McGraw-Hill, Inc., New York, 549 p.
Kokaly, R.F., 2001. Investigating a physical basis for spectroscopic
estimates of leaf nitrogen concentration, Remote Sensing of Environment, 75:153-161.
Kokaly, R.F., and R.N. Clark, 1999. Spectroscopic determination of leaf
biochemistry using band-depth analysis of absorption features
and stepwise multiple linear regression, Remote Sensing of Environment, 67:267-287.
Kupiec, J., and P.J. Curran, 1993. AVIRIS spectra correlated with the
chlorophyll concentration of a forest canopy, Summaries of the
Fourth Annual JPL Airborne Geoscience Workshop 1993, 25-29
October, Washington, D.C., (JPL Publication 93-26, Vol. I), pp.
105-108.
Lawrence, W.T., Y.E. Shimabukuro, and B. Gao, 1993. Analysis of
hyperspectral AVIRIS image data over a mixed-conifer forest in
Maine, Proceedings of the 7th Brazilian Remote Sensing Symposium 1993, 10-14 May, Curitiba, Brazil, pp. 129-155.
Luther, J.E., and A.L. Carroll, 1999. Development of an index of balsam
fir vigor by foliar spectral reflectance, Remote Sensing of Environment, 69:241-252.
Martin, M.E., S.D. Newman, J.D. Aber, and R.G. Congalton, 1998.
Determining forest species composition using high spectral resolution remote sensing data, Remote Sensing of Environment,
65:249-254.
and
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
-
McClellan, T.M., M.E. Martin, J.D. Aber, J.M. Melillo, K.J. Nadelhoffer,
and B. Dewey, 1991. Comparison of wet chemistry and near infrared reflectance measurements of carbon-fraction chemistry and
nitrogen concentration of forest foliage, Canadian Journal of Forest Research, 21:1689-1693.
Nelson, R.F., R.S. Latty, and G. Mott, 1984. Classifying northern forests
using Thematic Mapper simulator data, Photogrammetric Engineering b Remote Sensing, 50(5):607-617.
Palacios-Orueta, A.P., and S.L. Ustin, 1996. Multivariate statistical
classification of soil spectra, Remote Sensing of Environment,
57(2):108-118.
Price, J.C., 1998. An approach for analysis of reflectance spectra,
Remote Sensing of Environment, 64:316-330.
Rock, B.N., D.L. Williams, D.M. Moss, G.N. Lauten, and M. Kim,
1994. High-spectral resolution field and laboratory optical
reflectance measurements of red spruce and eastern hemlock
needles and branches, Remote Sensing of Environment,
471176-189.
Shen, S.S., G.D. Badwar, and J.G. Carnes, 1985. Separability of boreal
forest species in the Lake Jennette area, Minnesota, Photogrammetric Engineering b Remote Sensing, 51(11]:1775-1783.
Tsai, F., and W. Philpot, 1998. Derivative analysis of hyperspectral
data, Remote Sensing of Environment, 61341-51.
University of Georgia, 2000. 7Ymber Mart-South: 4th Quarter 2000,
Yearly Summary and Virginia StumpagelDelivered Prices, Daniel
B. Warnell School of Forest Resources, Athens, Georgia, 306022152, 6 p.
Wessman, C.A., J.D. Aber, and D.L. Peterson, 1989. An evaluation of
imaging spectrometry for estimating forest canopy chemistry,
Remote Sensing of Environment, 10(8):1293-1316,
White, J.D., C.K. Glenn, and J.E. Pinder 111, 1995. Forest mapping at
Lassen Volcanic National Park, California, using Landsat TM data
and a geographical information system, Photogrammetric Engineering b Remote Sensing, 61(3):299-305.
Yoder, B.J., and R.E. Pettigrew-Cosby, 1995. Predicting nitrogen and
chlorophyll content and concentrations from reflectance spectra
(400-2500 nm) at leaf and canopy scales, Remote Sensing of Environment, 53:199-211.
Zhoa, G., and A.L. Maclean, 2000. A comparison of canonical discriminant analysis and principal component analysis for spectral transformation, Photogrammetric Engineering b Remote Sensing,
66(7):841-847.
(Received 22 August 2000; accepted 11May 2001)
Forthcoming Articles
Adrian L. Bugg, R a y D. Spencer, a n d Alex Lee, Applying
G I s for Developing Regional Forest Agreements i n
Australia.
Kai-Yi Huang, A Synergistic Automatic Clustering Technique (SYNERACT) for Multispectral Image Analysis.
Tao Cheng, Fuzzy Objects: Their Changes a n d Uncertainties.
Hiroko Mizukoshi a n d Masamu Aniya, Use of ContourBased DEMs for Deriving a n d Mapping Topographic
Attributes.
Michele Crosetto, Francesco Crosetto, a n d Stefano
Tarantola, Optimized Resource Allocation for GISBased Model Implementation.
Kazuo Oki, Hiroyuki Oguma, a n d Mikio Sugita, Subpixel
Classification of Alder Trees Using Multitemporal
Landsat Thematic Mapper.
Amod Sagar Dhakal, Takaaki A m a d a , Masamu Aniya, a n d
Rishi Ram Sharma, Detection of Areas Associated with
Flood a n d Erosion Caused b y a Heavy Rainfall Using
Multitemporal Landsat TM Data.
Albert J. Peters, Elizabeth A. Walter-Shea, Lei Ji, Andr6s
Vifia, Michael Hayes, a n d M a r k D. Svoboda,Drought
Monitoring w i t h NDVI-Based Standardized Vegetation
Index.
Christ0pherJ.S. Ferro a n d Timothy Warner, Scale a n d Texture i n Digital Image Classification.
Jonathan H. Smith, James D. Wickham, Steven V.
S t e h m a n , a n d Limin Yang, Impacts of Patch Size a n d
Land-Cover Heterogeneity o n Thematic Image Classification Accuracy.
Susumu Hattori, Keiichi Akimoto, Clive Fraser, a n d
Harutaka Imoto, Automated Procedures w i t h Coded
Targets i n Industrial Vision Metrology.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING