Remote Sensing of Environment 80 (2002) 143 – 156
www.elsevier.com/locate/rse
A comparison of methods for monitoring multitemporal vegetation
change using Thematic Mapper imagery
John Rogana,*, Janet Franklina, Dar A. Robertsb
a
Department of Geography, San Diego State University, San Diego, CA 92182-4493, USA
b
Department of Geography, University of California, Santa Barbara, CA 93106, USA
Received 1 March 2001; received in revised form 21 June 2001; accepted 21 July 2001
Abstract
Forested ecosystems in California are undergoing accelerated change due to natural and anthropogenic disturbances. Change detection is
a remote sensing technique used to monitor and map landcover change between two or more time periods and is now an essential tool in
forest management activities. We compared the ability of two linear change enhancement techniques, the Multitemporal Kauth Thomas
(MKT) and Multitemporal Spectral Mixture Analysis (MSMA), and two classification techniques, maximum likelihood (ML) and decision
tree (DT), to accurately identify changes in vegetation cover in a southern California study area between 1990 and 1996. Supervised
classification accuracy results were high ( > 70% correct classification for four vegetation change classes and one no-change class) and
showed that (1) the DT classification approach outperformed the ML classification approach by 10%, regardless of the enhancement
technique used, and (2) using DT classification, MSMA change fractions [i.e., green vegetation (GV), nonphotosynthetic vegetation (NPV),
shade, and soil] outperformed MKT change features (i.e., change in brightness, greenness, and wetness) by 5%. D 2002 Elsevier Science
Inc. All rights reserved.
Keywords: Vegetation change; Spectral mixture analysis; Decision tree classification; Remote sensing
1. Introduction
The importance of mapping, quantifying, and monitoring changes in the physical characteristics of forest cover
has been widely recognized as a key element in the study
of global change (Nemani & Running, 1996). Forested
ecosystems actively regulate biosphere dynamics by
means of contributions to the carbon and nitrogen cycles
(Aber, 1992), the hydrological cycle (Running & Gower,
1991; Running & Nemani, 1991), and the surface energy
balance (Choudhury, 1994). In addition, the economic and
ecological importance of forests is well understood
(Franklin & Woodcock, 1997; Franklin, Woodcock, &
Warbington, 2000; Skole & Tucker, 1993). However, the
location and rates of forest structural change and the
degree to which landscapes respond to these disturbances
require extensive investigation (Borak, Lambin, & Strahler,
2000; Lambin, 1998).
* Corresponding author. Tel.: +1-619-594-8034; fax: +1-619-594-4938.
E-mail address: [email protected] (J. Rogan).
Among the large variety of forested ecosystems, Mediterranean-type ecosystems (MTEs) comprise a vegetation
form with worldwide importance and distribution and are of
special interest to global change research in terms of
biodiversity, desertification, water resources, and urbanization (Hill, Megier, & Mehl, 1995; Moreno & Oechel, 1995;
Sala et al., 2000; Verstraete & Schwartz, 1991). In southern
California, for example, accelerated anthropogenic disturbances, such as urban expansion (Stephenson & Calcarone,
1999), fire suppression (Minnich, Barbour, Burk, & Fernau,
1995), and the invasion of nonnative plants (O’Leary &
Westman, 1988), interact with natural disturbances, such as
fire and postfire succession (Keeley, Fotheringham, &
Morais, 1999; Zedler, Gautier, & McMaster, 1983) and pest
infestation (Hope, 1995), to significantly alter vegetation
cover in forests, woodlands, and shrublands.
Despite the recognized biotic diversity of southern
California and emphasis placed on the need for forest
monitoring activities by resource agencies (e.g., Southern
California Association of Governments and USDA Forest
Service), there has been a lack of comprehensive, large-area
change detection studies in the region (Levien et al., 1999;
0034-4257/01/$ – see front matter D 2002 Elsevier Science Inc. All rights reserved.
PII: S 0 0 3 4 - 4 2 5 7 ( 0 1 ) 0 0 2 9 6 - 6
144
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Waller, 1999). Satellite imagery, however, provides a viable
source of data from which updated landcover information
can be extracted in order to inventory and monitor changes
in vegetation cover in southern California (Franklin &
Woodcock, 1997; Franklin et al., 2000; Roberts, Green, &
Adams, 1997).
Three properties of vegetation influence the spectral
quantity and quality of solar reflected radiation received
by satellite sensors: abundance, composition, and condition
(Stow, 1995). Numerous researchers have addressed the
problem of accurately mapping changes in these properties
in a variety of forested habitats (Adams et al., 1995; Collins
& Woodcock, 1996; Coppin & Bauer, 1996; Gunilla,
Olsson, Austerheim, & Grenne, 2000; Radeloff, Mladenoff,
& Boyce, 2000; Roberts, Batista, Pereira, Waller, & Nelson,
1998; Singh, 1989; Sujatha, Dwivedi, Sreenivas, & Venkataratnam, 2000). These studies have reported either the
statistically significant correlations between change-related
features and satellite spectral response or the spatial accuracy and associated errors of a classified landcover changemap, but few have quantitatively compared the numerous
change enhancement and classification techniques available
(Macleod & Congalton, 1998; Mas, 1999; Smits, Dellepiane, & Schowengerdt, 1999).
A review of change detection studies in MTEs revealed
that, in addition to the standard problems encountered in
change studies (i.e., registration error, variation in atmospheric illumination, and sensor variability), the effects of
topography, heterogeneous vegetation types, and interannual phenological variability (caused by variable precipitation patterns) can produce errors in identifying interdate
change (Levien et al., 1999; Shoshany, 2000; Waller, 1999).
Most studies in MTEs have employed the two most widely
used change detection approaches, image differencing and
postclassification comparison, producing varied results.
Postclassification change detection techniques have significant limitations (Macleod & Congalton, 1998; Stow,
1995) because the comparison of landcover classification
for different dates does not allow the detection of subtle
changes within landcover classes (Biging, Colby, & Congalton, 1998). Further, the change map product of two
classifications often exhibits accuracies similar to the product of multiplying the accuracies of each individual classification (Stow, 1995).
A spectral mixture analysis (SMA) approach was
employed by Hill et al. (1995), resulting in high agreement
between vegetation and soil fractions and field-based
information for a single date. Multitemporal SMA (MSMA;
Adams et al., 1995; Roberts et al., 1998) approaches have
proved useful in two recent, but non-Mediterranean, change
detection applications (Adams et al., 1995; Roberts et al.,
1998). Because the application of MSMA to change detection is a relatively new concept in remote sensing, it will be
discussed in more detail in Section 2.
Several change detection studies have shown that interdate changes in vegetation properties are best identified
when image data are enhanced using vegetation indices
prior to image differencing (Coppin & Bauer, 1996; Fung,
1990; Mas, 1999; Radeloff et al., 2000). In addition,
changes in landcover are most often associated with a
combination of indices rather than any single index or
change feature (Cohen & Fiorella, 1998).
Levien et al. (1999) have established an operational
forest change mapping program in California employing
the Multitemporal Kauth Thomas (MKT) transform (Collins
& Woodcock, 1996), resulting in high decision tree (DT)
classification accuracy (89%). The MKT approach, however, was developed to detect forest canopy change due to
insect damage and drought but is being applied to all types
of landcover change in this statewide program. The objective of this research, therefore, was to compare the performance of two change enhancement techniques (i.e., MKT and
MSMA) and two classification techniques [i.e., maximum
likelihood (ML) and DT], in terms of classification accuracy, for identifying changes in vegetation cover in a
southern California study site.
2. Methods
2.1. Study site
The study area chosen for this project was the Descanso
Ranger District (DRD) in Cleveland National Forest, San
Diego County (Fig. 1). The 1142-km2 DRD is located 8 km
from the US –Mexico Border, and elevations range from
155 to 1931 m at Mount Laguna. Mean annual precipitation
is low (600 mm) and is correlated with elevation, with
moderate temperatures, varying seasonally between hot, dry
summers and cool, wet winters typical of MTE climates.
Climate, edaphic, and topographic factors combine to produce a complex mosaic of diverse vegetation including
chaparral, woodlands, montane conifer forest, wetlands,
coastal sage scrub, and grasslands. In addition, DRD supports over 90 threatened and endangered species of plants
and animals (Stephenson & Calcarone, 1999).
Fire is the leading ecological disturbance agent in DRD,
and fire frequency and location have shifted from upper
slopes to surrounding lowlands (shrublands) along the
urban –suburban interface over the last 100 years (Stephenson & Calcarone, 1999). Reduction of ground fires in
montane conifer habitats has resulted in a proliferation of
shade-tolerant understory trees (Minnich et al., 1995; Stephenson & Calcarone, 1999), potentially affecting regional
hydrology, future fire patterns, and insect/disease outbreaks.
Further, nonmetropolitan population growth in the San
Diego region has been rapid in the last 10 years (i.e.,
approximately 2% per annum). As a result, the DRD, which
contains a high proportion of private land holdings within its
administrative forest boundary (i.e., one-third; Fig. 1), is
subject to rapid changes in land use and landcover, especially in middle elevation oak woodlands. Timber harvest-
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
145
Fig. 1. Location of the study site (DRD, southern California), showing the distribution of private vs. federal land within the perimeter of the DRD.
ing has not occurred on any significant scale in more than
100 years because of the slow growth and limited extent of
conifer forest in this mountain range.
types present in DRD [e.g., photosynthetic and nonphotosynthetic vegetation (NPV) spectra and bare soil spectra].
2.2. Satellite data
2.3. Vegetation change categories and landcover
data collection
A 24 June 1990 Landsat Thematic Mapper scene (path
40, row 37) and an 8 June 1996 scene were acquired for
change detection analysis. These near-anniversary image
dates were used in the operational change mapping program
(Levien et al., 1999) to ensure a minimum number of
exogenous effects (i.e., low cloud cover, low soil moisture
content, and minimal sun angle effects—both dates had
solar elevations of 60 and azimuths of 101). An outline of
the data processing flow is shown in Fig. 2.
Geometric correction of the image pair involved approximately 50 ground control points, producing an average root
mean square error of 12.6 m (less than 0.5 pixels). Following nearest neighbor resampling, to preserve original pixel
values, the image pair was subjected to an absolute atmospheric correction. Digital numbers were converted to
radiance and input to the 6S software package (Tanre et
al., 1990), which models the atmospheric parameters necessary for conversion of radiance to reflectance. These parameters were then applied to the imagery to perform the
correction. Finally, the reflectance values for both dates
were rescaled from 0 to 255 (i.e., 8-bit ground-leaving
reflectance). This calibration step was important because
image pixel values must be spectrally compatible with fieldmeasured reflectance spectra for linear unmixing to occur.
The 500 field reflectance spectra used in this study were
collected in another region of southern California (Roberts,
Dennison, Ustin, Reith, & Morais, 1999). This spectral
library represented a diverse range of natural landcover
Five vegetation change categories were examined in this
study: (1) no change (static spectral/spatial response over
time), (2) vegetation increase (natural vegetation growth,
most rapid when following a recent disturbance), (3)
vegetation decrease (destructive fire scarring, vegetation
clearing, or insect drought mortality between image dates),
(4) change in nonvegetated regions (clearing and development within urban and suburban areas), and (5) recharge
or flooding in reservoirs and montane lakes. The five
change classes are presented in terms of their hypothesized
spectral change characteristics (Table 1). This landcover
change classification scheme was adapted from the operational USDA Forest Service, California Department of
Forestry change monitoring program (Levien et al., 1999)
and represents an intermediate step between identifying
landcover change related to spectral change and defining
the specific nature (‘‘from-to’’) and cause of that change.
The DRD was initially stratified into approximate change
vs. unchanged areas (by subtracting a 1990 NDVI image
from a 1996 image and setting a threshold for ‘‘changed’’ at
1.5 standard deviations from the mean) and vegetation
lifeform categories (i.e., conifer, chaparral, hardwood, scrub,
and nonforest) based on an existing vegetation map. Following this, a stratified adaptive cluster sampling (SACS)
protocol (Thompson, 1991) was used to identify locations
where field data were collected for training the ML and DT
classifiers and for accuracy assessment of the five change
categories. This ensured that a large sample (i.e., >600 field
146
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Fig. 2. Outline of methods used in this research (input data = light grey boxes, processes = white boxes, and output products = dark grey boxes).
plots) was collected that was representative of both the
desired change categories and the dominant vegetation lifeform types. The SACS approach originated in ecological
applications where populations with patchy or rare distributions require efficient and representative reconnaissance of
disproportionate sample sizes (Thompson & Seber, 1996).
SACS has particular utility for sampling disturbed locations
(changed landcover) because they usually represent a minor
portion of the target population (most of the land area has
not changed) and are often clustered.
Random sampling with equal proportions within these
strata identified the initial sample field plots (circular field
plots had a radius of 30 m). For each observed plot in the
initial sample, adjacent plots in the four cardinal directions
were sampled only if the initial plot contained a change
class of interest. In this way, the sampling effort was
Table 1
Change categories and hypothesized changes in MSMA fractions and MKT features
Hypothesized spectral characteristics and dynamics
Forest change class
MSMA fractions
MKT features
No change (nocng)
Consistent shade, soil, NPV, and GV
fractions over time
Increase in GV and possible increase in
shade in vegetation due to plant
architectural shade
Decrease in soil and NPV over time
due to increased ground cover
Consistent brightness, greenness, and wetness features
over time
Increase in greenness as GV is more abundant
Vegetation increase (veginc)
Vegetation decrease (vegdec)
Change in nonvegetated
regions (novegcng)
Lake recharge (lakerec)
Decrease in GV and plant architectural shade
Increase in soil and NPV due to decreased
vegetation ground cover over time
Decrease in GV due to land clearing
and construction
Increase in shade, soil, and possibly NPV
Decrease in GV and NPV
Increase in shade and possibly soil due to
increased sedimentation
Decrease in wetness due to increased absorption SWIR
wavelengths by vegetation
Decrease in brightness due to increased canopy cover
Decrease in greenness and increase in wetness due to
vegetation removal
Increase in brightness due to soil exposure
Decrease in greenness and wetness due to urbanization
Increase in brightness due to land clearing
Decrease in brightness and greenness due to flooding
Decrease in wetness due to water absorption of
SWIR wavelengths
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
concentrated in local areas of high abundance of the rare
event, i.e., change clusters (Brown & Manly, 1998). This
increased sample sizes for all change classes and reduced
data collection costs (Stehman & Czaplewski, 1998). At
each field plot, the following information was recorded:
UTM easting and northing, dominant lifeform, and observer-interpreted change class and disturbance event/agent
(for example: fire scar, land clearing, flooding etc).
Once field work was completed, the data were randomly
divided to provide 80 points per class for training the image
classifiers ( 400 total) and 30 points per class (150 total)
for accuracy assessment. SACS designs provide an additional advantage to studies utilizing ‘‘data-hungry’’ classification approaches (e.g., DTs and artificial neural networks)
because they facilitate the collection of sufficient quantities
of in situ data for classification training and validation (Smits
et al., 1999). A possible drawback of this approach is the
potential loss of information attributable to intracluster
spatial correlation among ground data (Dobbertin & Biging,
1996; Stehman & Czaplewski, 1998). This approach, however, is considered statistically unbiased with increasing
sample size (Thompson, 1991).
2.4. Image enhancement
In order to produce a suite of change features for change
classification, the 1990 and 1996 images were enhanced
using the MKT and MSMA linear transformations. The
MKT transform develops orthogonal indices based on a
library of soil spectra and assumes no interaction between
subpixel components (Crist & Ciccone, 1984; Kauth &
Thomas, 1976). Linear spectral mixture models, however,
derive mixing lines using pure spectra from the image
(image endmembers) or from a library of field-measured
or laboratory spectra (reference endmembers; Adams,
147
Smith, & Gillespie, 1993) under the assumption of linear
subpixel mixing. These two approaches, in contrast with
other reorganizations of multidimensional feature space,
such as principal components analysis, are based on physical scene materials.
The MKT transform produced three spectral features
representing changes in brightness (DB), greenness (DG),
and wetness (DW) based on TM reflectance factor coefficients developed by Collins and Woodcock (1996) (Fig. 3).
For a complete description of the MKT process, see (Collins
and Woodcock (1996)). Few studies, to date, have used the
MKT in change detection, but they show the MKT produces
change components that are capable of accurately extracting
physical scene characteristics from spectral features (Collins
& Woodcock, 1996; Levien et al., 1999). These pixel-level
approaches to change detection, however, have proved
ineffective in some applications where influence from
background soil and the forest floor confound the spectral
signal of vegetation at subpixel scales (Roberts et al., 1998).
MSMA, therefore, presents an alternative for recovering the
vegetation change signal.
Prior to performing MSMA on both TM images, the
reference library spectra were manipulated to derive candidate reference endmembers for green vegetation (GV), soil,
shade, and non-photosynthetic vegetation (NPV) based on
an endmember optimization technique (Roberts et al., 1998).
This iterative technique compares the relative brightness of
each candidate reference endmember to the brightest candidate image endmembers selected to derive a suitable set of
endmembers to be used in the unmixing process (Fig. 4). For
a full description of this MSMA process, see Roberts et al.
(1998). Because reference endmembers represent the spectra
of known materials, they have the advantage of providing
physically based, standardized measures of fractional abundance. Linear mixing models measure the reflectance of a
Fig. 3. MKT spectral change features (1990 – 1996).
148
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Fig. 4. Optimized reference endmembers for shade, soil, GV, and NPV used to unmix the 1990 and 1996 image pair.
pixel in each spectral band as a linear combination of the
reflectances of its component endmembers, weighted by
their respective surface proportions (Ichoku & Karnieli,
1996). A mixed spectrum, Pil, is modeled as the sum of N
Fig. 5. MSMA change fractions (1990 – 1996).
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
endmembers, Pkl, weighted by the fraction of the endmember, fki, within the field of view at pixel I, i.e.,
Pil ¼
N
X
fki Pkl þ eil
ð1Þ
k¼1
Unmodeled portions of the image spectrum are expressed as
a residual term, eil, at wavelength l (Eq. (1)).
Four reference endmembers [i.e., shade (deep water), GV
(dense oak stand), NPV (senesced grassland), and soil
(bright-bare soil)] were chosen from the optimization procedure (Fig. 2) to represent the fractional cover for the
image pair. Four endmembers were chosen because previous
studies have found the dimensionality of TM data suited to
spectral unmixing using shade, GV, NPV, and soil fractions
(Adams et al., 1995; Roberts et al., 1998). The RMS values
were examined for the image pair and were found to be
within the threshold of error over all areas, with the
exception of recently burned areas (charcoal on the surface
was poorly modeled) and urban areas (a nonnatural landcover type). The 1990 image fractions were then subtracted
from 1996 image fractions, resulting in change fraction
149
images representing change in shade (DShade), GV (DGV),
NPV (DNPV), and soil (DSoil) (Fig. 5).
2.5. Classification and evaluation
The MKT change feature and MSMA change fraction
data sets were classified using supervised ML and univariate
DT algorithms to assign all pixels in the change image data
sets into the five landcover change categories. Most landcover mapping applications have used either parametricsupervised or unsupervised classification algorithms to
identify spectrally distinct groups of data (Smits et al.,
1999). The DT classifier has only recently been used in
landcover mapping studies (Borak & Strahler, 1999; Friedl
& Brodley, 1997; Hansen, Dubayah, & DeFries, 1996;
Hess, Melack, Filsos, & Wang, 1995). It has, however,
seldom been applied in change mapping studies (Roberts
et al., 1998).
The univariate DT classification approach employs treestructured rules that recursively partition the data set feature
spaces into increasingly homogenous classes based on a
splitting criterion. In this approach, possible splits of each
Fig. 6. Univariate DT used to classify the MSMA change fraction data (see Table 1 for definitions of change classes).
150
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Fig. 7. Univariate DT used to classify the MKT change feature data.
independent variable are examined, and the particular split
threshold value of a single variable that produces the largest
deviance measure is chosen to recursively partition the
dependent data (Breiman, Friedman, Olshen, & Stone,
1984; Franklin, 1998). Predictor variables that have already
been used in the tree may be reexamined and potentially
reintroduced into the tree structure. As a result, hierarchical,
nonlinear relationships within the data are derived (Friedl,
Brodley, & Strahler, 1999). Trees were pruned to an
optimum size based on cross-validation using subsets of
the training data. This results in a parsimonious tree model
that does not overfit the training data. The efficiency of tree
classifiers, when compared to ML methods, has been
attributed to their nonparametric nature, in that they do
not require assumptions regarding the distributions of the
input data (Borak & Strahler, 1999). The ML classifier
applies a single classification operator to each observation
and assumes multivariate normally distributed data.
The four resulting landcover change maps (labeled MKTML, MKT-DT, MSMA-ML, and MSMA-DT) were each
assessed for accuracy using contingency matrices. The
kappa (k) statistics, used here, considers all cells in a
contingency matrix, providing a correction for the proportion of chance agreement between the training sites and test
Table 2
Contingency matrix for MKT-ML classification
No change
Vegetation increase
Vegetation decrease
Change in nonvegetated areas
Lake recharge
Total
Reference total
Class total
Correct total
22
25
25
22
56
150
30
30
30
30
30
150
17
14
20
10
24
85
Producer’s
accuracy (%)
User’s
accuracy (%)
77.27
56.00
80.00
45.45
42.86
56.67
46.67
66.67
33.33
80.00
Overall accuracy: 56.67
k
0.4922
0.36
0.6
0.2188
0.6809
0.4583
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
151
Table 3
Contingency matrix for MKT-DT classification
No change
Vegetation increase
Vegetation decrease
Change in nonvegetated areas
Lake recharge
Total
Reference total
Class total
Correct total
37
23
26
21
43
150
30
30
30
30
30
150
22
15
23
14
29
103
Producer’s
accuracy (%)
User’s
accuracy (%)
59.46
65.22
88.46
66.67
67.44
73.33
50.00
76.67
46.67
96.67
k
0.6460
0.4094
0.7177
0.3798
0.9533
0.6083
Overall accuracy: 68.67
data sets (Congalton & Mead, 1983). The k is a standard
statistic to evaluate overall classification accuracy, providing a more conservative estimation than simple percent
agreement value. Other accuracy statistics such as Producer’s Accuracy or omission error (indicating how well training set pixels were classified), User’s Accuracy or
commission error (indicating the probability that a classified
pixel actually represents that category in reality), and
Overall Accuracy (the total number of correctly classified
pixels divided by the total number of reference pixels) were
also evaluated. In addition, the spatial accuracy and extent
of several of the mapped change categories (i.e., no change,
vegetation increase, and vegetation decrease) was examined
using the fire perimeter database.
3. Results
Figs. 6 and 7 show the pruned DTs used to classify the
MSMA change feature data and the MKT change fraction
data, respectively. Change classes with little variability, such
as vegetation decrease and lake recharge, have few terminal
nodes, whereas vegetation increase has several terminal
nodes as a result of greater intraclass variability. These
classification rules make intuitive sense and allow for useful
interpretation in studying the relationships among change
classes and the spectra change features and fractions.
most accurately identified change category followed by
vegetation decrease and no change. The least accurate class
is urban change. The remaining individual class accuracies
are considerably lower as a result of this poor characterization. The DT classification provides for a substantially higher
overall accuracy (69%) and individual class accuracies over
ML (Table 3). The individual class accuracy trends are
similar to the MKT-ML results, with lake recharge having
the highest and urban change having the lowest accuracy.
3.2. MSMA-ML and MSMA-DT contingency matrices
The overall classification accuracy of the MSMA-ML
classification is 65%, with an overall k value of 0.57
(Table 4). The least accurate class is urban change (it
exhibited high commission error). Vegetation increase also
exhibits high commission error (44%) and was most frequently mislabeled as no change in the reference data set. In
general, those change classes that could be considered to
exhibit high intraclass variability are least accurately mapped.
The overall classification accuracy for the MSMA-DT
classification is 72%, which is 7% higher than that of
MSMA-ML, and has a 7% higher k value (Table 5). Individual class accuracies are also all higher, with the exception
of lake recharge, where k values are equal (82%). Individual
class accuracy trends follow those of MSMA-ML, with the
least accurate class being urban change.
3.3. Comparison of MSMA and MKT change
enhancement techniques
3.1. MKT-ML and MKT-DT contingency matrices
The overall accuracy of the MKT-ML classification is low
(57%), with a comparably low k value of 0.46 (Table 2).
Individual class accuracies reveal that lake recharge is the
While the DT classification technique produced higher
overall and individual class accuracies, compared to ML,
Table 4
Contingency matrix for MSMA-ML classification
No change
Vegetation increase
Vegetation decrease
Change in nonvegetated areas
Lake recharge
Total
Reference total
Class total
Correct total
19
44
24
26
37
150
30
30
30
30
30
150
20
17
21
14
26
98
Producer’s
accuracy (%)
User’s
accuracy (%)
45.45
70.83
80.77
73.68
70.27
66.67
56.67
70.00
46.67
86.67
Percent overall accuracy: 65
k
0.52
0.48
0.63
0.38
0.82
0.5667
152
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Table 5
Contingency matrix for MSMA-DT classification
No change
Vegetation increase
Vegetation decrease
Change in nonvegetated areas
Lake recharge
Total
Reference total
Class total
Correct total
44
27
25
21
33
150
30
30
30
30
30
150
23
20
23
16
26
108
Producer’s
accuracy (%)
User’s
accuracy (%)
52.27
74.00
92.00
76.00
78.79
76.67
66.00
76.67
53.00
86.87
k
0.66
0.59
0.72
0.45
0.82
0.65
Overall accuracy: 72
there were also systematic differences between the two
change enhancement techniques. MSMA produced a 5%
higher classification accuracy than MKT. Interestingly, with
regard to the best-classified landcover change class, lake
recharge, MKT is 13% more accurate at mapping flooded
areas than MSMA, suggesting that changes in scene brightness and wetness are more indicative of flooding than
changes in shade (the shade fraction is typically indicative
of the presence of water bodies). Several class accuracy
results were almost equal (i.e., no change and vegetation
decrease). Notable exceptions to this general agreement can
be seen in the change categories that are assumed to exhibit
high intraclass spectral variability such as vegetation
increase and urban change. MSMA was 19% more accurate
in mapping vegetation increase, despite the high commission
errors in this case. MSMA was 8% more accurate than MKT
in mapping urban change, a highly variable class.
3.4. Examination of MSMA-DT change using fire
perimeter data
Because the MSMA-DT approach yielded the most
accurate vegetation change map, it was used to examine
whether all of the vegetation change related to wildfire
between 1990 and 1996 was accurately mapped. This was
accomplished by overlaying the fire perimeter data set
(minimum mapping unit = 10 ha) on polygons derived from
the MSMA-DT classification, representing vegetation increase and decrease. All of the fire scars that occurred in the
nongrassland areas were colocated with vegetation
decrease pixels, with few exceptions (Fig. 8). Several
small fires, however, had occurred prior to the 1990 image
acquisition and in the mid-1990s but were not revealed in
the final change map. These fires had occurred in areas of
the DRD dominated by scrubland and grassland, where
Fig. 8. Accuracy assessment of vegetation change using a fire perimeter database.
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
postfire vegetation regrowth proceeds rapidly. The 6-year
temporal resolution of this study meant that we could not
map all landscape-scale disturbances in DRD.
3.5. Examining the spatial autocorrelation of commission
and omission errors
Spatial autocorrelation is present in an image data set
when nearby pixel values are more similar than expected
by chance. The high commission errors that occurred in the
vegetation increase class occurred with a ‘‘salt-and-pepper’’
appearance in the shrubland areas of the study site. This
pattern was examined for spatial autocorrelation, using
Moran’s I, to determine whether the pattern was random
or nonrandom. Moran’s I is a weighted product –moment
correlation coefficient, where the weights reflect geographic proximity (Griffith, 1987; Anselin & Getis, 1993).
Values larger than zero indicate positive spatial autocorrelation, while values smaller than zero indicate negative
spatial autocorrelation.
A significant Moran’s I coefficient of 0.18 was calculated, within the shrubland lifeform class, for the vegetation
increase commission errors, indicating significant spatial
correlation. This error analysis reveals that precipitation
effects are clustered but highly localized in the shrubland
lifeform category in the image. Therefore, it may be possible
to correct or reduce precipitation effects on interannual
vegetation greenness in grassland – shrubland areas by stratifying change analysis by lifeform and adjusting the overall
map accuracy results accordingly. In addition, errors of
omission in the vegetation decrease category were investigated and were found to be highly significantly autocorrelated (Moran’s I coefficient = 0.26). Visual inspection of
these errors showed that terrain-induced shading has a large
effect on detecting fire-related vegetation decrease in rugged
terrain. In addition, the high spatial autocorrelation of error
in complex terrain areas reveals a need for terrain correction
in future change detection work.
4. Discussion and conclusions
4.1. Comparison of classification approaches
The DT classification outperformed the ML classification, regardless of the linear change enhancement algorithm. Nonparametric classifiers have frequently been found
to yield higher classification accuracies than parametric
classifiers because of their ability to cope with nonnormal
distributions and intraclass variation found in a variety of
spectral data sets (Carpenter, Gopal, Macomber, Martens, &
Woodcock, 1999; Friedl et al., 1999). A good example of
this can be seen in the DT produced for MKT features
(Fig. 7) where the vegetation increase class was associated
with several terminal nodes (predicted by change in brightness; greenness and wetness). ML classification results
153
might have been improved if unequal prior probabilities
had been assigned for the change map classes (Strahler,
1980). The use of prior probabilities for training data is
considered important in environments where one or two
classes are dominant spatially (Borak & Strahler, 1999).
ML classification requires prior probabilities, while DTs
learn them from the distributions in the training data.
In this study, prior probabilities were not provided for
the ML classification because (1) beyond the knowledge
that a no change class will be the largest category, they
cannot readily be identified in rapidly changing regions
(i.e., use of probabilities may be more suited to postclassification change landcover mapping studies where they can
be more easily computed), and (2) our goal was to compare
automated change detection techniques for landcover
change mapping, where a minimal level of information
about a landscape must be adequately dealt with (Carpenter
et al., 1999).
4.2. Comparison of change enhancement approaches
Irrespective of the classification algorithm used to classify change enhanced imagery, MSMA techniques outperformed those of MKT. Similar to MSMA, MKT orthogonal
transformations are representative of physical scene materials and spatially and temporally robust, as has been demonstrated in previous studies (Collins & Woodcock, 1996;
Levien et al., 1999). However, MKT coefficients are fixed
and cannot be iteratively selected in the same fashion as
MSMA reference endmembers. Despite the claim that NPV
spectra and image fractions are highly correlated to those of
bare soil (Roberts et al., 1997), the addition of this fourth
dimension to represent known scene-based information
appears invaluable to vegetation change studies (i.e., detecting vegetation decrease caused by fire) and provides yet
another advantage over MKT.
The hypothesized relationship between change categories and MSMA fractions and MKT features presented in
Table 1 were, for the most part, correct. The no change
category was strongly associated with consistent MKT
features and MSMA fractions between 1990 and 1996,
but classification accuracies for this class were lower than
expected due to large errors of omission caused by confusion with the vegetation increase category. The most
accurate category, lake recharge, was closely associated
with changes in shade and NPV fractions.
The vegetation decrease class was the second most
accurate change class examined due to the discrete
spectral and spatial nature of vegetation reduction caused
by fires in DRD. Fire scars were related to every
vegetation decrease change image examined in this study.
An important caveat is that not all wildfires were detected
using the 1990 –1996 image dates, as noted previously.
This result raises important questions about temporal
resolution considerations of change detection studies in
fire-affected regions.
154
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
Urban changes caused considerable confusion among
linear transformations that were principally designed for
vegetation mapping studies. This confusion was caused by
strong intraclass spectral variability (i.e., changes in urban
areas can be associated with vegetation clearing or increases
associated with landscaping).
Finally, the vegetation increase class was associated
with increases in greenness and GV and decreases in
brightness and soil. The commission errors of this class
were associated with shrub and grassland areas of DRD
because these vegetation types are highly sensitive to
variations in the precipitation regime. The effectiveness
of change detection is dependent upon the change signal
being greater than the interdate noise. Errors of commission found in the vegetation increase class were caused by
sizeable differences in precipitation between 1990 and
1996 in DRD. The Descanso ranger station, for example,
received 30% more precipitation in 1996 than in 1990.
Interannual differences in vegetation productivity associated with precipitation are difficult to separate from vegetation regrowth following disturbance.
4.3. Examining commission and omission errors
Understanding the key interactions between vegetation
cover and dynamic disturbance agents is critical in identifying desired vegetation cover conditions and regional management priorities. Increasingly, a higher degree of
automation is desired from change enhancement and classification techniques (Carpenter et al., 1999; DeFries, Townshend, & Hansen, 1999; Smits et al., 1999). Further, the
categories of landcover change are many and varied,
depending on the map user’s needs and the study area
involved. High commission errors within the vegetation
increase class were evident due to interannual variation in
precipitation between 1990 and 1996 (Roos, 1992). Some
authors (Cihlar, 2000; Hall & Goetz, 1991) have suggested
that a ‘‘phenological correction’’ could be feasible based on
seasonal trajectories established for similar vegetated
ground targets. Correcting for this type of error remains
an open area in remote sensing research because it has been
shown to be an significant cause of omission error in both
change detection and postclassification change approaches
(Mas, 1999; Waller, 1999).
Finally, misregistration error may have contributed to the
salt-and-pepper effect found in the classified images despite
our efforts to significantly reduce it (Townshend, Justice,
Gurney, & McManus, 1992). Low-pass filters, however, can
reduce the effects of misregistration on change-related
classifications, while preserving regions of continuity in
the image data, and will be examined in future work. In
addition, a postclassification approach may have improved
the overall accuracy of these results (Mas, 1999; Roberts
et al., 1998) but was not considered in this research due to
our emphasis on testing change enhancement and classification algorithms in a map-updating, operational context.
Acknowledgments
The authors wish to acknowledge support from NASA
Grant LCLUC 99-0002-0126 and to thank C. Fischer of
California Department of Forestry and Fire Protection
(CDF), L. Levien of USDA Forest Service-Sacramento, P.
Roffers of Pacific Meridian Resources, and A. Getis, A.
Hope, D. Apalatea, and D. Stow of San Diego State
University for their help in this research and preparation of
this manuscript.
References
Aber, J. D. (1992). Nitrogen cycling and nitrogen saturation in temperate
forest ecosystems. Trends in Ecology and Evolution, 7, 220 – 223.
Adams, J. B., Sabol, D. E., Kapos, V., Almeida-Filho, R., Roberts, D. A.,
Smith, M. O., & Gillespie, A. R. (1995). Classification of multispectral
images based on fractions of endmembers: application to land-cover
change in the Brazilian Amazon. Remote Sensing of Environment, 52,
137 – 154.
Adams, J. B., Smith, M. O., & Gillespie, A. R. (1993). Imaging spectroscopy:
interpretation based on spectral mixture analysis. In: C. M. Pieters, & P.
Englert (Eds.), Remote geochemical analysis: elemental and mineralogical composition ( pp. 145 – 166). New York: Cambridge Univ. Press.
Anselin, L., & Getis, A. (1993). Spatial statistical analysis and geographic
information systems. In: M. M. Fischer, & P. Nijkamp (Eds.), Geographic information systems, spatial modelling, and policy evaluation
( pp. 35 – 49). New York: Springer-Verlag.
Biging, G. S., Colby, D. R., & Congalton, R. G. (1998). Sampling systems
for change detection accuracy assessment. In: R. S. Lunetta, & C. D.
Elvidge (Eds.), Remote sensing change detection: environmental monitoring methods and applications ( pp. 281 – 308). Ann Arbor, MI: Ann
Arbor Press.
Borak, J. S., Lambin, E. F., & Strahler, A. H. (2000). The use of temporal
metrics for landcover change detection at coarse spatial scales. International Journal of Remote Sensing, 21 (6 – 7), 1415 – 1432.
Borak, J. S., & Strahler, A. H. (1999). Feature selection and landcover
classification of a MODIS-like data set for a semiarid environment.
International Journal of Remote Sensing, 20 (5), 919 – 938.
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees (p. 356). Belmont, CA: The Wadsworth
Statistics and Probability Series, Wadsworth International Group.
Brown, J. A., & Manly, B. J. F. (1998). Restricted adaptive cluster sampling. Environmental and Ecological Statistics, 5, 49 – 63.
Carpenter, G. A., Gopal, S., Macomber, S., Martens, S., & Woodcock, C. E.
(1999). A neural network method for efficient vegetation mapping.
Remote Sensing of Environment, 70, 138 – 152.
Choudhury, B. J. (1994). Synergism of multispectral satellite observations
for estimating regional land surface evaporation. Remote Sensing of
Environment, 49, 204 – 214.
Cihlar, J. (2000). Landcover mapping of large areas from satellites: status
and research priorities. International Journal of Remote Sensing, 21
(6 – 7), 1093 – 1114.
Cohen, W. B., & Fiorella, M. (1998). Comparison of methods for detecting
conifer forest change with Thematic Mapper imagery. In: R. S. Lunetta,
& C. D. Elvidge (Eds.), Remote sensing change detection: environmental monitoring methods and applications ( pp. 89 – 102). Ann Arbor,
MI: Ann Arbor Press.
Collins, J. B., & Woodcock, C. E. (1996). An assessment of several
linear change detection techniques for mapping forest mortality using
multitemporal Landsat TM data. Remote Sensing Environment, 56,
66 – 77.
Congalton, R. G., & Mead, R. A. (1983). A quantitative method to test for
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
consistency and correctness in photo interpretation. Photogrammetric
Engineering and Remote Sensing, 49 (1), 69 – 74.
Coppin, P. R., & Bauer, M. E. (1996). Digital change detection in forest
ecosystems with remote sensing imagery. Remote Sensing Reviews, 13,
207 – 234.
Crist, E. P., & Ciccone, R. C. (1984). Application of the tasseled cap
concept to simulated Thematic Mapper data. Photogrammetric Engineering and Remote Sensing, 50, 343 – 352.
DeFries, R. S., Townshend, J. R. G., & Hansen, M. (1999). Continuous
fields of vegetation characteristics at the global scale. Journal of Geophysical Research, 104, 16911 – 16922.
Dobbertin, M., & Biging, G. S. (1996). A simulation study of the effect of
scene autocorrelation, training sample size and sampling method on
classification accuracy. Canadian Journal of Remote Sensing, 22 (4),
360 – 367.
Franklin, J. (1998). Predicting the distribution of shrub species in southern
California from climate and terrain-derived variables. Journal of Vegetation Science, 9, 733 – 748.
Franklin, J., & Woodcock, C. E. (1997). Multiscale vegetation data for the
mountains of southern California: spatial and categorical resolution. In:
D. A. Quattrochi, & M. F. Goodchild (Eds.), Scale in remote sensing
and GIS ( pp. 141 – 168). Boca Raton, FL: CRC/Lewis Publishers.
Franklin, J., Woodcock, C. E., & Warbington, R. (2000). Multi-attribute
vegetation maps of Forest Service lands in California supporting resource management decisions. Photogrammetric Engineering and Remote Sensing, 6, 1209 – 1217.
Friedl, M. A., & Brodley, C. E. (1997). Decision tree classification of
landcover from remotely sensed data. Remote Sensing of Environment
61, 399 – 408.
Friedl, M. A., Brodley, C. E., & Strahler, A. H. (1999). Maximizing landcover classification accuracies produced by decision trees at continental
to global scales. IEEE Transactions on Geoscience and Remote Sensing, 37 (2), 969 – 977.
Fung, T. (1990). An assessment of TM imagery for land-cover change
detection. IEEE Transactions in Geoscience and Remote Sensing, 28
(4), 681 – 684.
Griffith, D. (1987). Spatial autocorrelation: a primer. Washington, DC:
Association of American Geographers Monograph.
Gunilla, E., Olsson, A., Austerheim, G., & Grenne, S. N. (2000). Landscape
change patterns in mountains, land use and environmental diversity,
mid-Norway 1960 – 1993. Landscape Ecology, 15, 155 – 170.
Hall, F. G., & Goetz, S. J. (1991). Large-scale patterns of forest succession
as determined by remote sensing. Ecology, 72 (2), 628 – 640.
Hansen, M., Dubayah, R., & DeFries, R. (1996). Classification trees: an
alternative to traditional landcover classifiers. International Journal of
Remote Sensing, 17 (5), 1075 – 1081.
Hess, L. L., Melack, J. M., Filsos, S., & Wang, Y. (1995). Delineation of
inundated area and vegetation along the Amazon floodplain with the
SIR-C synthetic aperture radar. IEEE Transactions on Geoscience and
Remote Sensing, 33 (4), 896 – 904.
Hill, J., Megier, J., & Mehl, W. (1995). Land degradation, soil erosion, and
desertification monitoring in Mediterranean ecosystems. Remote Sensing Reviews, 12, 107 – 130.
Hope, A. S. (1995). Monitoring ecosystem response to global change: high
spatial resolution digital imagery. In: J. M. Moreno, & W. C. Oechel
(Eds.), Global change and Mediterranean type ecosystems ( pp. 239 –
254). Heidelberg: Springer-Verlag.
Ichoku, I., & Karnieli, A. (1996). A review of mixture modeling techniques for sub-pixel landcover estimation. Remote Sensing Reviews, 13,
161 – 186.
Kauth, R. J., & Thomas, G. S. (1976). The tasseled cap—a graphic description of the spectral-temporal development of agricultural crops as seen
by Landsat. In: Proceedings of Remotely Sensed Data ( pp. 41 – 57).
West Lafayette, IN: Purdue University.
Keeley, J. E., Fotheringham, C. J., & Morais, M. (1999). Reexamining
fire suppression impacts on brushland fire regimes. Science, 284,
1829 – 1832.
155
Lambin, E. F. (1998). Modeling and monitoring land-cover change processes in tropical Regions. Progress in Physical Geography, 21 (3),
375 – 393.
Levien, L. M., Fischer, C. S., Roffers, P. D., Maurizi, B. A., Suero, J.,
Fischer, C., & Huang, X. (1999). A machine-learning approach to
change detection using multi-scale imagery. In: Proceedings of ASPRS
Annual Conference, (1,22, Portland, OR). Bethesda, MD: American
Society for Photogammetry and Remote Sensing (ASPRS).
Macleod, R. D., & Congalton, R. G. (1998). A quantitative comparison of
change-detection algorithms for monitoring eelgrass from remotely
sensed data. Photogrammetric Engineering and Remote Sensing, 64
(3), 207 – 216.
Mas, J. F. (1999). Monitoring land-cover changes: a comparison of change
detection techniques. International Journal of Remote Sensing, 20 (1),
139 – 152.
Minnich, R. A., Barbour, M. G., Burk, J. H., & Fernau, R. F. (1995). Sixty
years of change in Californian conifer forests of the San Bernardino
Mountains. Conservation Biology, 9 (4), 902 – 914.
Moreno, J. M., & Oechel, W. C. (Eds.). (1995). Global change and Mediterranean-type ecosystems (p. 527). New York: Springer-Verlag.
Nemani, R., & Running, S. W. (1996). Global vegetation cover changes
from coarse resolution satellite data. Journal of Geophysical Research,
101, 7157 – 7162.
O’Leary, J. F., & Westman, W. E. (1988). Regional disturbance effects on
herb succession patterns in Coastal Sage Scrub. Journal of Biogeography, 15, 775 – 782.
Radeloff, V. C., Mladenoff, D. J., & Boyce, M. S. (2000). Effects of
interacting disturbances on landscape patterns: budworm defoliation
and salvage logging. Ecological Applications, 10 (1), 233 – 247.
Roberts, D. A., Batista, G. T., Pereira, J., Waller, E. K., & Nelson, B. W.
(1998). Change identification using multitemporal spectral mixture
analysis: applications in Eastern Amazonia. In: C. Elvidge, & R. Lunetta (Eds.), Remote sensing change detection: environmental monitoring applications and methods ( pp. 137 – 161). Ann Arbor, MI: Ann
Arbor Press.
Roberts, D. A., Dennison, P., Ustin, S. L., Reith, E., & Morais, M. (1999).
Development of a regionally specific library for the Santa Monica
Mountains using high resolution AVIRIS data. In: Proceedings 8th
AVIRIS Earth Science Workshop (p. 6). Pasadena, CA: JPL.
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, 223 – 240.
Roos, M. (1992). The hydrology of the 1987 – 1992 California drought.
Technical Paper, State of California Resources Agency, Department
of Water Resources, Division of Flood Management, p. 25.
Running, S. W., & Gower, S. T. (1991). FOREST-BGC, A general model of
forest ecosystem processes for regional applications: II. Dynamic carbon allocation and nitrogen budgets. Tree Physiology, 9, 147 – 160.
Running, S. W., & Nemani, R. R. (1991). Regional hydrologic and carbon
balance responses of forests resulting from potential climate change.
Climatic Change, 19, 349 – 368.
Sala, O. E., Chaplain, F. S. III, Armesto, J. J., Berlow, E., Bloomfield, J.,
Dirzo, R., Huber-Sanwald, E., Huenneke, L. F., Jackson, R. B., Kinzig,
A., Leemans, R., Lodge, D. M., Mooney, H. A., Oesterheld, M., Poff,
N. L., Sykes, M. T., Walker, B. H., Walker, M., & Wall, D. H.
(2000). Global biodiversity scenarios for the year 2100. Science,
287, 1770 – 1774.
Shoshany, M. (2000). Satellite remote sensing of natural Mediterranean
vegetation: a review within an ecological context. Progress in Physical
Geography, 24 (2), 153 – 178.
Singh, A. (1989). Digital change detection techniques using remotely sensed
data. International Journal of Remote Sensing, 10 (6), 989 – 1003.
Skole, D. L., & Tucker, C. (1993). Tropical deforestation and habitat fragmentation in the Amazon: satellite data from 1978 and 1988. Science,
260, 1905 – 1910.
Smits, P. C., Dellepiane, S. G., & Schowengerdt, R. A. (1999). Quality
assessment of image classification algorithms for land-cover mapping: a
156
J. Rogan et al. / Remote Sensing of Environment 80 (2002) 143–156
review and a proposal for a cost-based approach. International Journal
of Remote Sensing, 20 (8), 1461 – 1486.
Stehman, S. V., & Czaplewski, R. L. (1998). Design and analysis for
thematic map accuracy assessment: fundamental principles. Remote
Sensing of Environment, 64, 331 – 344.
Stephenson, J. R., & Calcarone, G. M. (1999). Southern California mountains and foothills assessment: habitat and species conservation issues.
General Technical Report GTR-PSW-172. Albany, CA: Pacific Southwest Research Station, Forest Service, U.S. Department of Agriculture, 402 pp.
Stow, D. (1995). Monitoring ecosystem response to global change: multitemporal remote sensing analyses. In: J. Moreno, & W. Oechel (Eds.),
Anticipated effects of a changing global environment in Mediterranean
type ecosystems ( pp. 254 – 286). New York: Springer-Verlag.
Strahler, A. H. (1980). The use of prior probabilities in maximum likelihood classification of remotely sensed data. Remote Sensing of Environment, 10, 135 – 163.
Sujatha, G., Dwivedi, R. S., Sreenivas, K., & Venkataratnam, L. (2000).
Mapping and monitoring of degraded lands in part of Jaunpur district of
Uttar Pradesh using temporal spaceborne multispectral data. International Journal of Remote Sensing, 21 (3), 519 – 531.
Tanre, D., Deroo, C., Duhaut, P., Herman, M., Morcrette, J. J., Perbos, J., &
Deschamps, P. Y. (1990). Description of a computer code to simulate
the satellite signal in the solar spectrum: the 6S code. International
Journal of Remote Sensing, 11, 659 – 668.
Thompson, S. K. (1991). Stratified adaptive cluster sampling. Biometrika,
78, 389 – 397.
Thompson, S. K., & Seber, G. A. F. (1996). Adaptive sampling (p. 265).
New York: Wiley.
Townshend, J. R. G., Justice, C. O., Gurney, C., & McManus, J. (1992).
The effect of image misregistration on the detection of vegetation
change. IEEE Transactions in Geoscience and Remote Sensing, 30,
1054 – 1060.
Verstraete, M. M., & Schwartz, S. A. (1991). Desertification and global
change. Vegetation, 91, 3 – 13.
Waller, E. K. (1999). A comparison of Landsat Multispectral Scanner imagery and aerial photography for identifying landcover change in coastal southern California. Master’s Thesis, University of California, Santa
Barbara, p. 131.
Zedler, P. H., Gautier, C. R., & McMaster, G. S. (1983). Vegetation change
in response to extreme events: the effect of short interval between fires
in California chaparral and coastal scrub. Ecology, 64 (4), 809 – 818.