Environmental Monitoring and Assessment (****) ***: 65 83
c Springer 2006
DOI: 10.1007/s10661-006-1071-z
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA USING
AERIAL PHOTOGRAPHS ALONG THE CACHE CREEK, CALIFORNIA
SHAOKUI GE1,2,, RAYMOND CARRUTHERS2, PENG GONG1
and ANGELICA HERRERA1,2
1
Department of Environmental Science, Policy, and Management, University of California,
Berkeley, California, USA; 2 Exotic and Invasive Weeds Research Unit, Western Regional Research
Center USDA, Agricultural Research Service, 800 Buchanan Street Albany,
California, USA
( author for correspondence, e-mail: abpyy9@r.postjobfree.com)
(Received 17 September 2004; accepted 24 January 2005)
Abstract. Natural color photographs were used to detect the coverage of saltcedar, Tamarix parvi ora,
along a 40 km portion of Cache Creek near Woodland, California. Historical aerial photographs from
2001 were retrospectively evaluated and compared with actual ground-based information to assess
accuracy of the assessment process. The color aerial photos were sequentially digitized, georeferenced,
classi ed using color and texture methods, and mosaiced into maps for eld use. Eight types of ground
cover (Tamarix, agricultural crops, roads, rocks, water bodies, evergreen trees, non-evergreen trees
and shrubs (excluding Tamarix)) were selected from the digitized photos for separability analysis
and supervised classi cation. Due to color similarities among the eight cover types, the average
separability, based originally only on color, was very low. The separability was improved signi cantly
through the inclusion of texture analysis. Six types of texture measures with various window sizes
were evaluated. The best texture was used as an additional feature along with the color, for identifying
Tamarix. A total of 29 color photographs were processed to detect Tamarix infestations using a
combination of the original digital images and optimal texture features. It was found that the saltcedar
covered a total of 3.96 km2 (396 hectares) within the study area. For the accuracy assessment, 95
classi ed samples from the resulting map were checked in the eld with a global position system (GPS)
unit to verify Tamarix presence. The producer s accuracy was 77.89%. In addition, 157 independently
located ground sites containing saltcedar were compared with the classi ed maps, producing a user s
accuracy of 71.33%.
Keywords: invasive species, saltcedar, tamarix parvi ora, monitoring, detection, aerial photograph,
remote sensing, California, texture analysis
1. Introduction
Biological invasions have increasingly become an important topic for environmental
scientists and natural resource managers, worldwide. Invasive species alter commu-
nity composition and ecosystem function, and cause severe threats to biodiversity
in the invaded areas. In order to understand the invasion mechanism and better con-
trol invasive plants and their spread, it is helpful to characterize the abundance and
distribution of an invasive species in areas of concern (Mooney, 1999). Although
ground-based investigation can accurately measure invasive species in a limited
66 SHAOKUI GE ET AL.
area, it is a time-consuming, labor-intensive work. In particular, ground-based as-
sessment is not always practical for inaccessible areas where invasive species often
exist. Thus, remote sensing can be an important tool to detect and characterize
invasive species over large areas (Everitt et al., 1995, 1996; Lass et al., 1996).
Saltcedars, Tamarix spp., were introduced to the United States in the 1800 s for
a variety of reasons including soil and wind erosion control. This shrub escaped
from cultivation and now poses a signi cant threat to the natural ecosystems in the
western USA (Brock, 1994; Di Tomaso, 1998; Bailey et al., 2001). It grows most
successfully along riparian zones, where it out-competes native vegetation.
This invasive shrub has spread widely along Cache Creek and other riparian
areas in Northern California (Figure 1). Its roots extend deeply into the soil where
it depletes surface and underground water. In addition to its extensive water use,
its invasion causes several other negative impacts such as salinization of soils, high
stream bank erosion, increased ooding and high re hazard to localized habi-
tats. Multiple stems and slender branches characterize the general morphology of
saltcedar, which may grow over 10 meters in height (Figure 2). It out-competes
and replaces native vegetation, resulting in high-density monospeci c-stands of
saltcedar along riparian corridors. In some areas, water consumption and salt
Figure 1. Study area (Cache Creek, California, USA, after California NRCS, (A) down part; (B)
middle part; and (C) up part for this study).
67
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
Figure 2. Showing Saltcedars patches.
buildup have highly altered native ecosystems to the point that few other species can
survive. It is therefore critical that effective techniques to limit saltcedar growth and
control its spread be developed and implemented (Di Tomaso, 1998; Cleverly et al.,
1997; Zavaleta, 2000; Bailey et al., 2001). Such attempts closely depend upon ac-
curate detection of this plant on a regional scale, thus control actions can be applied
widely throughout entire watersheds. However, except for some preliminary tests
using digital videography, no research has been reported on the use of aerial photog-
raphy in monitoring saltcedar distributions across wide areas or entire watersheds.
It was demonstrated that high spatial resolution remotely sensed data was a valu-
able data source to detect invasive species infestations across wide areas (Gausman
et al., 1977; Everitt et al 1995; Lass et al., 1996). Wide area assessment is of-
ten important because invasive species are commonly distributed in an aggregated
fashion providing a patchwork pattern in speci c habitats within an invaded area.
In particular, for invasive and noxious plant weed, small-patches of infested habitat
are very important and cannot be ignored, since they often are newly infected areas
under high-intensity control or eradication. Existing studies of invasive species have
primarily been based on the re ectance difference of targets with the expectation
that vegetation types could be distinguished from surrounding neighboring plants
based on color or color infrared photographs (CIR). For example, having compared
the application of color photos with CIR aerial photography, Everitt and Deloach
(1990) found that the foliage of Tamarix ramosissima turned golden brown during
the late fall/ early winter period so that its re ectance curve was much different
from those of the associated native vegetation and soil background. Both types of
photography (color and CIR) were used to distinguish Tamarix ramosissima from
associated native vegetation. Color photography, however, is more widely avail-
able than CIR using both lm and digital technology. High resolution commercial
satellite images are usually taken along polar orbits with a narrow footprint. It takes
multiple overpasses of those satellites to acquire images over a river system that
could take many different directions in their meanderings. Therefore, the cost of
68 SHAOKUI GE ET AL.
data acquisition is high. In consideration of these factors, color aerial photogra-
phy was thought to be the most cost-effective remote sensing tool for the study of
saltcedar in our riparian environment and several sets of historical photographs of
Cache Creek were available.
In this study, color aerial photography was used to map invasive Tamarix parv-
i ora along an east running portion of Cache Creek, near Woodland, California
(see Figure 1). Traditionally, for speci c feature extraction and delineation, manual
photo interpretation is the most common method used by landmanagers interested
in invasive plant control. However, manual interpretation is both time-consuming
and labor-intensive, and it is almost impossible to recognize small patches of vege-
tation over wide areas or at a watershed scale of interest. Therefore, we developed
an image based classi cation procedure to extract Tamarix covers for large areas
along riparian corridors. Here, not only color but also image texture was used to
detect these invasive plants. Texture analysis was considered important because
color alone was not suf cient to separate Tamarix cover from adjacent bene cial
plant species. In general, texture refers to the tonal variation in an image; and thus
texture features help enhance separability among classes (Anys et al., 1995; Mather
et al., 1998; Smits et al., 1999; Jonathan et al., 2001).
Texture extraction algorithms have been developed from remotely sensed data
for various spatial resolutions (Haralick et al., 1973; Haralick, 1979; Gong et al.,
1992; Hassan and He 1995; Riou et al., 1997; Saatchi et al., 2000, Nyoungui
et al., 2002; Xu et al., 2003). All these studies demonstrated that texture in a
local pixel neighborhood can be used to improve vegetation discrimination in high-
resolution images (Hudak et al., 1998; Podest and Saatchi, 2002). In this paper
we report our analysis methods and experimental results. In addition, we evaluate
some commonly used texture measures for improving land cover classi cation from
aerial color photographs and assess the accuracy of extracted Tamarix parviflora
distributions through ground-based veri cation.
2. Materials and Methods
Historical aerial photographs were used retrospectively to assess saltcedar infes-
tations along an important waterway in Northern California and then linked back
to ground based validation sampling conducted following this analysis. The pho-
tographs were actually taken for assessing gravel distribution along Cache Creek
but were timed such that they correspond with peak bloom and thus allowed an
assessment of the saltcedar infestation in this area. A total of 29 color photos were
taken along Cache Creek in April 2001 using a Zeiss RMK TOP 15 camera with
a Zeiss Pleogon A3/4 lens. The camera was physically mounted in a twin-engine
aerial platform that was piloted by American Aerial Survey, Inc. The scale of the
photos was 1:12,000. At the time of photography, the saltcedar was not yet leafed out
and only owers were present on these plants. The owers were pink in color, mak-
69
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
ing them distinctive from other associated riparian vegetation and other background
materials in the study area. Because the photographs were taken in response to ow-
ering, it was possible to mix saltcedar with some blooming fruit trees in adjacent
agricultural orchards and thus careful discrimination between riparian zones and
agricultural elds was needed to ensure that only saltcedar was extracted. Although
it was possible to scan the photos in great detail, we chose to scanned the photos to
1 meter resolution. The scanned photos were georeferenced to 1-meter resolution
Digital Ortho Quarter Quads (DOQQ) from USGS with a second order polynomial
function. After classi cation, they were mosaiced together to provide a continuous
photographic map. Based on the color similarity and relationships among various
types of plants in the area, vegetation was initially divided into eight categories:
saltcedar, evergreen trees, non-evergreen trees, shrubs, crops, bare elds and hills
(including agriculture and rangeland), water bodies (including wetlands), rocks and
roads. The procedure used for mapping the saltcedar cover is outlined in Figure 3.
The image processing used to distinguish the tamarix from associated vegetation
and other background substrates required four distinct steps: (1) Texture extraction
using six algorithms; (2) Texture subset selection based on separability measured
by Bhattacharya distances (BD, Shaban and Dikshit, 2001); (3) Best texture deter-
mination based on classi cation accuracy; and (4) A post-classi cation processing
procedure, were used as the basis of invasive habitat determination. Following this
classi cation procedure, the original eight categories were reduced to six cate-
gories, by merging rocks and roads with bare elds and hills. Individual classi ed
photos were then mosaiced into three discrete sections of a 40km long portion of
Cache Creek: the downstream, middle and up-stream segments of the riparian cor-
ridor. Finally, using completely random samples 100 points were extracted from
these classi ed maps and used to conduct overall accuracy assessments, including
calculating producer s and user s accuracies.
2.1. C OMPARING COLOR SEPARABILITY
The original image was scanned in three layers (blue, green and red) using 8 bits per
layer. Then, twenty 1-meter pixels of each cover type were randomly sampled from
these digital images, and then was replicated three times. The color differences
among land covers were determined using analysis of variance, and a Tukey s
multiple range test was used to examine the statistical signi cance between means
at a 0.95 probability level.
2.2. TEXTURE MEASUREMENTS
Texture measurements were derived from the digitally processed photographs based
on a gray level co-occurrence matrix (GLCM) (Gong et al., 1992; Jensen, 1996).
GLCM is usually a probability matrix whose elements are indexed by gray-level
values of any pixel-pair (a, b) at a xed distance and angle in a pixel neighborhood.
70 SHAOKUI GE ET AL.
Scan photo with 1000 dpi to form 3-layer-images of 1-foot pixel size
Resampling scanned photos to pixel sizes of 1.0 meter
Try different texture algorithms with 3*3,5*5,7*7 and 9*9 windows.
Overall and tamarix
separabilities
Determine texture for inclusion in supervised classification by MLC
Based on overall and tamarix accuracy to determine the best texture
Aggregate into six categories including trees, shrubs, bare fields, tamarix,
crops, and water
Based on habitat info to calibrate and validate classified results manually
interpretation.
Mosaicing and outputting map of invasion cover for Tamarix distribution
Figure 3. Procedures for the tamarix detection.
Given N as the number of gray levels and Pa,b (i, j) as the joint probability of pixel
a with gray level i and pixel b with gray level j. For each pixel neighborhood of
a certain size, a GLCM was constructed with its ith row and jth column having
probability Pa,b (i, j). The relative spatial relationship between a and b is xed for
each GLCM. For simplicity, we will use P(i, j) instead of Pa,b (i, j). The following
71
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
functions were used to develop texture algorithms:
p (i, j )
Homogeneity: Homo = (1)
1 + (i j )2
Contrast: Con = (i j )2 p (i, j ) (2)
Dissimilarity: Dis = p (i, j ) abs (i j ) (3)
Mean: Mean = i p (i, j ) (4)
2
i
Standard deviation: Sdev = p (i, j ) i (5)
n
Entropy: Ent = p (i, j ) log p (i, j ) (6)
2.3. B ETWEEN -CLASS SEPARABILITY WITH DIFFERENT WINDOW SIZES
Bhattacharya distance (BD) was applied to represent the separability between class
pairs of training samples. BD values range between 0.0 and 2.00. If it is less than
1.00, the two cover classes are very similar and thus hard to separate. When this is
the case, the classes should be merged or one of the cover classes discarded. BD
values between 1.00 and 1.9 indicate that the two classes could be separated at least
to some extent. BD values between 1.90 and 2.00 indicate very good separation
between the two classes in question. In this study, texture measures were calcu-
lated and compared at different window sizes to determine the optimal window
size for texture selection to distinguish invasive saltcedar from other surface cov-
ers. Comparisons were based on maximum, minimum and average separabilities of
training samples with different texture measures obtained in window sizes of 3 * 3,
5 * 5, 7 * 7, and 9 * 9 pixels. A texture set was determined to be appropriate for
classi cation when the separability was greater than 1.90. Speci cally, the separa-
bility between Tamarix and other classes were given speci c attention in order to
determine whether it was separable from all other cover types. Based on the over-
all separability and Tamarix separability with each non-tamarix class, the optimal
window size and texture feature were chosen for mapping tamarix the distribution
along Cache Creek. This texture assessment was combined with the original image
layers to jointly classify different pixel-based patterns to extract the invasive cover
areas. Due to variation in light condition and topographical changes within our
large study area, texture could be different from location to location, even for the
same vegetation type. Therefore, in order to obtain accurate classi cation results,
different textures were assessed and used across individual photos. As an example,
a photo taken from the middle reach of the creek was used to illustrate such image
processing.
72 SHAOKUI GE ET AL.
2.4. S ELECTION OF THE OPTIMAL TEXTURE ALGORITHM
Various texture features selected as described previously, were added as additional
channels to the original image classi cation process. The maximum likelihood
classi er (MLC) was then used in the nal classi cation, and class signatures
were derived from images based on pixel color values and the selected texture
determined by training samples. Although this study focused primarily on tamarix
covers, the overall classi cation accuracy was still considered. Therefore, through
the balanced assessment of the overall classi cation accuracy and the classi cation
accuracy of tamarix, the optimal texture feature was selected for inclusion in the
nal classi cation.
2.5. H ABITAT - BASED CALIBRATION AND ACCURACY ASSESSMENT
Due to color similarities among different cover types, non-tamarix habitat areas
such as orchards had the potential to be misclassi ed as saltcedar. However, Tamarix
primarily grows only along the riparian corridor adjacent to the ood plane. Thus,
large rectangular orchard areas were masked out from image classi cation in order
to improve the classi cation process. After orchard exclusion, accuracy assessment
was analyzed using independent eld validation data. For producer s accuracy,
random samples were selected from the classi ed tamarix cover class. A differential
eld GPS unit was then used to locate speci c areas where pixels were identi ed
as saltcedars by the proposed classi ers. A eld crew located appropriate points
and made on-site determination as to whether saltcedar was present or absent at
each site. As the majority of the saltcedar in this location are medium to large
shrubby trees comprised of several square meters of cover area, the submeter GPS
accuracy was considered adequate for this assessment along with aerial photograph-
based maps that allowed individual tree recognition in the eld. For user s accuracy
assessment, we randomly located tamarix patches in the eld. The coordinates of
these locations were then measured and the sample sites were then compared with
the classi ed tamarix covers back in the laboratory where we veri ed how many
samples were found on the classi cation map to calculate user s accuracy. Based on
the calibrated and verify mapping of tamarix cover, the full area of invasive tamarix
within our study area was estimated.
3. Results and Analysis
3.1. O RIGINAL DATA ANALYSIS
Using twenty sampled pixels with three replications (60 total samples), all eight
cover types were compared for their gray-level values. Statistically, there were
high correlations among the three scanned color layers in each image. Correlation
73
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
Figure 4. Comparisons on separabilites of different covers by color composition.
coef cients of the rst layer versus the second and the third layers was 0.96 and
0.92, respectively. Therefore, to reduce the information redundancy, we only used
the blue layer to compare the gray-level difference of covers and derive the texture
features. Using analysis of variance, it was determined that there was signi cant
difference at the overall level, however, there existed tonal confusion among some
covers. The Turkey multi-comparison demonstrated that some covers partly over-
lap, which led to confusion in image classi cation (Figure 4). For example, tamarix
sometimes was confused with water and other shrubs, because of their color sim-
ilarity and spatial overlap where some mixing occurred within the 1 meter pixels.
Only when differences were far from the zero separation line, were these covers
easily separated. Therefore, just using gray-level or color alone, it was dif cult
to separate the tamarix from some other shrubs and water. There was also some
overlap of gray level among trees, shrubs and crops. Tamarix therefore was not
signi cantly different from shrubs, trees, or even water bodies. Clearly it was im-
possible to distinguish tamarix from its associated vegetation and background using
color alone, at this level of spatial resolution.
74 SHAOKUI GE ET AL.
3.2. E FFECTS OF WINDOW SIZES ON SEPARABILITY
Window size provided a functional range of texture features. The separability of
the original images was just 1.81 using the overall average. The separability of
Tamarix from other covers was 1.88. The separability between Tamarix and water
was only 1.03. When textures were combined with the original data, the separability
was signi cantly improved between different cover types. To some extent, the
level of change depended partly on the window size (Tables I IV). Therefore,
window size played an important role in capturing spatial pattern information to
improve separability among covers. If a window size is too small, some spatial
information contained in larger-scale patterns might be lost. However, if a window
size is too large, small cover patterns would be mixed with other patterns. Also,
the effect of a texture function may be in uenced by window size. For example,
TABLE I
Texture separabilities comparison with 3 3 windows
Tamarix separability
Overall separability
Covers of Min. Covers of Min.
Textures Max. Min. Average separability Max. Min. Average separability
Raw 2.00 1.03 1.81 Rocks-roads 2.00 1.03 1.88 Tamarix-water
Hom 2.00 1.20 1.91 Trees-shrubs 2.00 1.56 1.91 Tamarix-water
Mean 2.00 1.15 1.89 Tamarix-waters 2.00 1.15 1.89 Tamarix-water
Con 2.00 1.29 1.95 Trees-shrubs 2.00 1.76 1.95 Tamarix-shrubs
Dis 2.00 1.29 1.95 Trees-shrubs 2.00 1.78 1.94 Tamarix-water
SD 2.00 1.29 1.96 Shrubs-trees 2.00 1.91 1.98 Tamarix-trees
Entropy 2.00 1.30 1.91 Shrubs-trees 2.00 1.48 1.97 Tamarix-trees
TABLE II
Texture separability comparison with 5 5 windows
Tamarix separability
Overall separability
Covers of Min. Covers of Min.
Textures Max. Min. Average separability Max. Min. Average separability
Raw 2.00 1.03 1.81 Rocks-roads 2.00 1.03 1.88 Tamarix-water
Hom 2.00 0.95 1.87 Tamarixs-water 2.00 0.95 1.87 Tamarix-water
Mean 2.00 1.30 1.91 Tamarix-water 2.00 1.30 1.91 Tamarix-water
Con 2.00 1.23 1.96 Trees-shrubs 2.00 1.92 1.98 Tamarix-trees
Dis 2.00 1.36 1.95 Rocks-roads 2.00 1.91 1.98 Tamarix-trees
SD 2.00 1.23 1.96 Shrubs-trees 2.00 1.91 1.98 Taamrix-trees
Entropy 2.00 1.43 1.95 Rocks-roads 2.00 1.90 1.96 Tamarix-water
TABLE III
Texture separability comparison with 7 7 windows
Overall separability Tamarix separability
Textures Max. Min. Average Covers of Min. separability Max. Min. Average Covers of Min. separability
Raw 2.00 1.03 1.81 Rocks-roads 2.00 1.03 1.88 Tamarix-water
Hom 2.00 0.95 1.86 Tamarix-water 2.00 0.95 1.86 Tamarix-water
Mean 2.00 1.28 1.90 Tamarix-waters 2.00 1.28 1.90 Tamarix-water
Con 2.00 1.24 1.96 Nonevergreen trees-shrubs 2.00 1.92 1.98 Tamarix-nonevergreen trees
Dis 2.00 1.36 1.94 Nonevergreen Trees-shrubs 2.00 1.91 1.98 Tamarix-nonevergreen trees
SD 2.00 1.29 1.96 Shrubs-nonevergreen Trees 2.00 1.91 1.98 Tamarix-nonevergreen trees
Entropy 2.00 1.33 1.93 Shrubs-nonevergreen trees 2.00 1.83 1.94 Tamarix-nonevergreen trees
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
75
76
TABLE IV
Texture separability comparison with 9 9 windows
Overall separability Tamarix separability
Textures Max. Min. Average Covers of Min. separability Max. Min. Average Covers of Min. separability
Raw 2.00 1.03 1.81 Rocks-roads 2.00 1.03 1.88 Tamarix-water
Hom 2.00 1.15 1.93 Nonevergreen trees-shrubs 2.00 1.86 1.96 Tamarix-water
Mean 2.00 1.22 1.89 Tamarix-water 2.00 1.22 1.86 Tamarix-water
Con 2.00 1.26 1.95 Nonevergreen trees-shrubs 2.00 1.88 1.97 Tamarix-water
SHAOKUI GE ET AL.
Dis 2.00 1.31 1.96 Nonevergreen trees-shrubs 2.00 1.91 1.98 Tamarix-Nonevergreen trees
SD 2.00 1.29 1.96 Shrubs-nonevergreen trees 2.00 1.91 1.98 Tamarix-Nonevergreen trees
Entropy 2.00 1.30 1.92 Shrubs-nonevergreen trees 2.00 1.72 1.92 Tamarix-water
77
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
the mean texture decreased separability in large window sizes. On the other hand,
entropy increases the separability in large window sizes. For the 3 3 window,
the overall average separability was improved from 1.81 to the range of 1.89
1.96; and the Tamarix separability was improved to 1.89 1.98. At this window
size, the standard deviation improved separability the most. For the 5 5 window,
the overall average separability was improved to 1.87 1.96, while the average
separability of Tamarix from other covers was improved to 1.83 1.98. The 7 7
and 9 9 window sizes had similar effects, but the corresponding separabilities
were a little lower than that of the 5 5 window. Therefore, the 5 5 window size
was nally chosen as the optimal window size for analysis and was then used to
extract texture features with these photographs, and was set as the standard window
size in all texture measurements along the Cache Creek. These textures were then
applied to classify all the images, the images were then mosaiced and used for
mapping the saltcedar and estimating overall area of infestation. Among different
texture features, those measuring heterogeneity such as the standard deviation and
dissimilarity played a more important role in improving separability than the other
measures discussed. Based on an overall average and the speci c separability of
Tamarix compared to other classes, the two homogeneity measures gave the poorest
performance.
3.3. T HE DETERMINATION OF THE OPTIMAL TEXTURE ALGORITHM
The overall accuracy and the speci c accuracy for Tamarix ( Table V) were used to
determine the optimal texture for detection and classi cation of invasive saltcedar
across large areas along Cache Creek. A high separability did not necessarily corre-
spond to high accuracy in all categories, but was true for saltcedar cover detection
as follows. The overall accuracy was calculated for all eight covers from different
textures for comparison with the raw image data. The accuracy for Tamarix was
singled out separately from other classes as this was the target species of most
importance in our analysis. The overall accuracy and accuracy for Tamarix across
different wondow sizes were sometimes inconsistent. For instance, the dissimilarity
TABLE V
Accuracy assessment of the classi cation with 5 5 window size
Average Overall Tamarix Min. Target Max Target
accuracy accuracy accuracy accuracy of Min. accuracy accuracy of Max. accuracy
Raw 90.99 91.38 87.92 78.49 Nonevergreen trees 99.46 Bare elds
Con 93.46 94.74 98.61 88.81 Roads 99.37 Bare elds
Dis 93.72 95.34 84.95 86.17 Nonevergreen trees 100.00 Bare elds
SD 94.37 95.58 97.43 80.79 Roads 100.00 Bare elds
Entropy 96.70 97.41 98.22 87.34 Nonevergreen trees 100.00 Bare elds
78 SHAOKUI GE ET AL.
texture with a 3 3 window size led to a very high overall accuracy; however, the
corresponding saltcedar accuracy was poor. Therefore, the best texture algorithm
was judged by balancing among the average, maximum and minimum accuracies
for both overall separability and for Tamarix separability. For most of the images,
the contrast texture was the best choice for saltcedar mapping. Therefore, it was
included in all image classi cation processes. Although the degree of improvement
in separability may vary from photograph to photograph due to topography and
light conditions, it was shown that texture analysis always improved invasive cover
detection, and that consecutive images usually shared the same optimal texture set
(Table VI).
3.4. P OST PROCESSING EDITING AND ACCURACY ASSESSMENT
Tamarix, as the focal cover type, only exists along or near creeks. However, in the
initial classi cation, some orchards were also classi ed as Tamarix due to similar
bloom characteristics. Tamarix was also occasionally confused with shrubs and
trees in the riparian zone, but rarely. Based on its habitat characteristics, upland
areas where orchards occur and cause apparent classi cation errors, were corrected
and edited as non-invasive covers, and thus the classi cation results were further im-
proved. In addition, three other non-objective covers (roads, rocks, and bare elds)
were merged together for nal assessment and analysis. These post-classi cation
processed images were then mosaiced with increased accuracy for nal assessment
and eld use.
To estimate producer s accuracy, we rst randomly sampled 400 pixels from
classi ed Tamarix covers along the creek. However, most of these points were
inaccessible, due to water bodies blocking and steep banks limiting access. A total
of 95 sample points were reached in the eld. Among these, 74 samples were
veri ed to be Tamarix cover (accuracy was 77.89%). Only 12 misclassi ed pixels
were caused by dif culties in separating the classes with the color and texture
methods used in the analysis. For example, there were 4 willow pixels classi ed as
saltcedar. There were other objective factors causing some errors that we hope to
eliminate in future assessments. For example, there were 5 sample points classi ed
as Tamarix that were removed at the time of the eld veri cation, because it was
found that there were dead patches of saltcedar around speci c sample locations
where local landowners had chemically treated and killed these plants. There were
also 4 errors in the samples caused by georeferencing problems along steep slopes,
for example, near a dam, extreme creek banks, or uneven road edges. For user s
accuracy assessment, we used a differential submeter GPS to collect 157 randomly
selected eld points where saltcedar occurred in patches along the creek. These
samples were then checked and compared with the classi ed saltcedar covers back
in the laboratory. There were 112 samples classi ed as Tamarix; however, there
were 45 samples not found as tamarix in the map. The resulting user s accuracy
was 71.33%. In this case, it was found that the misclassi cation was caused to some
79
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
TABLE VI
The optimal texture measure with different photographs
Photo No. UTM coordinates Texture Photo no. UTM coordination Texture
1 605535E, 4288576N DEV 16 576180E, 4289000N CON
608430E, 4285739N 580283E, 4284835N
2 603344E, 4288607N DEV 17 574660E, 4290559N CON
606494E, 4285566N 578809E, 4286403N
CON + DEV
3 601775E, 4289077N 18 573141E, 4294140N CON
605690E, 4285193N 577164E, 4288138N
4 600170E, 4287628N CON 19 572622E, 4292646N CON
604096E, 4283738N 576464E, 4288832N
5 598642E, 4286268N CON 20 571687E, 4294644N CON
602592E, 4282305N 575469E, 4290885N
6 594683E, 4284239N CON 21 570776E, 4296627N CON
598181E, 4280859N 574429E, 4293042N
7 592608E, 4284969N CON 22 569786E, 4298835N CON
596098E, 4281569N 573439E, 4294552N
8 590332E, 4285671N CON 23 568570E, 4300738N
593838E, 4282197N 572346E, 4296928N
DIS + DEV
9 588410E, 4286137N 24 567690E, 4302862N
591860E, 4282713N 571462E, 4299014N
CON + DIS + DEV 25
10 586529E, 4286626N 566908E, 4303654N
590029E, 4283206N 570825E, 4300572N
11 584459E, 4287248N DEV 26 566424E, 4304561N CON
587874E, 4283773N 570175E, 4301795N
12 583178E, 4287125N DEV 27 565631E, 4306054N CON
586253E, 4284205N 569296E, 4302458N
13 581129E, 4286824N CON 28 564882E, 4306530N CON
584054E, 4283989N 568845E, 4302605N
14 579046E, 4286667N 29 563348E, 4307980N CON
582038E, 4283782N 567205E, 4304042N
15 577736E, 4287528N CON 581816E, 4283467N
581816E, 4283467N
Note: not using any texture.
degree by the spread of this invasive species after the photos were taken or small
patch sizes that may not have been easily visible in the 1 meter pixels. These photos
were taken in 2001, but the retrospective analysis and nal accuracy assessment
were not completed until the summer of 2004. Although saltcedar is a relatively
slow growing shrub, the three-year time delay certainly added to this error. In all, we
found that there were 13 misclassi ed samples near big patches of saltcedar on the
photos in 2001 that could have been caused by Tamarix dispersal after the photos
80 SHAOKUI GE ET AL.
were taken. If these samples were checked in 2001, the corresponding accuracy
would certainly have been higher. If we exclude these 13 samples for the accuracy
assessment, then the accuracy would have been 77.78%. In the future, in order
to have more exact invasion information, the invasive map should be updated and
veri ed each year. It was also determined that 32 samples of isolated saltcedar were
not found on the resulting classi cation map. Such errors were mainly caused by
small patches of saltcedar. These small infestation areas were not easy to capture
in the image analysis process due to our 1 meter pixel resolution and the texture
analysis conducted in the relatively large 5 5 window size. This might be improved
by conducting the analysis under ner image resolution, however, constraints of
image processing and handling made that level of detail in the analysis impractical
and cost prohibitive. Finally, using these analyses and assessing the overall classi ed
Tamarix cover along a 40 km section of Cache Creek, the total area of invasive
saltcedar was estimated to be 396 hectares (3.96 km2 Table VII).
It was found that there were many more saltcedar invaded areas in the middle
section of the creek and that the invaded areas in up-stream section were less in
area and less dense. Through correlation analysis, it was found that the invasive
areas were strongly linked with bare soil areas of and water adjacent to the creek
(including wetland) (Table VIII). Saltcedar infestation negatively correlates with
the amount of bare eld area, and positively with the area ooded by water. In the
middle section, the area of water in the creek is large but the water ow is slow.
TABLE VII
Invasive characteristics in different parts along Cache Creek in 2001
Tamarix areas Length of water Invasive degree
Creek parts (Square meters) Body (Meter) (Square meters / meter)
Down stream 522***-***** 40.00
Middle part 918***-***** 67.54
Up stream 521***-***** 36.11
TABLE VIII
Correlation coef cients among different cover types
Non-evergreen Bare Water bodies
Tamarix Evergreen (including shrubs) Crops elds (including wetland)
0.58 0.77 0.47 0.99
Tamarix 1.00 0.96
0.99 0.69 0.78
Evergreen 1.00 0.97
0.93 0.85 0.92
Non-evergreen 1.0000
1.00 0.59
Crops 0.70
1.00 0.99
Bare elds
Water bodies 1.00
81
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
Possibly, such water condition forms a habitat condition bene cial to saltcedar
seed germination and invasion along the creek The invasion of Tamarix also com-
petitively reduces native vegetation in infested areas leading to a further negative
correlation with both evergreen and non-evergreen plant covers in these areas.
4. Discussion
Previous studies of invasive plant species detection using remote sensing and image
analysis have primarily emphasized the use of re ectance or color properties alone.
In those studies results have provided some but limited success. In situations where
effective detection has been achieved, it has been attributed to substantial differences
in re ectance caused by changing phenological characteristics between the species
of interest and the associated environment (Everitt and Deloach, 1990; Everitt
et al., 1995; Lass et al., 1996). In this research, color differences from natural color
photographs were not suf ciently large between the invasive target species and
other associated background vegetation to allow accurate classi cation. Texture
analysis increased the separability between invasive cover and these background
plants and other habitat features.
We intended to nd proper texture(s) from six different measures. In fact, the
spatial dispersion patterns in plant ecology are typically scale-dependent, and thus
we hoped to nd an appropriate texture measure to describe the spatial pattern
of gray levels of a pixel neighborhood at different scales of study. That is, after
evaluating several different scales (different window sizes), we applied the optimal
texture assessment to extract the spatial neighbor relationships with a particular
window size (5 5 pixels). This texture was then used as an additional layer in image
classi cation and improved the average separability (minimum accepted value set
to 1.90). However, occasionally in some images, a selected window size, using only
one texture did not reach a separability of over 1.90. Thus we included two or three
different texture measurement to capture the optimal texture for comparison and
thus ensured that the separability was always greater than 1.90.
Sometimes the separability improvement was not much different among textures
producing similar results as seen in the contrast and dissimilarity texture measures.
The optimal texture assessment method could only be determined after the full
accuracy assessment was completed, and it was expected to have a higher overall
accuracy and the best accuracy for the speci c classi cation of Tamarix. During this
process, we also assessed the use textures other than the selected optimal texture(s)
to improve classi cation accuracy. In the assessment process, we found that none
of the other textures increased accuracy more than 2%, even when using all six
texture measures simultaneously to do the classi cation. Therefore in the end, we
used the optimal texture subset alone, instead of all the texture measures combined
to extract the Tamarix cover. These results suggest that gray-level values of digitized
photographs in combination with a single texture feature can be effectively used
82 SHAOKUI GE ET AL.
in an automatic procedure to detect invasive Tamarix in large areas of infestation.
Such detection is possible for not only large tracts of invasive areas but also for
small patches where local land managers are interested in detecting and eradicating
these plants.
5. Conclusions
In this study, we demonstrated that texture could be used to improve the separability
between invasive Tamarix parvi ora and associated vegetation along an example ri-
parian corridor in California. Texture features were useful in developing a relatively
automatic procedure for detection and recognition of Tamarix using historical color
aerial photographs. We derived six different texture algorithms from a gray-level
co-occurrence matrix (GLCM), which proved useful for improving classi cation
results. The optimal window size was 5 5 among 4 other bracketing window
sizes. This optimal size was chosen based on the separability of cover types with
special attention paid to the separability of Tamarix when compared to other cover
types. The producer s and user s accuracies were found to be 77.89% and 71.33%,
respectively. Totally, there was 396 hectares (3.96 km2 ) of Tamarix cover along the
invaded Cache Creek drainage.
Acknowledgments
We are grateful to Dr. Ruiliang Pu for his help on the use of PCI software and
discussions on the application of remote sensing for use in mapping invasive species.
We also thank Mr. Justin Weber, Mr. Joel Garza and Ms. Julie Garren for their
eld validation work. In particular, we thank USDA-ARS # 5325-22000-017-00D
(Biology and Control of invasive weeds on the Western United States) and USDA-
CSREES-IFAFS # 00-521**-**** (Biologically-based control for the area-wide
management of exotic and invasive weeds).
References
Anys, H. and He, D. C.: 1995, Evaluation of textural and multipolarisation radar features for crop
classi cation, IEEE Trans. on Geosci. Remote Sens. 33, 1130 1181.
Bailey, Joseph K., Schweitzer, J. A. and Whitham, T. G.: 2001, Salt cedar negatively affects biodi-
versity of aquatic macroinvertebrates, Wetlands 21, 442 447.
Brock, J. H.: 1994, Tamarix spp. (Salt Cedar), An Invasive Exotic Woody Plant in Arid and Semi-arid
Riparian Habitats of Western USA, In: L. C. de Waal, L. E. Child, C. P. M. Wade, J. H. Brock,
Landscape Ecology Series; Ecology and Management of Invasive Riverside Plants, John Wiley
and Sons Ltd.; John Wiley and Sons, Inc.; Chichester, UK/New York, USA, 1994, pp. 27 44.
Cleverly, J. R., Smith, S. D., Sala, A. and Devitt, D. A.: 1997, Invasive capacity of Tamarix ramo-
sissima in a Mojave desert oodplain: The role of drought, Oecologia (Berlin) 111, 12 18.
83
TEXTURE ANALYSIS FOR MAPPING TAMARIX PARVIFLORA
Di Tomaso, J. M.: 1998, Impact, biology and ecology of saltcedar (Tamarix spp.) in the southwest
United States, Weed Tech. 12, 326 336.
Everitt, J. H. and Deloach, C. J.: 1990, Remote sensing of Chinese Tamarisk (Tamarix chinensis)
and associated vegetation, Weed Sci. 38, 273 278.
Everitt J. H., Escobar, D.E. and Davis, M. R.: 1995, Using remote sensing for detecting and mapping
noxious plants, Weed Abstr., 44, 639 649. CAB internal.
Everitt, J. H., Escobar, D. E, Alaniz, M. A, Davis, M. R. and Richardson, J. V.: 1996, Using spatial
information techniques to map Chinese tamarisk (Tamarix chinensis) infestations, Weed Sci. 44,
194 201.
Gausman, H. W., Everitt, J. H., and Gerbermann Bowen R. L.: 1977, Canopy re ectance and lm
image relations among three south Texas rangelands plants, J. Range Manage. 30, 449 450.
Gong, P., Marceau, D. J. and Howarth, P. J.: 1992, A comparison of spatial feature extraction algo-
rithms for land-use classi cation with SPOT HRV data, Remote Sens. Environ. 40, 137 151.
Haralick, R. M.: 1979, Statistical and structural approaches to texture, Proc. IEEE, 67, 786 804.
Haralick, R. M., Shanmugam, K. and Dinstein, I.: 1973, Textural features for image classi cation,
IEEE Tran. Geosci. Remote Sens. 33, 1170 1181.
Hassan, A. and He, D. C.: 1995, Evaluation of textural and multi-polarization radar features for crop
classi cation, IEEE Tran. Geosci Remote Sens. 33, 1170 1181.
Hudak, A. T. and Wessman C. A.: 1998, Textural analysis of historical aerial photography to char-
acterize woody plant encroachment in South African Savanna, Remote Sens. Environ. 66, 317
330.
Jonathan, C.-W. C., Defries, R. S. and Townsend, R.G.: 2001, Improved Recognition of Spectrally
Mixed Land Cover Classes Using Spatial Textures and Voting Classi cations, CAIP 2001, LNCS
2124, 217 227.
Lass, L. W., Carson, H. W. and Callihan, R. H.: 1996, Detection of yellow starthistle (Centaurea sol-
stitialis) and common St. Johnswort (Hypericum perforatum) with multispectral digital imagery,
Weed Tech. 11, 248 256.
Mather, P. M., Tso, B. C. K. and Koch, M.: 1998, An evaluation of Landsat TM spectral data and
SAR Derived textural information for Litho logical discrimination in the Red Sea Hills, Sudan,
Int. J. Remote Sens. 19, 587 604.
Nyoungui, A. N., Tonye, E. and Akono, A.: 2002, Evaluation of speckle ltering and texture analysis
methods for land covers classi cation from SAR images, Int. J. Remote Sens. 23, 1895 1925.
Podest, E. and Saatchi, S.: 2002, Application of multiscale texture in classifying JERS-1 radar data
over tropical vegetation, Int. J. Remote Sens. 23, 1487 1506.
Riou, R. and Seyler, F.: 1997, Texture analysis of tropical rain forest infrared satelliate images, PE
and RS. 63, 515 521.
Saatchi, S. S., Nelson, B., Podest, E. and Holt, J.: 2000, Mapping land cover types in the Amazon
Basin using 1 km JERS-1 mosaic, Int. J. Remote Sens. 21, 1183 1200.
Shaban,M. A. and Dikshit, O.: 2001, Improvement of classi cation in urban areas by the use of
textural features: The case study of Lucknow city, Uttar Pradesh Int. J. Remote Sens. 22, 565
593.
Smits, P. C. and Annoni, A.: 1999, Updating Land-cover maps by using texture information from
very high-resolution space-borne imagery, IEEE Trans. Geosci. Remote Sens. 37, 1244 1254.
Stohlgren, T. J., Binkley, D., Chong, G. W. Kalkhan, M. A., Schell, L. D., Bull, K. A. Otsuki, Y.,
Newman, G., Bashkin, M. and Son, Y.: 1999, Exotic plant species invade hot spots of native plant
diversity, Ecol. Monog. 69: 25 46.
Xu, B., Gong, P., Spear R. and Seto, E.: 2003, Comparison of different gray level reduction schemes
for a revised texture spectrum method for land-use classi cation using IKONOS imagery, PE.
RS. 69, 529 536.
Zavaleta, E.: 2000, The economic value of controlling an invasive shrub, Ambio 29, 462 467.