Comparing Texture Analysis Methods

through Classification
Philippe Maillard

Abstract used for the experiments. Next comes the results and their
The development and testing of two techniques of texture analysis followed by the main conclusions.
analysis based on different mathematical tools—the semi-
variogram and the Fourier spectra—are presented. These are Background
also compared against a benchmark approach: the Gray-Level Reed and du Buf (1993) claim that most development in texture
Co-occurrence Matrix. The three methods and their implemen- has been concentrated on feature extraction methods (some-
tation are briefly described. Three series of experiments have times called channel-based methods) which seek to extract rel-
been prepared to test the performance of these methods in evant textural information and map it onto a special dedicated
various classification contexts. These contexts are simulated channel called a feature. The authors classified the various fea
by varying the number, type and visual likeness of the texture ture extraction methods as belonging to one of three possible
patches used in classification tests. More specifically, their classes: feature-based, model-based, or structural. Cocquerez
ability to correctly classify, separate, and associate texture and Philipp (1995) have used a similar classification of image
patches is assessed. Results suggest that the classification segmentation methods which they compare in varioussitua-
context has an important impact on performance rates of all tions (including textured images).
methods. The variogram-based and the Gray-Tone Depen- In feature-based methods, characteristics of texture (such as
dency Matrix methods were generally superior, each one in orientation, spatial frequency, or contrast) are used to classify
particular contexts. homogeneous regions in an image. Model-based methods rely on
the hypothesis that an underlying process governs the arrange
Introduction ment of pixels (such as Markov chains or Fractals) and try to
As scientists and researchers of the remote sensing community extract the parameters of such processes. Structural methods
began to use high spatial resolution data, it soon became clear assume that a texture can be expressed by the arrangement of
that spectral-based methods of computer classification and seg- some primitive element using a placement rule. Feature-based,
mentation were doomed to yield unsatisfactory results. At model-based, and hybrid methods have overwhelmingly domi-
high resolution, conceptual objects like forests or pasture usu- nated the scene in the last 20 years or so. One of their findings
ally show significant variations in their pixel values (Strahler was that, although so many different methods have been devel
et al., 1986). Stationary in nature, these variations can give rise oped, no rigorous quantitative comparison of their results had
to an apparently regular spatial pattern referred to as texture ever been done, which is a major theme of the present work.
(Kittler, 1983). One of the key elements that the interpreters use Because Bela Julesz (1965) has shown evidence that human
to identify and analyze images is clearly the spatial arrange- perception of texture could be modeled using second-order sta-
ment of color and tone that form natural visual entities: visual tistics (although he would later change his theory for the “tex-
texture (Haralick et al., 1973; Pratt et al., 1978). ton” approach; see Julesz (1981)) many researchers have explored
Because there is no universally accepted definition of second-order statistics as possible features for texture analysis.
visual texture, one has to choose a definition that best reflects Among the most common second-order statistics that have been
the objective or the results being sought. The definition used are the co-occurrence matrix, the spatial-autocorrelation,
adopted here was given by Pratt (1991, p. 505): natural scenes the covariogram, and the semi-variogram.
containing semi-repetitive arrangements of pixels. The prob- The frequency domain approach, also referred to as the
lem of analyzing and classifying texture has generated a wealth Fourier Spectra approach, has been a long time favorite for tex-
of studies and techniques that are seldom compared in a sys ture analysis. From the early attempts at using it as a texture
tematic way. This study is an experimental analysis of the analysis tool by Rosenfeld (1962) to the recent use of Gabor
problem of classifying texture using different mathematical functions as filters in the frequency domain to create frequency-
tools. In particular, the specific classification context is ana- and orientation-specific texture features (e.g., Fogel and Sagi,
lyzed in terms of the effect of between-class variation and num- 1989; Jain and Farrokhinia, 1991; Manjunath and Ma, 1996),
ber of classes on classification accuracy. To achieve the latter, a the Fourier transform offers infinite possibilities not only for
special experimental framework has been prepared and experi- texture analysis but for applications requiring the analysis of
mental results are presented and discussed. spatial frequencies and their orientation.
The paper is organized in six sections. A short background In order to evaluate a technique, it is necessary to have some
review of feature extraction methods for texture analysis fol- base for comparison. In this research, the comparison will take
lows the introduction. Then the three approaches are described
individually and compared through sample data. The fourth
section describes the experimental framework and the data
Photogrammetric Engineering & Remote Sensing
Vol. 69, No. 4, April 2003, pp. 357–367.
Universidad Federal de Minas Gerais, Departamento de Carto 0099-1112/03/6904–357$3.00/0
grafiá, Av. Antônio Carlos, 6627, Belo Horizonte MG 31270- © 2003 American Society for Photogrammetry
091, Brazil ([email protected]). and Remote Sensing

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2003 357

the form of another technique that has already been widely After numerous tests using different ways to transform the
accepted by the scientific community and obviously performs variogram into texture features, the most promising approach
well. This method was proposed by Haralick et al. (1973) who was found to be the averaging of selected distance lag intervals.
have named it Gray-Tone Spatial-Dependence Matrices, also The SRPD texture feature extraction routine can be summarized
known as Gray-Level Co-occurrence Matrices (GLCM). Not only in the following steps:
do almost all the authors in visual texture analysis quote the
• For every pixel in the image, a neighboring window (32 by 32
GLCM, but many have already used it as a comparison technique.
pixels) is considered and four directional variograms (0°, 45°,
Among them, Davis et al. (1979), Conners and Harlow (1980), 90°, and 135°) are computed for all possible combinations in
Pratt (1991), Bonn and Rochon (1992), Wu and Chen (1992), that window.
Reed and du Buf (1993), Dikshit (1996), Franklin et al. (2001) • The maximum lag size is equal to one half the window size.
and Zhang (2001) have either used the GLCM method as a com- • The mean Square-Root Pair Difference is used as a semi-vario-
parison or have described it in their review. gram estimation.
It was decided that all three methods be rotation-invariant • Six values (features) of the SRPD are computed by giving more
so that the particular orientation of texture would not be con- weight to the values corresponding to the smaller lags; in other
sidered even though it was observed that considering particular words, by computing the SRPD features over regular intervals
on a logarithmic scale. Figure 1a illustrates such a scale with
orientations can increase classification accuracy (Franklin and
six lag ranges while Figures 1b and 1c illustrate the semi-vari-
Peddle, 1989; Maillard, 2001). This was important especially ance graph before and after the averaging, respectively.
for the crops and waves classes for which the factors control- • These values are computed for all four directions for a total of
ling their orientation are difficult to predict. 24 features
• The 24 directional features are then transformed to 18 rotation-
Description of the Three Texture Feature Extraction Methods invariant features: for each lag, the mean, standard deviation
and sum of perpendicular ratios ([g0/g90  g90/g0  g45/g135
The Variogram Approach  g135/g45] where g is the estimate of variance) are computed
Many authors have already shown the potential of the vario- where the latter two parameters are meant to preserve anisotropy
gram as a texture analysis approach (Serra, 1982; Woodcock et in the data.
al., 1988; Ramstein and Raffy, 1989; Miranda et al., 1992; Atkin-
son, 1995; St-Onge and Cavayas, 1995; Lark, 1996). On the one
hand, the variogram is related to other statistical approaches The Fourier Approach
like the autocorrelation function and the fractal Brownian The Fourier-based method proposed by Stromberg and Farr
motion (Xia and Clarke, 1997). On the other hand, it is compu- (1986) was surely the first really successful use of the Fourier
tationally simple and easy to interpret as a graph. One point in Transform through the application of a series of ring (band-
which the variogram appears more appropriate is that only pass) filters before applying the inverse transforms from which
weak stationarity is assumed, in other words, the expectation the texture features resulted. This was consistent with the find-
only has to be constant locally (Woodcock et al., 1988). ings of Richard and Polit (1974) and Harvey and Gervais (1978)
It appears, however, that most techniques using the vario- in their psychophysical experiments. But, as Caelli (1982)
gram do so in the geostatistical manner, i.e., a model is usually would later show, it lacked the orientation component. This
applied whose parameters are taken as a way of describing the was largely corrected by the method proposed by Wilson and
semi-variogram curve. In Remote Sensing images some texture- Spann (1988), which divided the spectral domain into a series
based variograms might be best modeled using the spherical of band-pass/orientation-pass filters defined by a finite prolate
model while others are best represented with an exponential spherical function. Others then suggested the use of a bank of
or even sinusoidal model. This poses a problem in terms of cre- Gaussian filters applied locally (as opposed to the whole image)
ating a systematic approach. One solution would be to use the in the frequency domain to create the textural features (Gorenic
“best” model type, selected as a texture feature. But using a and Rotman, 1992). The most popular Fourier-related texture
nominal scale feature would cause problems further down the analysis method is without doubt the “wavelet” transform
classification process. This would also imply that a battery of which uses a similar approach but replaces the Gaussian filters
models would have to be fitted for all pixels of all texture sam- by the FT of a Gabor function, also applied locally (Reed and
ples, and the cost in terms of computing would be high. For Wechsler, 1990; Jain and Farrokhinia, 1991; Dunn et al., 1994;
these reasons and because others have already pursued that line Manjunath and Ma, 1996). The method proposed here makes
of research, the “traditional” function representation of sill and use of these most recent findings but also attempts to simplify
range has not been considered here. significantly these methods and make them more computation-
Another point that has received attention is the alternate ally efficient.
use of the mean square-root pair difference (SRPD(h)) function In an effort to simplify the approach and to avoid a two-
proposed by Cressie and Hawkins (1980) as a semi-variance dimensional transform, the Fourier approach was imple-
estimator which is resistant to outliers. Lark (1996) has also mented through appending all lines of pixels in the window in
shown that for four different classes of texture (urban, farm- four directions (0°, 45°, 90°, and 135°). In this approach, the fre-
land, woodland, and meadow), when tested for normality, the quencies between zero and the number of lines appended are
SRPD(h) function scored much better than the g(h) function, a artifact and irrelevant because the appending of lines of image
fact confirmed by an earlier study by the author (Maillard, pixels is an artificial process. These artifact frequencies were
2001). eliminated. The following steps summarize the computation
Considering these findings, a number of considerations of the Fourier texture features:
were taken to implement the texture feature extraction based • For every pixel in the image, a neighboring window is consid-
on the variogram: ered and four directional transforms (0°, 45°, 90°, 135°) are com-
• a rather large window had to be used in order to cover larger puted for all appended lines; frequencies between 0 and 32
distance lags (up to 32 pixels), were eliminated; a window size of 32 by 32 pixels is used for
• the texture feature set had to be rotation-invariant but had to the 0° and 90° directions and of 64 by 64 pixels for the 45° and
preserve anisotropy, and 135° directions.
• the behavior of the SRPD graph near the origin had to receive • Six Gaussian filters are applied to the transform to create the
special attention because it bears a special significance in terms following texture frequency features: very low (centered on 1
of micro-texture (Serra, 1982; Jupp et al., 1989; Xia and cycle per window of 32 pixels), low (2 c/w), mid-low (4 c/w),
Clarke, 1997). mid-high (8 c/w), high (14 c/w), and very high (25 c/w). An

358 April 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 (a)

(b) (c)

Figure 1. The construction of texture features for the Variogram approache to texture
classification. (a) The logarithmic scale used to average SRPD values. (b) The original
SRPD graph showing all the values. (c) The six SRPD values based on averaging the semi-
variance according to a logarithmic scale.

average value is computed for each filtered result of each direc- lick et al., 1973). Because many of the features first described by
tion for a total of 24 directional features. Figure 2 illustrates Haralick are highly correlated among themselves (Haralick et
the process of filtering; in Figure 2a the Gaussian filters are al., 1973), a pre-selection was done to reduce the 14 possible
presented while Figure 2b shows the effect of applying the measures to less than half. The selection was done by combin-
filters on a sample transform of a forest image.
• The 24 directional features are then transformed to 18 rotation-
ing all the features used by many different research teams that
invariant features: the mean, standard deviation and sum of have used the GLCM method. Table 1 gives a listing of the
perpendicular ratios are computed for each frequency band. authors considered and the texture features they have used.
Analyzing the table revealed that the most commonly used fea-
tures are in decreasing order of popularity: Angular Second
The Gray-Level Co-occurrence Matrices Moment, Entropy, Inertia (initially contrast), Correlation, and
The Gray-Level Co-occurrence Matrices (GLCM) method was Inverse Difference Moment.
implemented in a manner similar to its original form (in Hara- Apart from the texture features used, pixel pair sampling dis-
tances have to be chosen with respect to the expected spatial fre-
quencies present in the images. The choice of sampling distance
is as important as the types of measurements. In order to be as
objective as possible, the sampling distances have been chosen
based on the visual analysis of the semi-variograms of the sample
texture patches. This analysis yielded the following distances:
three, six, and twelve pixels. In their original setting, Haralick et
al. would choose a particular sampling distance and then rotate it
by steps of 45 degrees so that, for a distance of three pixels, the
x,y sampling distances setting would be (3,0), (3,3), (0,3), and
(3,3) for the 0°, 45°, 90°, and 135° orientations, respectively.
Then, for each sampling distance, the mean and standard devia-
(a) tion would be computed over the four orientations instead of
using each orientation separately. Therefore, the features are not
orientation-specific but still account for some effect of anisot-
ropy. This approach is meant to obtain rotation-invariant features
similar to those adopted for the variogram and Fourier methods.
The following steps summarize the implementation of the
GLCM method:

• For every pixel in the image, a neighboring window (32 by 32

pixels) is considered and matrices are computed for three differ-
ent pixel pair sampling distances (3, 6, and 12 based on analysis
of the variograms) and four directions (0°, 45°, 90°, and 135°).
• Five measurements are computed for each matrix: contrast,
angular second moment, inverse difference moment, entropy,
and correlation for a total of 60 features.
(b) • The 60 “directional” features are transformed into 30 rotation-
invariant features by computing the mean and standard devia-
Figure 2. Filtering the Fourier transform of a single (forest) tion over the four original directions.
pixel. (a) The six Gaussian filters. (b) Result of applying the
six filters on the Fourier transform of the neighboring window Comparison of the Three Approaches
of a forest pixel (0º direction). Displaying the texture features from the three methods as
images would be of little help in understanding how these

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2003 359

 TABLE 1. LIST OF THE GLCM FEATURES USED IN SOME TEXTURE ANALYSIS STUDIES
Author(s) and Year Feature Used (See Legend at Foot of Table)
Haralick et al., 1973–Three Experiments (a) ASM, Con, Cor (b) ASM, Con, Cor, Ent (c) ASM, Con, Cor, SSq, IDM, SAv, SV,
SEnt, Ent, DV, DEnt
Haralick, 1979 ASM, Con, Cor, IDM, Ent
Davis et al., 1979 ASM, Con, Cor, Ent
Conners and Harlow, 1980 ASM, Cor, IDM, Ent
Rosenfeld and Kak, 1982 ASM, Con, IDM, Ent
Gonzalez and Wood, 1992 ASM, Cor, IDM, Ent, Other
Sali and Wolfson, 1992 ASM, Con, IDM, Ent, DEnt
Wu and Chen, 1992 ASM, Con, Cor, IDM, SAv, SV, SEnt, Ent, DV, DEnt, IMC
Anys and He, 1995 SSq, ASM, Ent, Con, Cor, Others
Dikshit, 1996 Ent, Con, IDM, ASM
Hay et al. (1996) ASM, Ent, IDM, SV
Legend: ASM: Angular Second Moment, Con: Contrast (or inertia), Cor: Correlation, SSq: Sum of Squares, IDM: Inverse Difference Moment,
SAv: Sum Average, SV: Sum Variance, SEnt: Sum Entropy, Ent: Entropy, DV: Difference Variance, DEnt: Difference Entropy, IMC: Information
Measure of Correlation, MCC: Maximum Correlation Coefficient, Other: Other measures.

multi-dimensional spaces can help separate textures of various i.e., forest, residential, desert, crops, shrub, and waves (samples
origin. A graph representation was preferred for its simplicity of these classes are presented in Figure 4).
and ease of interpretation. Figure 3 illustrates the data gener- The semi-variogram (Figure 3a) displays variance as a
ated by each method as a series of graphs for six texture classes: function of lag distance. The sill reached by each texture

(a) (b)

(c)
Figure 3. Six texture classes mapped from the three different methods of texture analysis. (a) Semi-variogram. (b) Fourier
transform. (c) Five measures taken from the GLCM.

360 April 2 003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 lard, 2001). Such a framework has already been proposed by the
MeasTex project through three aspects: (1) the implementation
of major texture processing paradigms, (2) a library of texture
images, and (3) a measure of their performance (MeasTex,
2001). In this work, some further refinements of such frame-
work are proposed through the use of a texture base exclu-
sively dedicated to visual textures found in high-resolution
images of the Earth and a series of specific experiments, both of
which are described below.

Data Set Description

The texture samples were extracted from scanned aerial photo-
graphs of the Emerald (desert samples) and Brisbane areas
(both in Queensland, Australia) and correspond to six different
generic classes of texture: forest, residential areas, arid envi-
ronment, agriculture, shrub, and ocean waves. Figure 4 pre-
sents the texture images used while the generic classes are
briefly described next.
Forest includes basically all natural tree formations, both
closed and open. Reforestation is not included in this class.
Residential has been preferred to Urban because it usually
covers more extensive areas. Although all residential texture
Figure 4. Data set composed of 36 square patches of tex- patches have been selected in the Brisbane area, it could
ture from six generic classes: forest (F6), residential areas incorporate other types of residential areas like shantytowns
(R6), desert (D6), crops (C6), shrub (S6) and waves (W6). or other suburbs.
Each column (T6(a) to T6(f)) contains one sample of each Desert is a rather broad class that includes natural areas with
generic class. scarce vegetation in which the soil, rock, and low vegetation
are responsible for the texture patterns.
Agriculture is another very broad class that was intended to
incorporate all structured forms of agriculture including
orchards. Pasture is a somewhat marginal member of this
clearly separates them from each other, but it is the behavior of class only when it shows some artificial pattern.
the curve near the origin that is most determinant as a discrimi- Shrub was basically meant to complement the forest class
nating factor. The most striking features are the height and for low and open naturally vegetated areas. This class justi-
irregularity of the Residential class, the cyclic shape of the fies itself by its extent in many countries like Australia.
Crops and, to a lesser extent, Waves classes. Interestingly, the Waves is a generic class of bodies of water under the influ-
latter reaches its peak at a distance of three pixels and then ence of wind or currents. Both ocean and lakes are included,
drops and stabilizes with an undulating pattern. Apart from the but rivers remain a marginal member when wide enough to
height of the sill and the level of regularity of the curve, it is the produce a visible wave pattern.
slope described by the first few points that better differentiate
the six texture classes. The fact that these first few data points Six samples of each generic class were selected for a total
have received more attention in generating the texture features of 36 texture image samples of 128 by 128 pixels each. To assess
made it possible to preserve this important aspect. the appropriate resolution, a local variance was computed for a
The one-dimension Fourier transform of the same six tex- series of resolutions in the same manner suggested by Cao and
ture classes is presented in Figure 3b. The graph was con- Lam (1997) and yielded a graph with a first break at about two
structed as described above by appending all lines (in the meters and a peak at about three. Scanning the 1:25,000-scale
horizontal direction in this case) and then eliminating the arti- black-and-white photographs at a resolution of 450 dpi yielded
fact frequencies. Interestingly, the average amplitude of each a ground resolution of about 1.4 meters and ensured that most
texture frequency curve follows the same progression order as spatial variability was preserved without generating unneces-
in the sill reached in the semi-variogram graph; in increasing sarily large image sets. All texture samples are presented in
order: Crops, Desert, Forest, Waves, Shrub, and Residential. As Figure 4.
before, the Crops class has a well-defined frequency peak but so As can be observed, each column of the data set contains a
has the Shrub class, a feature not easily perceived in the semi- different sample of each generic texture class (named T6(a) to
variogram. The high frequency part of the curves (lower end) is T6(f)) while each row is composed of six different samples of a
much more confused and clear patterns are hard to identify. single generic class (named F6, R6, D6, C6, S6, and W6).
For the Grey-Level Co-occurrence Matrix, a series of mean- Although the choice criteria were somewhat suggestive, care
standard deviation plots were constructed for each co-occur- was taken to pick samples with low within-class variability but
rence based measure: contrast, angular second moment, high within-generic class variability. For example, each indi-
entropy, inverse difference moment, and correlation. The vidual Forest sample had to be homogeneous (texture-wise),
mean marks the center of each plot line while the height is but the generic class is represented by diversified samples of
determined by the standard deviation. Generally speaking, all forest.
five measures appear to separate well the six classes analyzed The first series of classifications will be carried out on each
except for two class pairs: Desert—Forest and Shrub—Waves. column of texture samples separately. This was a way of simu-
The former can only be distinguished by the correlation and the lating six different classification contexts (as, for instance, with
latter by the contrast and the inverse difference moment. different photographs). The second series of classifications
will be performed on each row of the set to simulate thematic
Experimental Methodology conditions in which one wishes to distinguish similar yet dif-
A special experimental framework is needed in order to evalu- ferent classes of texture. The last series of classifications will
ate and compare texture analysis methods (Smith, 2001; Mail- include all 36 samples.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2 003 361

The Comparison Tool that no special consideration has been taken for the borders and
Classification is a statistical tool used to assign an object to one edges of the texture samples can be seen as another source of
of a number of possible groups or classes with the help of a bias. In this case the bias is partly wanted in order to assess
how each method behaves with respect to texture edges. To
decision rule (James, 1985). The classification technique is
evaluate better how these edges alter the results, a separate
(usually) a point-dependent algorithm, meaning that only one Kappa statistic was computed where the edges (and their zone
pixel is considered at a time. In order for it to consider region- of influence) have been removed from the computation. This
based information such as texture, this information has to be is consistent with the fact that many authors consider that
first extracted in the neighbourhood of each pixel and coded on texture analysis and texture edge recognition should be
a different plane, which is called a texture feature here and is approached as different tasks. Pratt (1991) suggested that
performed here by one of three texture processing algorithms. boundaries should be established prior to attempting texture
Classification was preferred over segmentation approaches for measurements. Serra (1982) treats the edge effect from what
various reasons: he calls the individual-textural paradigm. Recent texture analy-
sis methods that integrate both region growing and edge detec-
● Classification algorithms have been used in Image Processing tion such as the one proposed by Pavlidis and Liow (1990) are
for quite some time and are probably the most widely used the consecration of such an approach.
pattern recognition techniques,
● The decision rules of classification algorithms are simple and Classify to Associate, Generalize, Merge, and Separate
easy to control whereas many region-dependant segmentation Although classification can be defined simply as a way of
algorithms use a threshold that can be hard to set, and assigning an object to one of a number of possible groups, it is
● The texture feature extraction (using one of the texture analysis
methods) can yield an analysis of the behavior of the various in fact a complex operation by which a geographical meaning is
textures for each method. given to a set of spectral and spatial measurements. According
to Nyerges (1991), four abstraction types have been used to sup-
In the present study, the problem is not to assign each pixel port meaning in geographical information: classification
to one of a group of possible classes in the test image because (“instance-of ” relationship), association (“member-of ” rela-
the class membership is already known for all pixels in the tionship), generalization (“kind-of ” relationship), and aggre-
image. An alternate formulation would be: can a given set of gation (“part-of ” relationship). Often, a classification process
variables describing the texture of a pixel makes its assignment is expected not only to classify but also to generalize and some-
to the “right” class possible? Or rather, to what extent can the times even to associate objects depending on the amount of
“right” texture class be assigned to the texture patches forming information available about the region being classified. Classi-
the test image given one of three methods of texture analysis? fication is used as a generalization tool by expecting it to elimi-
To ensure the most objective approach, a “blind” technique has nate unnecessary details, sometimes through the use of filters
been used for defining the training class samples. Training was either prior to or after the classification process. Association is
done through a mixture of systematic and cluster sampling: six- sometimes expected as in the case of opposing slopes of a hill
teen equally spaced clusters of 25 pixels (5 by 5) ensured that receiving different amounts of light and having different spec-
the class was well represented (2.44 percent) and that the sam- tral signatures but that should nonetheless be associated to the
ples were spread throughout the image. Although other studies same class.
have shown that clusters of up to 25 pixels reduce bias due to On another level of reasoning, the process of classification
autocorrelation (Congalton, 1988), it can still be expected is one that has both the functions of joining and separating
within the clusters because texture processing is always based objects as belonging to the same class or not. In a general land-
on neighboring windows of pixels. use/land-cover type classification, the different types of forest
For classification to be considered a testing tool, a number might not be known or relevant, in which case the classifier is
of assumptions had to be made, six of which are described expected to be able to join all forest types into one broad class.
here: Not having the information on all forest types present also pro-
● Bayes’ rule is optimal only if minimizing Total Errors of Classifi-
hibits effective sampling. In another situation, a user might
cation (TEC) is considered the “best” criterion; for instance, it want not only to differentiate the forest types but also their
does not take classification error cost into account (James, 1985). health status. Can a single classification algorithm be an effi-
● In its normal form, the Bayes’ rule assumes that all variables cient tool for both situations? Part of the answer might lie in the
from the measurement vector have a normal distribution. How- choice of the number and type of features (or variables) being
ever, as was pointed out by Scheffé (1959), because the sample used. Experiments have been selected to assess both types of
size is reasonably large (400) and is the same for all textures, situation as well as the generalization problem.
the assumption of normality can be relaxed (Clark and Hosking,
1986). Some consideration of normality has already been taken Description of the Experiments
by preferring the Square-Root Pair Difference (SRPD) to the
straight semi-variance because it has been found to have a more It is clear, from the observations above, that no single experi-
normal distribution and is more resistant to outliers (Lark, 1996). ment could assess all of these problems. Mike James (1985)
● TEC and its complement, Total Success of Classification (TSC), points out that classification methods have been criticized for
are acceptable ways to express the error measure through a working well during tests but giving poor results in a “real” sit-
confusion matrix. TEC is obtained by subtracting the number of uation but that this situation usually results from an inadequate
correctly classified objects from the total number of pixels. test procedure. To ensure adequacy of the test procedure, three
However, the Kappa statistic also takes the mere chance factor different classification experiments have been prepared in
into account (Foody, 1992) and so it was preferred as a classifi- order to assess all of the above aspects. These experiments are
cation accuracy value. explained below.
● It is usually assumed that the conceptual objects being classified
have low within-class variability but, in the case of texture, ● First experiment: evaluating the different approaches for sepa-
such a statement would have to be reviewed because texture rating and classifying sets of texture patches having very differ-
variability is not yet a well defined concept. ent visual characteristics and belonging to different generic
● All a priori probabilities are equal because all texture samples classes of objects.
are of equal size. ● Second experiment: assessing the separation capability of the
● Overall accuracy and the edge effect. The fact that the training methods by comparing their performance in separating sets of
areas are part of the texture sample being classified brings a similar textures belonging to the same generic class and evaluat-
bias in the classification results and so these pixels were with- ing their relative performance according to the textural context
drawn from the overall accuracy computation. Also, the fact (e.g., forest versus residential).

362 April 2 003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 TABLE 3. RESULTS OF PAIRWISE COMPARISON OF KHAT VALUES FOR OVERALL
RESULTS OF TABLE 2; S ⫽ SIGNIFICANT AT THE 99% LEVEL OF CONFIDENCE, NS ⫽
NOT SIGNIFICANT AT THE 99% L.O.C.

GLCM vs VGM VGM vs FFT FFT vs GLCM

(a) Z Z Z
Data set Statistic Result Statistic Result Statistic Result
T6a ⫺44.60 S 24.21 S 20.50 S
T6b ⫺24.26 S ⫺7.40 S 31.68 S
T6c 1.96 NS ⫺11.59 S 9.64 S
T6d ⫺53.16 S 18.31 S 34.78 S
(b) T6e ⫺4.20 S 2.35 S 1.85 NS
T6f ⫺28.02 S 6.13 S 21.90 S

1960; Congalton, 1991; Foody, 1992). Analysis of Table 3

(c) reveals that differences as low as 1 percent or less can still be
Figure 5. Classification results from all three methods using significant given their context and the large samples used. In
the T6a texture subset. (a) The GLCM approach. (b) The VGM fact, only two pairwise comparisons did not prove signifi-
approach. (c) The FFT approach. cantly different at the 99 percent level of confidence (but would
be at 95 percent).
Given the slightly superior (but mostly significant) success
rate of the GLCM method, the first experiment has failed to dem-
onstrate that the alternative methods proposed can bring any
● Third experiment: separating a mixture of both different and improvement. These results suggest that the GLCM method,
similar textures and testing the capacity of good association to when used properly, is a very efficient approach to simple tex-
test the performance of each method for the classification of a
large number of classes (36) belonging to both very similar and
ture classification problems where the texture characteristics
very different geographic realms. are visually very different.
The second best method, the VGM approach, shows slightly
poorer results but the difference in Khat (4 percent at the most
Results and Discussion and about 1 percent on average) can be considered small but
First Experiment: Separating a Set of Very Different Textures important, given the fact that the differences are significant
The first experiment has been organized in the following man- and apply to four subsets out of six.
ner: first, the gray-level dependency matrix method (GLCM) The same reasoning can be applied to the FFT results which
was applied, then the variogram method (VGM), and finally the are systematically the lowest for both kinds of result (overall
Fourier-based approach (FFT). In order to give more reliability and no edges). The fact that, in the frequency domain, the high
to the results obtained, the classification tests were carried out frequencies (which correspond to small lags or sampling dis-
on all six similar (but different) sets of six texture patches each. tances in the other methods) are usually regarded as noise can
Figure 5 shows the graphical results of the best results obtained perhaps explain the poorer performance of the FFT method
with the T6a set for each method while Table 2 shows the tabu- which is typically used to separate signal (larger frequencies)
lar results of six sets of textures samples. from noise.
The results clearly show that all three methods of texture In all three methods and for all six texture sets, edges
processing approaches have very good potential for classifica- account for about 20 percent of misclassification errors on
tion purposes because, in all cases, the Kappa statistics ob- average. This is a very significant number which tends to con-
tained are mostly over 70 percent (one exception), and over 90 firm the fact that texture boundaries have a major role in classifi-
percent if edges are not computed. Table 3 presents a pairwise cation errors (Ferro and Warner, 2002) and should perhaps be
test of significance between the three Khat’s obtained for each extracted prior to classification as suggested by various
set. The test uses the normal deviate (Z statistics) to determine authors (Pavlidis and Liow, 1990; Pratt, 1991; Jones, 1994). This
if the confusion matrices are significantly different (Cohen, also outlines the role of patch size, hence resolution: in a texture
patch of 128 by 128, considering a texture analysis window of
32 by 32, the zone of influence of the texture edge is equal to 1282
TABLE 2. CLASSIFICATION SCORES (KAPPA STATISTICS) FROM THE
⫺ [128 ⫺ 32]2 or 7168 pixels in this case, which is about 44 per-
CLASSIFICATION OF SIX TEXTURE SETS; WHITE COLUMNS SHOW OVERALL cent of the total number of pixels of the patch. This percentage
RESULTS WHILE GRAY COLUMNS SHOW THE RESULTS WITHOUT CONSIDERING THE drops to 24 percent for a 250 by 250 patch, 12 percent for a 500
EDGES. BEST RESULTS ARE IN BOLD AND THE LAST ROW SHOWS THE by 500 patch, and 6 percent for a 1000 by 1000 patch.
AVERAGE SCORE
Second Experiment: Separating Sets of Similar Textures
GLCM Method VGM Method FFT Method
In the second series of experiments, six texture sets composed
Texture K̂ ⫻ 100 K̂ ⫻ 100 K̂ ⫻ 100 of similar texture samples were created (Figure 4) in order to
set Overall No edges Overall No edges Overall No edges test the capacity of the three methods to separate textures
T6a 90.1% 99.8% 86.1% 99.4% 82.2% 93.1%
belonging to the same generic classes. Additionally, this series
T6b 79.2 98.0 80.8 99.3 73.9 95.5 of experiments helps characterize the particular aspects of tex-
T6c 76.0 99.1 78.5 99.9 76.4 97.1 ture that are better taken into account by the three different
T6d 79.8 100.0 75.8 95.5 67.6 95.3 methods.
T6e 75.4 99.4 74.8 98.9 74.4 92.1 Table 4 shows the compiled results for all six texture sets
T6f 76.6 99.3 75.2 99.5 70.1 93.1 for both overall and no edges results. In four out of six texture
Average 79.5 99.3 78.5 98.8 74.0 94.4
sets, the VGM method performed better by yielding superior TSC
scores of ⬇1% (F6), ⬇4% (D6), ⬇2.5% (C6), and ⬇2% (W6), all

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2 003 363

 TABLE 4. CLASSIFICATION SCORES (KAPPA STATISTICS) FROM THE CLASSIFICATION OF THE SIX TEXTURE SETS EACH COMPOSED OF TEXTURE PATCHES BELONGING
TO THE SAME GENERIC CLASS. BEST RESULTS FOR EACH ARE IN BOLD

GLCM Method VGM Method FFT Method

K̂ ⫻ 100 K̂ ⫻ 100 K̂ ⫻ 100
Texture Set Overall No edges Overall No edges Overall No edges
F6 (forest) 77.7% 93.4% 78.6% 93.3% 72.3% 87.0%
R6 (residential) 76.7 93.2 60.1 81.9 46.7 64.3
D6 (desert) 74.1 91.7 77.9 95.4 75.3 87.5
C6 (crops) 92.6 100.0 95.0 100.0 85.4 98.4
S6 (shrub) 83.1 97.3 73.7 92.4 69.4 87.4
W6 (wave) 86.9 98.2 88.8 99.6 84.4 95.3

of which were significant at the 99 percent confidence level. As the other experiments (Table 5). This was predictable to a cer-
for the other two cases (R6 and S6), the GLCM method was supe- tain extent because, by increasing greatly the number of
rior by about 16.6 percent and about 9.4 percent, respectively. classes, the chance of misclassification was also increased.
In all cases, the VGM and the GLCM methods were significantly Both the VGM and GLCM methods produced Khat results of
better than the FFT approach by a margin of 2.5 percent to 30 about 65 percent to 67 percent. The FFT came in last with a score
percent in Khat scores. The results for no edge scores are as approximately 8 to 10 percent lower. If these results are overall
high as 100 percent in some cases. The difference between over- quite similar, the detailed analysis of their graphical counter-
all and no edge results are about 14 percent on average for the part shows that the behavior of each method can be different.
three methods which can be considered high given the sample Figure 6 shows the difference image between the classification
size. results of the three methods and what would be the ideal clas-
The first texture set for which the GLCM proved superior, sification. The most striking difference lies in the size and fre-
R6, is also characterized by relatively low Khat for all three quency of wrongly classified pixels and in the patches they
methods. One conclusion that this brings is that the R6 texture form. In the GLCM results, these patches are relatively large,
set is a poor candidate with ill-defined textures. Another obser- infrequent, and more concentrated around the edges of the tex-
vation is that the fact that the GLCM method includes a broader ture patches; hence, the Khat of 93.8 percent for the no edge
variety of measurements is possibly the reason why it scores bet- result. In the VGM classified image, these patches appear
ter whereas the other methods are more “specialized.” In the smaller on average but more frequent and sometimes give a
second texture set for which the GLCM shows superior results, speckled impression. However, the edges and borders account
S6, the situation is different but still keeps similar elements. for a significant part of errors (a Khat difference of 26.2 percent).
Although the Khat scores are higher (roughly between 65 per- The classification result generated with the FFT feature set
cent and 80 percent), a visual inspection of the individual tex- appears to suffer even more from a salt and pepper look: the
ture patches (see Figure 4) reveals that some of them are not patches are mostly small but very frequent. While about 23 per-
very homogeneous, having sometimes a dual textural charac- cent of misclassified pixels are attributable to edges and bor-
teristic (patches #3, #4, and #6 in particular), which tends to ders, another 20 percent are found within the central parts of
give more weight to cues other than the simple spacing of the texture patches. In all cases the differences between the
apparent objects on the background scene (which is the basis three methods are quite significant, as can be seen in the pair-
for the variogram approach). In this regard, the GLCM method wise test of significance of Table 6.
has a definite superiority over the other two. This suggests that The observations above suggest the following facts:
measurement type diversity can prove an important asset for
textures that are not necessarily blessed with a homogeneous ● The GLCM method scores higher than the other two methods for
visual aspect. complex classification situations,
As for the FFT method, its generally poorer performance ● The VGM method yields comparable scores but the difference
can be attributed to two different facts. The first one is inherent from the GLCM is significant,
to the approach (or its implementation) that was chosen, in- ● The fact that patches of misclassified pixels are generally larger
volving the appending of consecutive lines in a semi two- but less frequent for the GLCM method suggests that the method
dimensional approach instead of the full two-dimensional Fou- is not easily affected by small differences and is spatially more
consistent, and
rier transform. This approach might have created undesired
● The FFT feature sets are more likely to be affected by small
artifacts (for instance, the phase of the frequencies is not variations in textures than the GLCM approach.
respected in this approach). The second one is that, unlike the
real time series for which the Fourier transform was developed, Reclassification of the set of 36 textures. Table 7 presents
spatial frequencies in these texture sets are ill-defined and the Khat results obtained after reclassification of the classified
often require a complex set of sine-like waves to describe results of Table 5 (third experiment), and Figure 7 shows the
square-like shapes (as in the case of residential areas). difference image of the generic classes reclassification.
Although, in the overall classification of the set of 36 textures
Third Experiment: Separating a Mixture of Both Different and Similar Textures
In the first part of this experiment, the whole texture set of Fig-
ure 4 has been classified to assess the capacity for each method
to deal with a complex situation where both different and simi- TABLE 5. KHAT SCORES FROM THE CLASSIFICATION OF THE SET OF 36 TEXTURES.
lar texture samples are mixed. In the second part, a reclassifica- GRAY COLUMNS SHOW NO EDGE RESULTS. BEST RESULTS ARE IN BOLD
tion has been performed to assess the good association capa-
GLCM Method VGM Method FFT Method
bility by observing the nature of the errors of the first phase:
i.e.,whether the wrongly classified pixels were at least within K̂ ⫻ 100 K̂ ⫻ 100 K̂ ⫻ 100
the good generic class or not. Overall No edges Overall No edges Overall No edges
The classification of the set of 36 textures yielded results 67.2% 93.8% 65.4% 91.6% 57.1% 80.3%
that are much poorer than those which had been achieved in

364 April 2 003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 (a) (b) (c)
Figure 6. Difference image for classification results of the 36 textures set (misclassified pixels are shown in white,
correctly classified one in gray). (a) The GLCM approach. (b) The VGM approach. (c) The FFT approach.

TABLE 6. RESULTS OF PAIRWISE COMPARISON OF KHAT VALUES FOR OVERALL gains from 18 percent to 28 percent accuracy when not consid-
RESULTS OF TABLE 5; S ⫽ SIGNIFICANT, NS ⫽ NOT SIGNIFICANT AT THE 99% LEVEL ering exact class membership but, in these cases, it appears to
OF CONFIDENCE be due to high within-class variability marked by variations of
GLCM vs VGM VGM vs FFT FFT vs GLCM the between-trees distances. As for the Desert generic class, a
combination of high within-class variability and low contrast
Z Statistic Result Z Statistic Result Z Statistic Result with relation to the other texture patches might have combined
19.6 S 87.5 S ⫺107.2 S to increase accuracy from 14 percent to 19 percent.

Conclusions
Three methods of texture classification have been tested in this
the GLCM method scored better, it was this approach that bene- paper, two of which have received a novel implementation: the
fited the least from the reclassification into generic classes mov- semi-variogram and the Fourier spectra. Both have been imple-
ing from an all-classes Khat statistic of 67.2 percent to 72.2 mented to be computationally efficient and to relate in some
percent (a difference of 5 percent) compared with the FFT way to psychophysical evidence about human vision (Mail-
method that increased from 57.1 percent to 66.1 percent (a dif- lard, 2001). All three methods have proved to be powerful tools
ference of 9 percent) or even the VGM whose TSC score increased for texture classification, but the gray-level co-occurrence
from 65.1 percent to 71.4 percent (a difference of 6 percent). matrix has shown superior results for dealing with simple situ-
Still, the results tend to show that all three methods cannot reli- ations where the textures are visually easily separable. The
ably be expected to perform good association and that not hav- semi-variogram was, however, slightly superior for distin-
ing proper training areas for all classes can be costly in terms of guishing very similar texture patches, but more extensive test-
classification errors. One could conclude that these methods ing is needed to confirm this. In complex situations (a large
are generally better at separating than associating. number of classes), the VGM and GLCM have performed almost
It is interesting to look at which generic classes have gained equally but with generally poorer results (but better than the
the most out of reclassification because it gives an insight on FFT). Much of this poorer performance can be attributed to bor-
the factors that might affect the texture classification accuracy. ders and edges, which tends to show the importance of using a
The Residential class gains from 31 percent to 50 percent of resolution finer than the “optimal” resolution as given, for
accuracy when accepting misclassified pixels that fall into instance, by a measurement like the local variance. The good
another Residential texture class as correctly classified. This is association test proved the GLCM method slightly superior but
understandable because this generic class stands out from the showed that none of the methods tested can be expected to per-
rest by its contrast and square-like objects. The Shrub class also form good association for classes not accounted for in the train-

TABLE 7. KAPPA STATISTICS OF THE GOOD ASSOCIATION ANALYSIS THROUGH THE RECLASSIFICATION OF THE 36 TEXTURES SET. BEST RESULTS ARE IN BOLD
GLCM Method VGM Method FFT Method
K̂ ⫻ 100 K̂ ⫻ 100 K̂ ⫻ 100
Generic Class correct class correct generic class correct class correct generic class correct class correct generic class
Forest 55.0% 66.0% 48.9% 59.3% 43.2% 58.9%
Residential 65.6 96.9 61.3 93.1 37.4 87.5
Desert 51.3 70.5 52.1 66.5 40.8 54.4
Crops 72.1 75.4 71.7 75.0 63.3 72.8
Shrub 63.9 81.4 54.9 82.5 49.2 76.1
Waves 62.8 67.8 68.2 77.0 64.1 76.6
All classes 67.2 72.2 65.4 71.4 57.1 66.1

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2 003 365

 (a) (b) (c)
Figure 7. Difference image for reclassification of the 36 textures set according to generic classes (misclassified
pixels are shown in white, correctly classified one in gray). (a) The GLCM approach. (b) The VGM approach. (c) The
FFT approach.

ing phase: the methods are better at separating than asso- Congalton, R.G., 1988. A comparison of sampling schemes used in
ciating. generating error matrices for assessing the accuracy of maps gener-
The project can be said to have reached three goals: (1) new ated form remotely sensed data, Photogrammetric Engineering &
Remote Sensing, 54(5):593–600.
implementations of mathematical tools for texture analysis
were realized with success, (2) insights on the behavior of these ———, 1991. A review of assessing the accuracy of classification of
tools for texture analysis has provided some understanding of remotely sensed data, Remote Sensing of Environment, 37:35–46.
the characteristics of texture that are difficult to classify, and (3) Conners, R.W., and C.A. Harlow, 1980. A theoretical comparison of
texture algorithms, IEEE Transactions on Pattern Analysis and
a special use of classification using Bayes’ theorem has proven
Machine Intelligence, PAMI-2(3):204–222.
an effective tool for testing and comparing the performance of
texture analysis methods. Cressie, N., and D.M. Hawkins, 1980. Robust estimation of the vario-
gram, Journal of the International Association for Mathematical
The proposed experimental design for testing and compar- Geology, 12:115–125.
ing the results proved most valuable because it clearly showed
Davis, L.S., S.A. Johns, and J.K. Aggarwal, 1979. Texture analysis using
the importance of context in at least two aspects: (1) number of generalized co-occurrence matrices, IEEE Transactions on Pattern
classes and generic groups and (2) spatial properties such as Analysis and Machine Intelligence, PAMI-1(3):251–259.
sample size and surroundings. Further studies will aim at test- Dikshit, O., 1996. Textural classification for ecological research using
ing these methods with other classifiers or segmentation ATM images, International Journal of Remote Sensing, 17(5):
schemes and in more realistic texture interpretation situations. 887–915.
Dunn, D., W.E. Higgins, and J. Wakely, 1994. Texture segmentation
Acknowledgments using 2-D Gabor elementary functions, IEEE Transactions on Pat-
This work is part of a doctoral thesis concluded in 2001 at the tern Analysis and Machine Intelligence, 16(2):130–149.
University of Queensland and was made possible through the Ferro, C.J.S., and T.A. Warner, 2002. Scale and texture in digital image
financial support of the Brazilian Federal Government (CAPES) classification, Photogrammetric Engineering & Remote Sensing,
and the Universidade Federal de Minas Gerais (UFMG). 68(1):51–63.
Fogel, I., and D. Sagi, 1989. Gabor filters as texture discriminator,
Biological Cybernetics, 61:103–113.
References
Foody, G.M., 1992. On the compensation for chance agreement in
Anys, H., and D.-C. He, 1995. Approche multipolarisation et texturale image classification accuracy assessment, Photogrammetric Engi-
pour la reconnaissance des cultures à l’aide de données radar neering & Remote Sensing, 58(10):1459–1460.
aéroporté, Canadian Journal of Remote Sensing, 21(2):138–157.
Franklin, S.E., and D.R. Peddle, 1989. Spectral texture for improved
Atkinson, P.M., 1995. Regularizing variograms of airborne MSS imag- class discrimination in complex terrain, International Journal of
ery, Canadian Journal of Remote Sensing, 21(3):225–233. Remote Sensing, 10(8):1437–1443.
Bonn, F., and G. Rochon, 1992. Précis de Télédétection: Volume 1, Franklin, S.E., A.J. Maudie, and M.B. Lavigne, 2001. Using co-occur-
Principes et Méthodes, Presses de l’Université du Québec, Qué- rence texture to increase forest structure and species composition
bec, Canada, 485 p. classification accuracy, Photogrammetric Engineering & Remote
Caelli, T., 1982. On discriminating visual textures and images, Percep- Sensing, 67(7):849–855.
tion & Psychophysics, 31(2):149–159. Gonzalez, R.C., and R.C. Woods, 1992. Digital Image Processing, Addi-
Cao, C., and N. Lam, 1997. Understanding the scale and resolution son-Wesley Publishing Company, Reading, Massachusetts, 716 p.
effects in remote sensing and GIS, Scale in Remote Sensing and Goresnic, C., and R.S. Rotman, 1992. Texture classification using the
GIS (D.A. Quattrochi and M.F. Goodchild, editors), CRC Press, cortex transform, CVGIP: Graphical Models and Image Proc-
Lewis Publishers, Boca Raton, Florida, pp. 57–72. essing, 54(4):329–339.
Clark, W.A.V., and P.L. Hosking, 1986. Statistical Methods for Geogra- Haralick, R.M., 1979, Statistical and structural approaches to texture,
phers, John Wiley and Sons, New York, N.Y., 518 p. Proceeding of the IEEE Transactions Systems, Man and Cybernet-
Cocquerez, J.P., and S. Philipp (editors), 1995. Analyse d’Images: fil- ics, 67:786–804.
trage et Segmentation, Masson, Paris, France, 457 p. Haralick, R.M., K. Shanmugan, and I. Dinstein, 1973. Texture feature
Cohen, J., 1960. A coefficient of agreement for nominal scales, Educa- for image classification, IEEE Transactions Systems, Man and
tional and Psychological Measurements, 20(1):37–40. Cybernetics, SMC-3:610–621.

366 April 2 003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

Harvey, L.O., Jr., and M.J. Gervais, 1978. Visual texture perception and Reed, T.R., and du H. Buf, 1993. A review of recent texture segmentation
fourier analysis, Perception & Psychophysics, 24(6):534–542. and feature extraction techniques, CVGIP: Image Understand-
Hay, G.J., K.O. Niemann, and G.F. McLean, 1996. An object-specific ing, 57(3):359–372.
Image-texture analysis of h-resolution forest imagery, Remote Reed, T.R., and H. Wechsler, 1990. Segmentation of textured images
Sensing of Environment, 55:108–122. and Gestalt organization using spatial/spatial-frequency represen-
Jain, A.K., and F. Farrokhnia, 1991. Unsupervised texture segmentation tations, IEEE Transactions on Pattern Analysis and Machine Intel-
using Gabor filters, Pattern Recognition, 24(12):1167–1186. ligence, 12(1):1–12.
James, M., 1985. Classification Algorithms, John Wiley and Sons, Lon- Richard, W., and A. Polit, 1974. Texture matching, Kybernetik, 16:
don, United Kingdom, 211 p. 155–162.
Jones, G., 1994. Image segmentation using texture boundary detection, Rosenfeld, A., 1962. Automatic recognition of basic terrain types from
Pattern Recognition Letters, 15:533–541. aerial photographs, Photogrammetric Engineering, 28(1):
115–132.
Julesz, B., 1965. Texture and visual perception, Scientific American,
212:38–48. Rosenfeld, A., and A.C. Kak, 1982. Digital Picture Processing, Volume
2, Academic Press, New York, N.Y., 351 p.
———, 1981. Textons, the elements of texture perception, and their
interaction, Nature, 290:91–97. Sali, E., and H. Wolfson, 1992. Texture classification in aerial photo-
graphs and satellite data, International Journal of Remote Sens-
Jupp, D.L.B., A.H. Strahler, and C.E. Woodcock, 1989. Autocorrelation
ing, 13(18):3395–3408.
and regularization in digital images, II: Simple image models, IEEE
Transactions on Geoscience and Remote Sensing, 27:247–258. Scheffé, S., 1959. The Analysis of Variance, Wiley, New York, N.Y.,
477 p.
Kittler, J., 1983. Image processing for remote sensing, Philanthropical
Transactions of the Royal Society of London, A309:323–335. Serra, J.P., 1982. Image Analysis and Mathematical Morphology, Aca-
demic Press, London, New York, N.Y., 610 p.
Lark, R.M., 1996. Geostatistical description of texture on an aerial
photograph for discriminating classes of land cover, International Smith, G., 2001. MeasTex: Overview of Framework. Image Texture
Journal of Remote Sensing, 17(11):2115–2133. Database and Test Suite, URL: http://www.cssip.uq.edu.au/staff/
meastex/meastex.html, July 2001 (last accessed 22 April 2002).
Maillard, P., 2001. Texture in High Resolution Digital Images of the
Earth, Ph.D. Thesis, The University of Queensland, Brisbane, St-Onge, B.A., and F. Cavayas, 1995. Estimating forest stand structure
Queensland, Australia, 348 p. from high resolution imagery using the directional variogram,
International Journal of Remote Sensing, 16(11):1999–2021.
Manjunath, B.S., and W.Y. Ma, 1996. Texture features for browsing
and retrieval of image data, IEEE Transactions on Pattern Analysis Strahler, A.H., C.E. Woodcock, and J.A. Smith, 1986. On the nature of
and Machine Intelligence, 18(8):837–842. models in remote sensing, Remote Sensing of Environment,
20:121–139.
MeasTex, 2001. Image Texture Database and Test Suite, URL: http://
www.cssip.uq.edu.au/staff/meastex/meastex.html (last accessed Stromberg, W.D., and T.G. Farr, 1986. A Fourier-based feature extrac-
22 April 2002). tion procedure, IEEE Transactions on Geoscience and Remote
Sensing, GE-24(5):722–731.
Miranda, F.P., J.A. MacDonald, and J.R. Carr, 1992. Application of the
semi-variogram textural classifier (STC) for vegetation discrimina- Wilson, R., and M. Spann, 1988. Image Segmentation and Uncertainty,
tion using SIR-B data of Borneo, International Journal of Remote Research Studies Press, John Wiley & Sons, Letchworth, Hertford-
Sensing, 13(12):2349–2354. shire, England, 180 p.
Nyerges, T.L., 1991. Representing geographical meaning, Map General- Woodcock, C.E., A.H. Strahler, and D.L.B. Jupp, 1988. The use of vario-
ization (B. Buttenfield and R. McMaster, editors) Longman Scien- gram in remote sensing and simulated image, II: Real digital
tific and Technical, Essex, England, pp. 59–85. images. Remote Sensing of Environment, 25:349–379.
Pavlidis, T., and Y.-T. Liow, 1990. Integrating region growing and edge Wu, C.-M., and Y.-C. Chen, 1992. Statistical feature matrix for texture
detection, IEEE Transactions on Pattern Analysis and Machine analysis, CVGIP: Graphical Models and Image Processing, 54(5):
Intelligence, 12(3):225–233. 407–419.
Pratt, W.K., 1991. Digital Image Processing, John Wiley & Sons, New Xia, Z.G., and K.C. Clarke, 1997. Approaches to scaling of geo-spatial
York, N.Y., 698 p. data, Scale in Remote Sensing and GIS (D.A. Quattrochi and M.F.
Goodchild, editors), CRC Press, Lewis Publischers, Boca Raton,
Pratt, W.K., O.D. Faugeras, and A. Gagalowicz, 1978. Visual discrimina- Florida, pp. 309–360.
tion of stochastic texture fields, IEEE Transactions on Systems,
Man and Cybernetics, SMC-8(11):796–804. Zhang, Y., 2001. Texture-integrated classification of urban treed areas
in high-resolution color-infrared imagery, Photogrammetric Engi-
Ramstein, G., and M. Raffy, 1989. Analysis of the structure of radiomet- neering & Remote Sensing, 67(12):1359–1365.
ric remotly-sensed images, International Journal of Remote Sens-
ing, 10(6):1049–1073. (Received 31 December 2001; revised and accepted 09 July 2002)

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING April 2 003 367