See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228771219

Texture feature extraction for classification of remote sensing data using

wavelet decomposition: A comparative study

Article · January 2004

CITATIONS READS

120 761

3 authors, including:

Jorge Recio
Universitat Politècnica de València
44 PUBLICATIONS 919 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Characterisation of forest structure by integrated analysis of methods based on LiDAR, terrestrial laser scanning and imagery View project

Analysis and assessment of forest structure parameters from LiDAR and other emergent techniques for modeling fuel potential View project

All content following this page was uploaded by Jorge Recio on 23 September 2014.

The user has requested enhancement of the downloaded file.

 TEXTURE FEATURE EXTRACTION FOR CLASSIFICATION OF REMOTE SENSING
DATA USING WAVELET DECOMPOSITION: A COMPARATIVE STUDY

L. A. Ruiz; A. Fdez-Sarría; J.A. Recio

Dept. of Cartographic Engineering, Geodesy and Photogrammetry. Politechnic University of Valencia.

Camino de Vera s/n 46022-Valencia (Spain) – (laruiz, afernan, jrecio@cgf.upv.es)

KEY WORDS: Texture classification, multiresolution analysis, wavelets, urban, vegetation

ABSTRACT:

The extraction of texture features from high resolution remote sensing imagery provides a complementary source of data for those
applications in which the spectral information is not sufficient for identification or classification of spectrally heterogeneous
landscape units. However, there is a wide range of texture analysis techniques that are used with different criteria for feature
extraction: statistical methods (grey level coocurrence matrix, semivariogram analysis); filter techniques (energy filters, Gabor
filters); or the most recent techniques based on wavelet decomposition. The combination of parameters that optimize a method for a
specific application should be decided when these techniques are used. These parameters include the neighbourhood size, the
distance between pixels, the type of filter or mother wavelet used, the frequency or the standard deviation used to create the Gabor
filters, etc. The combination of parameters and the texture method used is expected to be key in the success and efficiency of these
techniques for a particular application.
In this study, we analyze several texture methods applied to the classification of remote sensing images with different types of
landscapes, as well as the optimal combination of parameters for each group of data. For this purpose, we created a database with
high resolution satellite and aerial images from two types of environments, representing two of the main applications of texture
analysis in remote sensing: Urban and forestry. The texture classes defined in urban applications involve heterogeneity and
symmetry, while in forest applications is important to know the type and density of vegetation. The results show that the type of
application determines the technique and the combination of parameters to be used for optimizing accuracy. The combination of
texture methods and spectral information improves the results of classification. Finally, some specific methods to correct the border
effect should be developped before these techniques can be applied in practice.

1. INTRODUCTION uniformity, rugosity, regularity, etc. A considerable number of

quantitative texture features can be extracted from images using
Multispectral information provided by airborne and satellite different methodologies in order to characterize these
sensors is succesfully used for creating and updating properties, and then can be used to classify pixels following
cartography for forest and agriculture uses, as well as for analogous processes as with spectral classifications.
monitoring urban sprawl. This information is valuable as a Many texture comparative studies can be found in the literature,
complement to the field data and the more traditional manual usually carried out by employing standard image databases for
interpretation of aerial photographs, allowing for an increase in the testing process. However, due to the lack of a widely
the efficiency of the processes by partially automatizing certain accepted benchmark, all experimental results should be
tasks, thus reducing costs of field data collection and improving considered to be applicable only to the reported setup. Using
the updating frequency due to the regularity of quality imagery images from the same database gives no guarantee of obtaining
data. comparable experimental results (Ojala et al., 2002).
In forestry and urban studies, multispectral classification In this article we describe the application of several texture
techniques provide suitable results when the classes defined feature extraction approaches to classify different images from
represent structural and spectral homogeneous units, provided two main environments: forest and urban landscapes. The
that the spectral response pattern of each class is sufficiently fundamental goals of this study were:
specific. This is the case of mountain areas where there are
dense forests with uniform growth and a predominance of one • To compare and evaluate four different approaches for
or few species. However, mediterranean ecosystems present a the extraction of texture features applied to the
wide structural and botanical diversity. A similar situation classification of a variety of images in different
occurs in most of the peripheral urban areas, where there is a environments, analyzing and assessing the different
strong structural diversity and, consequently, an important methodological parameters involved in the process.
spectral variability in the urban landscape units. This makes the • To study the potential of these techniques in order to
process of classification using only spectral information more classify (1) mediterranean forest landscape units with
difficult, and some methods for the extraction of structural different density and types of vegetation, and (2) urban
information from each type of unit are required. sprawl units.
Texture analysis offers interesting possibilities to characterize • To assess the potential sinergy of the combination of
the structural heterogeneity of classes. The texture of an image texture and spectral data from high resolution satellite
is related to the spatial distribution of the intensity values in the images, in order to classify complex landscapes.
image, and as such contains information regarding contrast,
 (Jain and Farrokhnia, 1991). A Gabor filters bank is composed
of a set of Gaussian filters that cover the frequency domain with
2. TEXTURE ANALYSIS METHODS different radial frequencies and orientations. In the spatial
domain, a Gabor filter h(x,y) is a Gaussian function modulated
In this chapter we will briefly describe the four methods used by a sinusoidal function:
for texture analysis and feature extraction: (1) Statistical
methods based on the grey level coocurrence matrix, (2) energy
filters and edgeness factor, (3) Gabor filters, and (4) wavelet h ( x, y ) =
1
⋅ exp[−
(x 2
+ y2 )] ⋅ exp( j 2πF (x cosθ + ysenθ ))
transform based methods. 2πσ 2
g 2σ g2
(3)
2.1 Grey level coocurrence matrix (GLCM)
where σg determines the spatial coverage of the filter. In the
The elements of this matrix, p(i,j), represent the relative frequency domain, the Gabor function is a Gaussian curve
frequency by which two pixels with grey levels "i" and "j", that (Bodnarova et al., 2002). The Fourier transform of the Gabor
are at a distance “d” in a given direction, are in the image or function is:
neighbourhood. It is a symmetrical matrix, and its elements are
expressed by H (u, v) = exp[−2π 2σ g2 ((u − F cosθ ) 2 + (v − Fsenθ ) 2 )] (4)

P(i , j ) (1)
p(i , j ) = Ng − 1Ng − 1
The parameters that define each of the filters are:
∑ ∑ P(i , j )
i =0 j =0
1. The radial frequency (F) where the filter is centered
in the frequency domain.
2. The standard deviation (σ) of the Gaussian curve.
where Ng represents the total number of grey levels. Using this 3. The orientation (θ).
matrix, Haralick (1973) proposed several statistical features
representing texture properties, like contrast, uniformity, mean, For the purpose of simplicity, we assume that the Gaussian
variance, inertia moments, etc. Some of those features were curve is symmetrical. The filter bank was created with 6
calculated, selected and used in this study. orientations (0º, 30º, 60º, 90º, 120º and 150º) and 3
combinations of frequency and standard deviation: F=0.3536
2.2 Energy filters and edgeness and σ =2.865, F=0.1768 and σ =5.73, F=0.0884 and σ
=11.444. This operation produced a total of 18 filters covering
The energy filters (Laws, 1985) were designed to enhance some the map of frequencies. Once the filters were applied and their
textural properties of the images. This method is based on the magnitude computed, the image was convolved by a Gaussian
application of convolutions to the original image, I, using filter (σ =5) to reduce the variance.
different filters g1, g2,...,gN , therefore obtaining N new images
Jn = I * gn (n = 1,...,N). Then, the energy in the neighbourhood 2.4 Wavelet transform
of each pixel is calculated. In order to reduce the error due to
the border effect between different textures, a post-processing The use of wavelet transform was first proposed for texture
method proposed by Hsiao y Sawchuk (1989) was used. This analysis by Mallat (1989). This transform provides a robust
method is based on the calculation, for each pixel of the filtered methodology for texture analysis in different scales. The
image Jn, of the mean and variance of the four square wavelet transform allows for the decomposition of a signal
neighbourhoods in which each pixel is a corner, and assigning using a series of elemental functions called wavelets and
as the final value for that pixel the mean of the neighbourhood scaling, which are created by scalings and translations of a base
with the lowest variance, which is supposed to be more function, known as the mother wavelet:
homogeneous and, consequently, should contain only one type s ∈ ℜ+ u ∈ ℜ
of texture (no borders).
1  x−u
ψ s ,u ( x ) = ψ 
The edgeness factor is a feature that represents the density of s  s 
edges present in a neighbourhood. Thus, the gradient of an
image I is computed as a function of the distance “d” between (5)
neighbour pixels, using the expression:
where “s” governs the scaling and “u” the translation. The
g (i, j, d ) = ∑{| I (i, j ) − I (i + d , j ) | + | I (i, j) − I (i − d , j ) | + (2) wavelet decomposition of a function is obtained by applying
( i , j )∈N
each of the elemental functions or wavelets to the original
+ | I (i, j ) − I (i, j + d ) | + | I (i, j ) − I (i, j − d ) |} function:
(6)
where g(i,j,d) represents the edgeness per unit area surrounding
1 * x − u 
a generic pixel (i,j) (Sutton and Hall, 1972). Wf ( s, u ) = ∫ f ( x) ψ  dx
ℜ s  s 
2.3 Gabor filters
In practice, wavelets are applied as high-pass filters, while
These filters are based on multichannel filtering, which scalings are equal to low-pass filters. As a result of this, the
emulates some characteristics of the human visual system. The wavelet transform decomposes the original image into a series
human visual system decomposes an image formed in the retina of images with different scales, called trends and fluctuations.
into several filtered images, each of them having variations in The former are averaged versions of the original image, and the
intensity within a limited range of frequencies and orientations latter contain the high frequencies at different scales or levels.
Since the most relevant texture information is lost in the The vegetation of this area is mainly composed of forest
lowpass filtering process, only fluctuations are used to calculate (Pinus halepensis) and mediterranean shrub, usually
texture descriptors. If the inverse transform is applied to the mixed, and mountain crops (Amigdalus communis, Olea
fluctuations, three reconstructed images, or details, are europaea, Ceratonia siliqua) sometimes forming flat
obtained: horizontal, vertical and diagonal. This process is terraces on the sides of the mountains. The trees of this
called multiresolution analysis. area are more scattered, in part because of a high
Regarding previous work in image texture analysis using recurrence of wildfires over the last several years. Nine
wavelet decomposition, different texture features have been classes were defined: high-density, mid-density and low-
extracted, sometimes from the fluctuations and in other cases density forest, high-density and low density shrub,
from the details, depending on the authors. Sometimes, basic cereals, almond trees, reforestation areas, and crops on
features directly extracted from the histogram were used, such terraces. The data were digital orthophotos with 1m of
as the local energy (Randen and Husoy, 1999) or variance filter spatial resolution, that were also mosaicked to form an
(Ferro and Warner, 2002). Simard et al. (1999), however, used image with a variety of zones (figure 1).
wavelet histogram signatures, while Van de Wouwer et al.
(1999) compared the energy, wavelet histogram signatures and
coocurrence features.
HD-forest MD-forest
We compared the use of fluctuations and details, and four
coocurrence features were calculated using them: variance,
inverse difference moment, contrast and correlation.
In a comparative study about the evaluation of the performance LD-forest HD-shrub
of texture segmentation algorithms based on wavelets, Fatemi-
Ghomi et al. (1996) stated that the identification of the most
appropriate parameters to use in a method is as important a
decision as the choice of which method to use. We also wanted Reforest. Terraces

to know, given our particular classification cases, the best group

of methodological parameters to solve each particular problem.
The following parameters were tested: the type of features, the
Almond Cereal
window or neighbourhood size, the type of wavelet, the
influence of the level of decomposition, and the use of the sum Figure 1. Orthoimages mosaic of forest area 2, Ayora (left), and
of the details or the fluctuations, or to consider them detail examples of eight of the classes defined (right).
independently. All these items will be analysed in the tests and
results section. 3. Forest 3: Located in the south of Menorca, one of the
Balearic islands in the western Mediterranean sea. The
landscape is composed of small forested areas (Pinus
3. TESTS AND RESULTS halepensis, Quercus ilex), and shrubs (Quercus coccifera,
Ulex, Pistacia lentiscus, Rhamnus alaternus), usually
The different texture analysis methods and parameters were combined with scattered trees (Olea europaea var.
evaluated for application in two environments: mediterranean sylvestris), pasture areas, crops and residential areas.
forested areas and growing urban areas. In this section, we will Seven forest and agricultural classes were defined: dense
first describe the testing areas and the type of image data used, forest, forest-shrub, dense shrub, scattered trees,
then we will analyze the selection of the texture parameters. herbaceous vegetation or weeds, cereal or pasture, and
Finally, we will compare the accuracy of the classification fallow; as well as two non-vegetation classes: residential
obtained with the specific methods used, as well as the spectral areas and sea. In this case, a high-resolution panchromatic
versus texture classification for one of the forest testing areas, satellite image (QuickBird) was used, but resampled to
where Quickbird images were available. 2.4 m to keep visual coherence of the texture classes
analysed, and to be able to compare them with the
3.1 Data and test areas multispectral image from the same satellite.
4. Urban: Located in the northern area of the city of
Imagery from a total of four areas was used for evaluation, Valencia, which has experienced an important urban
three forested and one urban, all in the mediterranean region of sprawl during the last several decades, and the
Spain. surrounding towns. The classes considered were: old
urban areas, new urban areas, more dispersed residential
1. Forest 1: Located at the Sierra de Espadán, Castellón, areas located outside of the city, industrial areas and
near the central mediterranean coast of Spain, with barren soil, horticulture, and citrus fruit orchards. A
dominance of forest (Pinus halepensis and Quercus panchromatic image captured by the satellite QuickBird
suber) and shrubs (Quercus coccifera, Ulex,...), olive tree was used, in this case resampled to 5 m.
crops and rocky areas. Seven classes were defined: high-
density forest, mid-density forest, areas combining forest- 3.2 Selection of methodological parameters
shrub, shrubs, scattered trees, scattered shrubs, and olive
trees. For the purposes of evaluation, a mosaic image was As we stated above, there are several methodological
created from aerial orthophotos scanned to 1m of spatial parameters that should be optimized for each type of
resolution. application (forest or urban). We will now describe the results
2. Forest 2: This area is located slightly south and west of obtained in the parameter selection process, method by method.
the previous one, in Ayora, Valencia, farther from the One of the most relevant parameters is the neighbourhood size,
coast and having a type of climate meso-mediterranean. which is obviously related to the spatial resolution of the
images. Therefore, a specific analysis is required for each of the
88
images with a different resolution.
87

Overall Accuracy (%)

• Coocurrence matrix method: The distance between pixels 86

(from 1 to 3) does not seem to effect on the results, so a 85

distance of one pixel was used. In general, the increase of 84

the window size rises the level of the accuracy in the inner 83

part of the texture areas, but produces a progressive 1+2 1+2+3 1+2+3+4 1+2+3+4+5

Groups of variables
increase in error due to the border effect. A neighbourhood
Daub4 Daub8 Coif12 Coif24
size of 25x25 was used, except for the forest area 3
(Menorca), where a size of 15x15 optimized the accuracy
results. Figure 2. Results for the selection of wavelet type and level of
• Energy filters and edgeness: A common window size of 7 decomposition used for the urban area. (Groups of variables:
pixels was used to apply the filters, while for the post- 1:Original image. 2:Textural variables from original image.
processing operation the window size ranged from 7 to 15 3:Variables from details of 1st level. 4:Variables from details of
pixels, depending on the area. The optimal distance for the 2nd level. 5:Variables from details of 3rd level).
edgeness factor was 3 pixels.
• Gabor filters: The main parameters are the standard
deviation of the filter, what has an interpretation similar to 3.3 Comparison of methods
the window size, and the frequency. After the selection
process, banks of filters with standard deviations of 2.86, The algorithm used in the classification process was the
5.73 and 11.44, and respective frequencies of 0.3536, maximum likelihhod classifier, and two sets of texture samples
0.1768 and 0.0884 were created. They were defined by the were defined for each area: a training set and a testing set, both
six dominant directions and then averaged to eliminate the independent and chosen to be representative of the different
orientation factor. classes considered. After the aforementioned selection of
• Wavelet based method: Four types of wavelet families variables, several combinations of groups of variables were
were tested, Daubechies 4 and 8, and Coiflet 12 and 24, as tested to compare the texture methods. The results of the
well as 3 different levels of fluctuations and details. The different classifications, in terms of overall accuracy, are shown
best results were obtained using the wavelet Coiflet-24 and in figure 3.
its reconstructed details form the 3 levels, because each As expected, due to the spectral heterogeneity of most of the
level provides texture information from a different scale classes, the lower accuracy levels correspond to the only
(figure 2). spectral classification that uses the four multispectral bands of
the QuickBird image (only for the area of Menorca). The
As a result of these preliminary tests, a reduction of the texture accuracy increases by combining different groups of texture
features to be used in the comparative classification process was variables.
made for each of the four methods tried.

MS+Textures 83,3

76,18
GLCM+WV+Gabor+Energy 82,06
81,4
88,41

73,92
GLCM+Energy 75,6
78,03
85,82
71,03
GLCM+Gabor 76,56
76,6
86,45
72,42
FOREST 1
GLCM+WV(Coif24) 75,16
80,05
87,24
FOREST 2
67,44 FOREST 3
Energy 57,35
62,96
70,7

65,7 URBAN
Gabor 61,73
68,4
68,14
53,75
WV(Coif24) 65,47
66,4
79,28
65,97
GLCM(8) 71,17
74,7
84,25

MS 61,9

50 55 60 65 70 75 80 85 90 95

Figure 3. Overall accuracy percentages obtained for the four test areas using different methods and combinations of texture variables.
Considering the different texture methods independently, it
cannot be stated that there is a universal method that is best for Class Producer’s User’s
all cases, since the results seem to depend on the type of accuracy accuracy
problem treated. However, they are usually better when Citrus orchards 80.07 86.18
statistical coocurrence features are used. The combination of New Urban 88.09 92.23
these statistical variables with any of the other methods, energy Horticulture 86.38 87.66
filters, Gabor filters or wavelets, produce a significant increase Old Urban 89.04 94.70
in the overall accuracy levels, especially with the latter. This is
Residential 96.70 98.83
problably due to the complementary condition of the methods
Industrial 94.10 65.99
based on filtering with respect to the direct statistical method
based on the GLCM. It is interesting to note that using only
three Gabor filters (three features) it is possible to obtain Table 1. Accuracy percentages of the classification of the urban
relatively good classification results. area using all the texture features (4 methods) combined.

In the forest areas, the texture classification provides accurate

results in those classes where there are mixed spectral
responses, such as reforestation, and where the density of
vegetation is a crucial factor, such as high, mid and low-density
forest. Some examples are shown in figure 4.

Figure 5. Texture classification of a detail image of the urban

area.

3.4 Spectral vs. texture classification

The classification of forest area 3 (Menorca) was done in two

steps. In the first step, the non-vegetation classes (residential
and sea) were masked out by taking advantage of the spectral
and radiometric properties of the QuickBird multispectral
Figure 4. Detail images of texture classification of mixed areas
image. The sea was masked by directly thresholding the
with reforested and mid-density forest (above); and three
infrared band, and the residential areas (also including roads
different levels of forest density (below).
and cliffs) were extracted by thresholding the third principal
component of the four bands. This is easily achieved using
Regarding the urban application, there are some classes that are
these images, due to their high radiometric resolution (11 bits).
accurately classified using texture methods, such as residential
Once the two masks had been applied over the panchromatic
areas and old urban areas, but there are many commission errors
image, the second step consisted of the vegetation classification
(34%) in the industrial class. It is difficult to create a
of the remaining areas. In addition, this comparative process of
representative texture signature of this area, probably because
classification was carried out using both texture and spectral
the spatial resolution used is not aproppiate for this class. Table
bands. Table 2 shows the comparative results in terms of
1 shows the specific accuracy levels for the different urban
producer’s and user’s accuracies.
classes, and figure 5 a detail of the classified image.

MULTISPECTRAL TEXTURES MS+TEXTURES

CLASS Producer’s User’s Producer’s User’s Producer’s User’s
Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy
Dense forest 54.11 57.02 58.00 78.38 53.67 82.92
Shrubs 62.26 55.39 88.46 94.76 88.90 94.21
Pasture-cereal 99.78 99.74 92.07 92.11 96.76 96.21
Scattered trees 41.96 42.25 85.27 75.22 87.00 75.67
Forest-shrub 21.45 25.56 73.71 54.79 76.10 52.47
Weeds 61.80 58.56 90.36 89.70 94.78 95.65
Fallow 98.69 97.42 87.34 91.02 97.13 100

Table 2. Results of the classification of Menorca using spectral variables, texture variables and a combination of both.
 Comparing the spectral and texture classifications in table 2, we important errors in the transition areas between texture
see that spectral classification is better suited for those units. Further work should be done to reduce this effect.
landscape units with a specific spectral response pattern and
well differenciated from the rest of the units, such as pasture REFERENCES
land and cereal crops, or fallow. The distribution of grey levels
in these two classes is very homogeneous, so they are more Fatemi-Ghomi, N., Palmer, P.L., Petrou, M., 1996. Performance
difficult to discriminate by texture methods. On the other hand, of texture segmentation algorithms based on wavelets.
texture techniques are very efficient in classifying lanscape Technical Report. Electronic and Electrical Engineering
units that contain a high spectral heterogeneity, such as Department. University of Surrey.
scattered trees, forest-shrub and dense shrub. These classes are
not very accurate when classified using only spectral band. Ferro, C.J. and Warner, T.A., 2002. Scale and texture in digital
Another interesting aspect is the the integration of spectral and image classification. Photogrammetric Engineering and Remote
texture bands for classification has a synergic effect on the Sensing, 68(1), pp. 51-63.
results, in some cases even improving the accuracy of both
Haralick, R.M., K Shanmugam and Dinstein, 1973. Texture
groups of classes.
features for image classification. IEEE Transactions on
However, it is important to note that the reported results refer to
Systems, Man. and Cibernetics. SMC-3 : pp. 610-622.
the inner areas of the texture units and not to the borders
between textures. In these areas, the border effect decreases the Hsiao, J.Y. and A.A. Sawchuk, 1989. Unsupervised image
overall accuracy to 47%. An example of this effect is shown on segmentation using feature smoothing and probabilisatic
the detail image of figure 6. Some previous tests have shown relaxation techniques. Computer Vision, Graphics and Image
how the post-processing operation, described for the energy Processing, vol. 48, pp. 1-21.
filters, increases the accuracy in the border areas in a 27% (Ruiz
et al., 2001) Laws, K.I., 1985. Goal-directed texture image segmentation.
Applications of Artificial Intelligence II, SPIE (548), pp.19-26.

Mallat, S.G., 1989. A theory of multirresolution signal

decomposition: The wavelet representation. IEEE Trans. Patt.
Anal. Machine Intell., 11(7), pp. 674-693.
Figure 6. Example of the border effect in texture classification. Ojala T, Mäenpää T, Pietikäinen M, Viertola J, Kyllönen J &
Huovinen S, 2002 Outex - New framework for empirical
4. CONCLUSIONS evaluation of texture analysis algorithms. Proc. 16th
International Conference on Pattern Recognition, Quebec,
This study has been focused on two main applications of texture
Canada, 1: pp. 701-706.
analysis in remote sensing: the classification of forest landscape
units and urban areas. The former holds a special interest for Randen, T. and Husoy, J.H., 1999. Filtering for texture
mapping forest areas, the latter is a first step in monitoring classification: A comparative study. IEEE Trans. Patt. Anal.
urban sprawl. Some important aspects can be concluded: Machine Intell., 21(4), pp. 291-310.
- The texture methods provide an alternative to the spectral Ruiz, L.A., Acosta, P., Fdez.-Sarría, A., Porres, M.J., Pardo,
methods for the classification of forest units with a high J.E., 2001. Mejora de la eficiencia en la segmentación de
spectral heterogeneity, or when the classes are defined by imágenes por texturas mediante filtros de energía. Proc. of the
differences in vegetation density. IX Spanish Congress on Remote Sensing, Lleida, pp. 477-480.
- In urban classification, the texture methods are useful for
discriminating old urban areas and new residential spots, Sutton, R.N. and E.L. Hall, 1972. Texture measures for
but they introduce important errors in the classification of automatic classification of pulmonary disease. IEEE
industrial areas, so spectral information should be used in Transactions on Computers, vol. C-21, pp. 667-676.
addition to texture.
- A universal criteria in order to use the idoneous texture Unser, M., 1995. Texture classification and segmentation using
extraction method for classification does not seem to exist. wavelets frames. IEEE Trans. Image Processing, 4(11), pp.
Therefore, the selection should be in funtion of the type of 1549-1560.
landscape units defined in each application.
- Furthermore, the combination of different texture methods Van de Vower, G., Scheunders, P., Van Dyck, D., 1999.
improves the classification results, especially when Statistical Texture Characterization from Discrete Wavelet
combining statistical methods based on the GLCM with the Representations. IEEE Trans. on Image Processing, 8(4), pp.
details of different levels obtained from the wavelet 592-598.
transform. The Gabor filters allow an important part of the
texture information to be condensed into a few variables. ACKNOWLEDGMENTS
- Before beginning the texture classification process, it is
The authors wish to thank the financial support provided by the
important to previously select the methodological
Spanish Ministry of Science and Technology and the FEDER
parameters and features to reduce the volume of data and to
(projects REN2003-04998 and BTE2002-04552), as well as to
optimize the discrimination power of these techniques.
the Politechnic University of Valencia (project 2002-0627).
- The main limitation for the standard application of texture
methods in image classification is probably the border
effect, inherent to texture analysis and which introduces

View publication stats