39
iamage level
ge levels are
ent with the
MULTI-SCALE SAR LAND USE/LAND COVER CLASSIFICATION BASED ON
CO-OCCURRENCE PROBABILITIES
e acquisition
cessary data
us make this
inning. The
>y additional
for a region
rages, which
[imitation to
nay play an
.fter natural
Yu ZENG a ’ b ’ c , Jixian ZHANG a , J.L.VAN GENDEREN c , Haitao LI a
a Chinese Academy of Surveying and Mapping, Beijing, 100039, P.R.China, zengyu@casm.ac.cn
b College of Geo-Information Science and Engineering, Shandong University of Science and Technology, Qingdao,
Shandong Province, 266510, P.R.China
c International Institute for Aerospace Survey and Earth Sciences (ITC),
P.O. Box 6, 7500 AA Enschede, The Netherlands
KEY WORDS: SAR, Land Use, Land Cover, Classification, Multi-scale, Texture Analysis, GLCM, GLCPs
piake of 26
ABSTRACT
Texture analysis plays an important role in the automated or semi-automated understanding and interpretation of digital imagery.
Grey-level co-occurrence probabilities (GLCPs) appear to be the most commonly used and are the most predominant method for
:g-
image texture analysis and presentation. This paper presents a multi-scale texture analysis method based on co-occurrence
probabilities, which is used for land use/land cover classification of single band and single polarization SAR imagery. In addition, it
introduces a geo-statistical model to estimate ground object scales which can be further used as the suboptimal selection for the
, 1993. The
-e, Vol. 364,
processing window size during the texture feature extraction. Experimental results prove the effectiveness of the proposed method.
1. INTRODUCTION 2. METHOD
l. Fault slip
California,
and GPS
f of America,
Synthetic Aperture Radar (SAR) is a powerful tool for resource 2.1 Grey Level Co-occurrence Probabilities
management and environmental monitoring applications. The
classification of SAR imagery is a crucial step for the The Grey Level Co-occurrence Probabilities (GLCPs) method,
understanding and interpretation of SAR imagery. Texture, a also known as Grey Level co-occurrence Matrix (GLCM), is a
representation of the spatial relationship of grey-levels in an second-order statistical texture analysis approach originally
image, is an important characteristic for the automated or semi- proposed by Haralick et al. (1973). Given a spatial window
iggered slip:
) earthquake
Letters, Vol.
automated interpretation of digital imagery, especially for the within the image, the co-occurrence method finds the
single band and single polarization SAR imagery. conditional joint probabilities, p(i j), of all pairwise
combinations of grey levels given the inter-pixel displacement
ing and GIS
Workshop on
Thailand,
Different methods have been proposed for the analysis of image vector (8 X , 8 y ), which represents the separation of the pixel pairs
texture, which can be categorized into four groups (Tuceryan in the x- and y-directions respectively. The set of grey-level co-
and Jain, 1993): statistical (such as grey-level co-occurrence occurrence probabilities (GLCPs) can be defined as:
probabilities (GLCPs) (Haralick et al., 1973) and grey-level run Py (l)
length), geometrical (including structural methods, and are AUI-yyp
applicable to artificial textures), model-based (such as markov ¡j ' J
a). Website:
1 (accessed
random fields (MRFs) (Manjunath and Chellappa, 1991; Deng where Py represents the frequency of occurrence between two
and Clausi, 2005) and Fractal features (Chaudhuri B B and g re y i eve ] S) / and for a given displacement vector (5 X , 8 y ), for
Sarkar N, 1995)) and signal processing (such as Gabor filters the specified window size. G is the number of quantized grey-
(Clausi and Jemigan, 2000), FFT, and Wavelet (Pichler et al., levels. Traditionally, the probabilities are stored in a grey level
1996, Pun and Lee, 2003)). Each method has its own pros and co-occurrence matrix (GLCM), where index (i,j) in the matrix
^correlation
actions on
950-959
cons, and there is no general agreement on an overall best represents the probability p(i,j). Statistics are applied to the
analysis method, which outperforms all the others on various GLCM to generate texture features which are assigned to the
tasks. Among these methods, the GLCP-based methods appear center pixel of the image window,
to be the most commonly used (Franklin, 2001) and are the
most predominant (Clausi, 2000; Clausi and Yue, 2004). Seven textural features are considered the most relevant among
ire Science
tion project
ivince. The
i>r providing
. the fourteen originally proposed by Haralick et al. and are used
e objective of this paper is to develop a multi-scale texture ¡ n our g^dy- homogeneity, angular second moment, standard
analysis method based on co-occurrence probabilities and used deviation, contrast, dissimilarity, entropy, and correlation. The
for land use/land cover classification of single band and single following equations define these features,
polarization SAR imagery. The rest of the paper is organized as
follows. Section 2 gives an overview on the methodology of Homogeneity (H OMO):
this research. Section 3 describes the experimental framework, j
results and their analysis and conclusions are drawn in Section V V —/?(/, /) №
4. ryi + (i-J) 2
