QUANTITATIVE TEXTURAL PARAMETER SELECTION FOR RESIDENTIAL
EXTRACTION FROM HIGH-RESOLUTION REMOTELY SENSED IMAGERY
J. Gu 3 ’ b *, J. Chen 3 , Q.M. Zhou c , H.W. Zhang 3 , L. Ma d
a National Geomatics Center of China, Baishengcun, Zizhuyuan, Haidian District, Beijing, China-
julietgujuan@ 163 .com
b Beijing Institute of Surveying and Mapping, 15 Yangfangdianlu, Haidian District, Beijing, China
^Department of Geography, Hongkong Baptist Univertity, Kowloon Tong, Kowloon, Hong Kong, China
d School of Remote Sensing and Information Engineering, Wuhan University, China
KEY WORDS: Texture, Residential Area, JM-Distance, Window Size, Quantization Level, Displacement, Orientation
ABSTRACT:
Residential areas show plenty of texture information on high resolution remotely sensed imagery. Appropriate description about this
texture information for discriminating residential class and its background is a key problem for improving the classification results.
Method for selecting proper texture parameters is presented in this paper. Based on the analysis of residential texture, grey level co
occurrence matrix (GLCM) and edge density (ED) approaches with candidate nine texture measurements (contrast, homogeneity,
dissimilarity entropy, energy, mean, standard deviation, correlation and edge density) is selected as candidate texture measurements.
The texture parameters are selected based on separability measured by Jeffries-Matusita distance (JM distance) between residential
and its background in corresponding texture space. IKONOS panchromatic imagery has been used as example and the optimal
texture parameters were selected by using the proposed method.
1. INTRODUCTION
Texture is an important image feature used in visual
interpretation of residential area from high-resolution remotely
sensed imagery. It is a property that relates to the nature of the
variability of pixel values (Anys and He, 1995) which requires
elaborate prior models that guide procedures for extraction.
Several studies have addressed that the addition of image
texture improves image classification. Karathanassi et al. (2000)
reported that the density of buildings can be discriminated
based on a simple, binary co-occurrence matrix. Myint and Lam
(2005) used lacunarity as a texture measure to improve
traditional spectral based classification accuracy. Gong and
Howarth (1990) applied edge detection and smoothing
techniques to generate a spatial pseudo-spectral ‘road-density’
band to supplement conventional spectral bands in classification.
Zhang et al. (2003) applied a similar approach to study urban
change in Beijing. In this case, line rather than edge detection
was applied since the former conforms to road patterns better.
Despite the fact that texture can be visually discriminated, there
is still no appropriate model for texture. It is more difficult to
quantify texture than spectral information as it involves
measurements of variability, pattern, shape and size (Cobum
and Roberts, 2004). There are two distinct types of methods to
extract texture information from an image, i.e. segmentation-
based and window-based. The segmentation-based methods
firstly segment an image into non-overlapping homogeneous
regions (segments), and then texture values are computed from
these segments. Such methods assume that residential areas are
homogenous and it can be obtained through segmentation
algorithms. However, most real residential regions in high-
resolution remotely sensed imagery do not present
homogeneous features. It is difficult to obtain “pure” residential
regions through segmentation algorithms. Another problem of
segmentation-based method is that texture values are usually
sensitive to the scale, but regions obtained from segmentation
are commonly with different sizes, so that the texture
parameters computed from these regions are not comparable.
The window-based method is the most prevalent technique.
Texture values is calculated from moving a fixed-size, odd-
numbered window through the image. The selection of window
size is important for computing texture parameters. Besides,
since most of texture is computed based on statistics, different
texture measurements also have different parameters needs to
be pre-set.
In this paper, the evaluation and optimization of parameters for
computing textural features to discriminate residential areas
from their background class based on Jeffries-Matusita distance
(JM-distance) is presented. Section 2 will introduce the texture
features and parameter selection method used in this paper.
Section 3 will give an example of the selection of texture
parameters by using IKONOS Panchromatic imagery.
Conclusions will be made in the final section.
2. METHODOLOGY
2.1 Candidate texture measurements for residential areas
The Haralick grey-level co-occurrence matrix (GLCM) is one of
the most popular methods for pixel variation statistics (Conners
and Harlow 1980). It uses a spatial co-occurrence matrix that
computes the relationships of pixel values in a certain window
size and uses these values to compute the second-order
statistical properties from these matrices (Haralick, 1979). Eight
second-order statistics derived from GLCM are mostly used in
remote sensing imagery analysis. They are contrast (CON),
homogeneity (HOMO), dissimilarity (DIS), entropy (ENT),
energy (also called angular second moment, ASM), mean
(MEAN), standard deviation (SD) and correlation (COR).
Details about the GLCM method are available in Haralick et al.
(1979). Equation 1 shows their definitions.