Satellite Image Classification Using Cascaded Architecture of Neural Fuzzy Network - Advanced Techniques in Knowledge Discovery and Data Mining - page 213

8.2 Input Acquisition

The objective of this section is to clarify how the input features are chosen and
what they are actually for. These input features are gray values, statistical features,
and spectral features. The varieties of gray values among channels are characteris-
tics that can be used to classify the ground cover type from data of multispectral
spaces into desired clusters. However, due to the complexities, mix-up samples,
and textural problems, we have to use statistical features from the co-occurrence
matrix and spectral features from wavelet decomposition to improve classification
accuracy. The following sections show how our input acquisition is implemented
and modified to get a better performance.

8.2.1 Features from the Spatial Domain by Multispectral Data

Taking advantage of multispectral data obtained from different sensors, we can
classify the data as required. Here we choose the gray values as our spatial fea-
tures. Because the variation of gray values among channels can serve as discrimi-
nating features, they can be used to classify the ground cover type in these multis-
pectral spaces into desired clusters. In details, the groups or clusters of pixel points
are referred to as information classes because they are the actual classes of data
that a computer can recognize.

8.2.2 Features from the Statistical Domain

8.2.2.1 Co-occurrence Matrix

Gray-level co-occurrence matrices constitute one of the basic approaches to the
statistical analysis of texture. By computing a set of gray-tone spatial-dependence
probability-distribution matrices for a given image block and extracting a set of
textural features from each of these matrices, the basic model attempts to take the
variation as a function of the direction of spatial distance.
Here is the second-order histogram: ˦˹ ˼ ˽˷

ʻʿʿʿ T , where i , j are gray val-
ues of pixels distance d apart and T is the angle (usually every

$

45 ) of the line
joining the centers of these pixels with the horizontal axis. These matrices are
symmetric: ˣ

. From this
matrix, a number of features can be computed. With G gray levels in the image,
the dimension of the matrix will be ˚ u . The ( i , j )th element, ˣ ˼˽ , of this matrix
is defined by

, that is, ˣ˼˽˷ ˣ˽˼˷

ʻʿʿʿʼ ʻʿʿʿʼ

T T

ʻʼ ʻ ʼ T ˣ T S