ASPRS 2012 Annual Conference Sacramento, California ♦ March 19-23, 2012 PIXEL CLASSIFICATION OF SATELLITE IMAGES USING A NOVEL PAIR WISE KERNEL FUNCTION SVM Biplab Banerjee, Surender Varma, Krishna Mohan Buddhiraju Satellite Image Analysis Lab, IIT Bombay Mumbai-400076 India biplab.banerjee@iitb.ac.in, gsvarma@iitb.ac.in, bkmohan@csre.iitb.ac.in ABSTRACT In this paper we have proposed a symmetric, positive semi definite kernel function for support vector machine classifier. Pixel classification is a form of supervised image segmentation where the actual object classes present in the image are known a priori. In case of satellite image, this prior information plays a huge role to estimate the actual statistics of different land covers. The state of the art kernels have the problem to clearly separate closely spaced data points, as in the case of image pixels of satellite images, where there are no sharp changes between two different regions in terms of the pixel intensity. The proposed kernel has overcome this difficulty with the previous kernels effectively and has good generalization capability. Experimental results establishes the fact when the proposed kernel based SVM has been used for supervised satellite image segmentation purpose. KEYWORDS: Support Vector Machine, Kernel function, Pixel Classification, Image Segmentation. INTRODUCTION Image segmentation is the process of diving a given image into its corresponding components. Each component represents a given region of interest. For intensity images four popular approaches are: threshold techniques, edge-based methods, region-based techniques, and connectivity-preserving relaxation methods. Threshold techniques, which make decisions based on local pixel information, are effective when the intensity levels of the objects fall squarely outside the range of levels in the background. Because spatial information is ignored, however, blurred region boundaries can create problem and the adjacent regions may turn out to form a single region in the output image. Edge-based methods revolve around contour detection. The image edges are obtained first using any first order or second order edge detection techniques. Now arises the problem of edge connectivity or edge linking. Some geometric techniques are used for this purpose to define the boundaries of the image objects. Hough transform has turned out to be an useful method in this aspect. Again the problem of blurring may cause some problem in edge linking. Region based segmentation techniques are based on the notion of similarity measures. The idea is to divide the image into some distinct regions in such a way that pixels within a given region share some common statistical characteristics. Region based techniques can broadly be classified into two different parts, 1. Region growing. 2. Region dividing and merging. In the first case, it is considered that each pixel of the image constitute an individual region. Then the pixels are combined with their neighbours based on some similarity recursively until some stopping criteria is satisfied. Sometimes, this process produces an over segmented image. Hence the similarity measures should be selected intelligently. In the second case, the whole image is considered initially to constitute a single region. Then the image is divided in each step based on some dissimilarity measures. One possible division scheme is called quad tree. This method starts at the root of the tree that represents the whole image. If it is found non-uniform, then it is split into four son-squares, and so on so forth. Conversely, if four son-squares are homogeneous, they can be merged as several connected components. The node in the tree is a segmented node. This process continues recursively until no further splits or merges are possible. A connectivity-preserving relaxation-based segmentation method, usually referred to as the active contour model, was proposed recently. The main idea is to start with some initial boundary shape represented in the form of spline curves, and iteratively modify it by applying various shrink/expansion operations according to some energy function. Although the energy-minimizing model is not new, coupling it with the maintenance of an elastic contour model gives it an interesting new twist. As usual with such methods, getting trapped into a local minimum is a risk against which one must guard.