Comparison of Kernel Support Vector Machine Method Using Aerial Photographs to Identify Slum Settlements in Squatter Area of South Magelang

Abstract

Detection of slums in urban informal settlements plays an important role in making the right decisions and ensuring equitable urban growth in highly developing cities. This paper presents a comparative study of four types of SVM kernel functions, namely, Linear, RBF, Polynomial, and Sigmoid for classifying slum areas based on texture feature using Gray Level Co-occurrence Matrix (GLCM) from very high resolution aerial images. The study is performed with a sample squatter area in South Magelang, Magelang City, Indonesia, and compares the performance of three essential GLCM features (mean, variance, and dissimilarity) on different window sizes. As can be seen, there is a considerable impact of kernel type selection in classification performance. The best performance in terms of overall accuracy was obtained by using RBF kernel with a relatively low accuracy of 62.26%, then comes Polynomial (57.78%), Linear (55.33%), and Sigmoid (49.72%). The above hierarchy of performance shows the fundamental non-linearity that exists between the morphology of slums and the textural representation of slums in the RGB image. Majority voting aggregation of the output of texture feature extraction for each kernel resulted in more coherent settlement maps. In summary, this study offers empirical evidence to guide the optimal use of SVM-based texture features in slum mapping, providing a screening procedure for the automated analysis of urban morphologies via remote sensing.