Super-pixel based segmentation of urban area and its detection using machine learning techniques

Abstract

In earth observation field, accurate and timely detection of urban area is requisite for military audit, digital mapping, national construction, and change detection which is a major responsibility of urban region planners and government agencies. First, adaptive gamma normalization is used to normalize any given input image without defining target. Adaptive simple linear iterative clustering along with texture feature is proposed for better segmentation of urban area from forest or land. Then, edge and homogeneous region extraction is performed. The comparison of performance results of Sobel operator with two different edge detection operators reveals improvement in detection. Texture feature obtained from gray-level co-occurrence matrix is used to train support vector machine, and further, the edges, which are present in an image, are clustered and labeled by self-organizing map. Then, this output is applied to support vector machine as an input. Finally, morphological hole filling operation makes buildings more perceptible.