A joint spatial and spectral SVM's classification of panchromatic images

The classification of very high resolution panchromatic images from urban areas is addressed. The spectral information, i.e. the gray level of each pixel, does generally not ensure a reliable classification. In this paper, we investigate the use of an area filter to extract information about the inter-pixel dependency. The classification is then performed using a Support Vector Machines (SVM) classifier. Using a linear composition of kernels, we define a kernel using both the spectral (original gray level) and the spatial information. A weighting parameter, controlling the relative importance of each feature, is introduced and tuned during the SVM’s training process. Experiments have been conducted on simulated panchromatic Pleiades data over Toulouse, France. Results obtained with the proposed approach is positively compared to those obtained with the standard use of gray value information only and classical SVM formulation.