Projecte final de carrera fet en col.laboració amb Ecole Nationale Supérieure d'Electronique et de Radioélectricité de Grenoble i Télécom - ENSIMAG

The classification of optical urban remote‐sensing images is addressed. Support Vector Machines (SVM) are proposed to classify hyperspectral images. An introduction to SVM is given in this report in order to help understand how they classify data according to the spectral information. Some kernel functions which are used to improve classification accuracy are presented as well. Then the use of spatial information through multiresolution decomposition is detailed. The objective of this report is to propose a methodology including the spatial information in the classification process and trying to evaluate and improve the accuracy of this classification. Spatial information is extracted from a wavelet analysis of the image. Finally experimental results are presented for each classification method: spectral, spatial and combining both spatial and spectral, and kernel parameters are selected in order to optimize the classification. After including the spatial information, classification accuracy has been improved.