Support vector machines for hyperspectral image classification with spectral-based kernels

Support Vector Machines (SVM) have been recently used with success for the classification of hyperspectral images. This method appears to be a robust alternative for pattern recognition with hyperspectral data: since the method is based on a geometric point of view, no statistical estimation has to be achieved. Then, SVM outperforms classical supervised classification algorithms such as the maximum likelihood when the number of spectral bands increases or when the number of training samples remains limited. Nevertheless, those kernel-based methods do not take into consideration the bio-physical meaning of the spectral signature. In fact, kernels are functions that are based on the quadratic distance between support vectors. Then, some modified kernels are presented to take into consideration the spectral similarity between support vectors to outperform SVM-based classification of hyperspectral data cube. Those kernels (that still suit Mercer's conditions) are based on the use of spectral angle to evaluate the distance between support vectors. Classifiers to compare have been applied to an image from the CASI sensor including 17 bands from 450 to 950nm representing an intensive argicultural region (Brittany, France). It appears that those kernels reduces false alarms that were induced by illumination effects with classical kernels.