Adaptive pixel neighborhood definition for the classification of hyperspectral images with support vector machines and composite kernel

Abstract

The pixel-wise classification of hyperspectral images with a reduced training set is addressed. The joint use of the spectral and the spatial information is investigated. The spectral information simply consists of the spectral value of each pixel. For the spatial information, we use an area filter to simplify the image and extract consistent connected components. These components are used to define an adaptive neighborhood for each pixel of the image. The vector median value of each component is defined as a spatial feature for the classification. support vector machines are used for the classification and a composite kernel is used to combine both the spatial and the spectral information. Experiments are conducted on AVIRIS hyperspectral data. The proposed approach provides significant improvements in terms of classification accuracy when compared with a standard statistical method (maximum likelihood) and with a SVM classifier using the spectral information alone. Robustness with respect to the size of the training set is also investigated.