Spectral-Spatial Graph Kernel Machines in the Context of Hyper spectral Remote Sensing Image Classification

The concept of spectral-spatial graph kernel machinesis proposed in this paper to expand capabilities of the currently implemented classification of hyper spectral images. The frame work is investigated to handle one of the most important challenges of hyper spectral remotely sensed classification especially when the performance is affected by Hughes phenomenon i.e. the curse of dimensionality will be arises when using traditional procedure to classify hyper spectral data.This novel method is concentrated on spatial graph kernel machines, grouping methods and non-uniform distribution informationfor hyper spectral image classification. The innovation of this work consists in: 1) introducing novel weighted spectral-spatial kernel, 2) computing graph kernel by taking into account the non-uniform distribution of the spatial-spectral information, 3) extending some grouped multivariate analysis methods to nonlinear kernel based version and 4) clarifying simultaneous spectral spatial graph kernels theoretical relationships.From empirical results, we conclude that the novel proposed grouped kernel approach meaningfuly enhances the classification performance, it greatly improves the classification overall and per class accuracies and it also provides classification maps with more homogeneous regions particularly in terms of limited training samples.