Spatial context driven manifold learning for hyperspectral image classification

Manifold learning techniques have been demonstrated to be successful in representing spectral signatures in hyperspectral images, which consist of spectral features with very subtle differences and often spatially induced disjoint classes whose neighborhood relations are difficult to capture using traditional graph based embedding techniques. Robust parameter estimation is a challenge in traditional kernel functions that compute neighborhood graphs e.g finding the optimal number of nearest neighbors. We address these challenges by proposing spatial context driven manifold learning methods. Empirically, the study reveals that use of spatial contextual information has a bearing on the structure of the graph Laplacian that in turn links image pixel observations to their manifold spaces. Further experimental results demonstrate an improvement in the classification performance compared to traditional manifold learning methods.