Improving Hyperspectral Image Classification using Graph Wavelets

We present a novel feature extraction method to improve classification of hyperspectral image, leveraging graph wavelet transform to address the shortcomings of classical wavelet transform that only works on a regularly 2-D or 3-D grid. In the proposed method, we first design an appropriate underlying graph connecting pixels with weights that reflect the hyperspectral image structure; secondly interpret the hyperspectral image as a signal on graph, and apply graph wavelet transform for processing hyperspectral images in graph spectral domain; finally, the transformed hyperspectral images are used to constitute the pixel features, and input these features to a classifier to get the class label. Graph can represent the spectral similarity and nonlocal spatial relation, moreover graph wavelets are sparse and localized in vertex/pixel domain, offering spectral-spatial intrinsic and discriminative features. The experimental results show the competitive performance compared with classical wavelet features.