Spectral–Spatial Feature Learning Using Cluster-Based Group Sparse Coding for Hyperspectral Image Classification

This paper presents a new spectral–spatial feature learning method for hyperspectral image classification, which integrates spectral and spatial information into group sparse coding (GSC) via clusters, each of which is an adaptive spatial partition of pixels. The clusters derived from the segmentation maps by the mean-shift algorithm are regarded as groups in GSC, where pixels within the same group are simultaneously represented by a sparse linear combination of a few common atoms in a given dictionary, thus enforcing spatial smoothness across the pixels in the same segmentation region to learn a spectral–spatial joint sparse representation. Finally, the recovered sparse representation can be viewed as a new feature and used directly for classification (e.g., by support vector machine). In comparison with other spectral–spatial classification techniques that exploit a fixed neighborhood system and force neighboring pixels to share a common sparsity pattern, the proposed method is more flexible and able to obtain adaptive spatial neighborhood correlations for spectral–spatial joint sparse coding. In addition, we also develop kernel GSC (KGSC) by incorporating the kernel trick into GSC to capture nonlinear relationships. The developed KGSC can also be applied to learning kernel sparse representation under the framework of the proposed spectral–spatial method, leading to a new spectral–spatial kernel sparse representation algorithm. Experimental results on three real hyperspectral datasets indicate that the proposed methods improve classification accuracy and provide distinctive classification maps, especially at small regions and boundaries in an image, compared with other similar approaches.