Cluster structured sparse representation for high resolution satellite image classification

Sparse Representation based model has achieved great success for image classification. The classical approach represents each visual descriptor as a sparse weighted combination of codebook words. While offering a sparse and robust representation for each single descriptor, this method however does not ensure that similar descriptors lead to similar representations. In this paper, we present a cluster structured sparse coding (CSSC) method by unifying sparse coding and structural clustering. This approach can encourage using the same codebook words for all similar descriptors in a group, providing a discriminative representation for the task of image classification. We evaluate our method on a challenging ground truth image dataset of 21 land-use classes manually extracted from high-resolution satellite imagery. Experimental results show that structural sparse representation yields higher accuracies in classification.