Hierarchical sparse representation for dictionary-based classification of hyperspectral images

The recent advances in sparse coding and dictionary learning have shown extremely good performances and robustness in high-dimensional classification problems. Most often, dictionary-based methods rely either on the reconstruction power of the dictionary or on the structure of the sparse representation. In this paper we jointly exploit the discrimination power of both approaches by combining the reconstruction error with the hierarchical information of the sparse codes collected during the learning stage. The proposed method performs similarly to state-of-the-art classifiers and outperforms them sharply in small sample situations, where the number of patterns used to learn the dictionaries is much smaller than the number of dimensions.