Hierarchical Discriminant Feature Learning Methods Based On Sparse Representations For Hyperspectral Image

Hyperspectral image(HSI) classification is a hot topic in the hyperspectral image processing field currently, the research focuses on finding different technical methods to make computers learn images intelligently. To build excellent image representations from raw hyperspectral data will help to improve the performance for HSI classification of land cover. However, a major problem we faced in HSI classification process is that the original feature space may not be the most effective space for representing the data. To solve this problem, three feature learning methods which are based on sparse representation(SR) and spatial pyramid representation(SPR) for HSI classification are proposed. Applying these proposed methods in three hyperspectral datasets which are Indian Pines, Salinas Scene and University of Pavia, we conduct experiments and achieve satisfying classification accuracies. Simultaneously, it also proves the effectiveness of our proposed feature learning methods for HSI classification problem according to experiments result obtained. The main contributions can be summarized as follows:1. We develop a hierarchical discriminant feature learning(HDFL) algorithm for HSI classification which is a deformation of spatial pyramid representation(SPR) based on sparse coding learned from the supervised dictionary in every layer of this model. This new method remarkably makes the pooling features achieved by hierarchical discriminant feature learning more separable. Furthermore, both the spectral and spatial information are considered to improve the classification performance.2. We propose a novel hierarchical discriminant feature learning approach based on scaling cut(SC) criterion for HSI classification. The proposed approach forms the features as a weighted sum of the mid-local features over all of the pyramid levels in every hierarchical layer. The weights which are obtained by using SC are selected to maximize the discriminative power. By using hyperspectral datasets, our approach showed high performance.3. We propose a new method feature learning for HSI classification based on divisive information theoretic feature clustering(DITC). Our method can exploit the compact and effective feature representations from the features achieved by the hierarchicaldiscriminant feature learning approach. Compared with original hierarchical discriminant features, features obtained by our method can reduce the time-consuming and improve the computational efficiency greatly.This work was supported by the National Natural Science Foundation of China(No. 61272282), and the Program for New Century Excellent Talents in University(NCET-13-0948)...