MULTI-VIEW SELF-REPRESENTATION BASED HYPERSPECTRAL BAND SELECTION

Hyperspectral image (HSI) can provide continuous spectral information in hundreds to thousands of bands, but adjacent bands are highly correlated and some bands may not carry discriminative information. Thus, effective band selection methods aiming to eliminate redundancy and retain key spectral information, have become an important research direction in HSI analysis, among which the self-representation (SR) has become a popular means. However, the existing SR-based methods are generally carried out within a single view, namely the feature space formed by stretching all pixels into a vector, which may not fully capture the information of HSIs in different views, and have a negative impact on the accurate selection of bands. Therefore, this paper proposes a multi-view self-representation (MVSR)based band selection method which conducts SR on multi-view feature space to more comprehensively represent the HSI and further improve band selection performance. Specifically, two views, including the feature space formed by original all pixels and that formed by super-pixels, are used to respectively conduct the SR based band selection, which can represent the HSI using important bands from different spatial-scales. In addition, the SR under two views are enforced to approximate a jointly shared representation coefficient matrix to consistently indicate the key bands in HSIs. Furthermore, an efficiently convergent iterative algorithm is designed to solve the joint consistent self-representation coefficients of MVSR. Experimental results on three benchmark datasets demonstrate that the proposed MVSR outperforms several band selection methods in HSI classification.