Class-Oriented Weighted Kernel Sparse Representation With Region-Level Kernel for Hyperspectral Imagery Classification

As a nonlinear extension of traditional sparse representation-based classifier (SRC), kernel SRC (KSRC) has shown its excellent performance for hyperspectral image (HSI) classification, by mapping the nonlinearly separable samples into high-dimensional feature space. However, the rich locality structure of HSI contains more discriminative information, which should be considered in KSRC. We intend to incorporate the locality structure and kernelmethod into a unified SR-based framework by a local spatial kernel. As a powerful texture descriptor, local binary patterns (LBP) was used to extract local feature for remote sensing. Region-level kernels are applied to calculate the distance between two LBP histogram features. To discover nonlinear similarity information between test and training samples, we integrate the LBP feature into spatial region-level kernel for HSI classification. Then, we propose a weighted kernel sparse representation classifier optimized via class-oriented strategy, which combines local structure information and SRC in the kernel feature space based on spatial region-level kernel. Experimental results on three open HSIs demonstrate that the proposed method achieves better classification performance than other state-of-the-art classification methods.