Hyperspectral Image Classification Based on a Novel LDSLTSR Method

Abstract

Hyperspectral image classification is a research hotspot in remote sensing image analysis and application, and its high dimensionality and few samples make it a challenging problem. In this paper, we propose to use linear dynamic system model combined with linear transformation and sparse representation to achieve hyperspectral image classification. First, we perform image preprocessing based on the intuition that the neighboring pixel of a pixel likely share similar spectral characteristics. Second, we establish the linear dynamic system model of each sub-cube image extracted from the image, and at the same time, introduce linear transformation and sparse representation principles into the established model, and propose a novel linear dynamic system combining linear transformation and sparse representation (LDSLTSR). Finally, we use the error between the linear transformation of the model and the sparse representation to achieve classification. The spatial and spectral information of the image is fully considered in our method, which is conducive to the classification effect. The experiments are conducted on public AVIRIS data and Pavia University data to demonstrate the effectiveness of the proposed method. The results show that the proposed LDSLTSR method effectively outperforms other current state-of-the-art methods.