Hyperspectral remote sensing image classification based on domain-level complementarity of spatial-spectral component.

Abstract

Hyperspectral image (HSI) generally suffers from "homogeneity" and "heterogeneity" issues, leading to low intra-class consistency, poor inter-class differentiation, and higher misclassification rates in feature classification. Effectively integrating global structure and local details of spatial-spectral information is key to addressing this. Existing methods struggle to achieve this integration, which consequently impacts the classification performance. In this paper, we propose a hyperspectral remote sensing image classification method based on the domain-level complementarity of the spatial-spectral component (D²FuPro). This method utilizes a dual-branching structure to obtain information on the global overall structure and local details in the panoramic and mesoscopic domains, respectively. This dual-domain spatial-spectral complementarity aims to alleviate the above issues. A panoramic domain low-rank feature extraction method (PLFE) is designed in the first branch of the D²FuPro method. This method preserves the global spatial-spectral structure by low-rank modeling of the HSI. It combines texture smoothing to suppress noise and spectral anomalies and enhances intra-class spectral consistency. The second branch designs a method of mesoscopic domain feature enhancement (MDFE), which acquires the spectral information of features within the local range of HSI and utilizes the spatial structure of neighboring pixels to model, optimize the spatial boundaries, and enhance inter-class differentiation. Finally, fusing dual-domain information from both branches enables the realization of complementary spatial-spectral information in the panoramic and mesoscopic domains. A comparative experimental validation and analysis on four commonly used classical hyperspectral datasets demonstrates that the D²FuPro method outperforms traditional and twelve more advanced classification methods in terms of classification accuracy.