FSSLN: A Lightweight Frequency–Spectral–Spatial Cooperative Learning Network for Hyperspectral Image Classification

Abstract

Hyperspectral images (HSIs) play a cruciala role in various remote sensing applications due to their rich feature information. However, existing methods mostly perform feature modeling from a single view, which limits their ability to effectively distinguish subtle differences among land-cover classes. To address this problem, this article proposes a lightweight frequency–spectral–spatial collaborative learning network (FSSLN) for HSI classification. Specifically, a collaborative dimensionality reduction strategy is first introduced to suppress band redundancy and enhance critical interband correlations, thereby producing more compact and discriminative spectral representations. Then, a multiview learning framework is constructed to jointly exploit information from the frequency, spectral, and spatial domains. The spectral–spatial view is used to efficiently capture local and global dependencies, whereas the frequency-domain view is employed to characterize frequency-domain patterns and fine-grained variations. In this way, feature discriminability is further improved. Experimental results on multiple HSI datasets demonstrate that FSSLN achieves superior performance while preserving clear advantages in model compactness and computational efficiency.