GPSMamba: Global-Perception Spatial-Semantic Enhanced Mamba for Hyperspectral Image Classification

Abstract

Hyperspectral image classification (HSIC) is a core task in remote sensing. Traditional convolutional neural networks (CNNs) are constrained by limited local receptive fields and struggle to capture long-range dependencies within hyperspectral images (HSIs). Although Mamba-based HSIC models can effectively model long-range dependencies, flattening 2-D HSI patches into 1-D sequences results in spatial information fragmentation, which creates a pervasive problem of lost spatial continuity and directional cues. To address these limitations, this article proposes a novel global-perception spatial-semantic enhanced mamba network (GPSMamba) for HSI feature extraction. Specifically, the dynamic mixed convolution (DMC) module computes affinity matrices between feature tokens and regional centers via global information embedding (GIE), then aggregates these matrices to generate spatially variant dynamic convolution kernels that jointly model the global context and local features of HSI. Considering the intrinsic spatial–spectral structure of HSI, the direction-aware mamba (DAM) module maintains intact spatial layouts via continuous 2-D scanning and direction-aware updates. Meanwhile, the decoupled spectral attention (DSA) module decouples HSI features along spatial and channel axes, producing dedicated spatial attention matrices and channel attention vectors, respectively. The channel branch captures interband spectral dependencies of HSI, while the spatial branch models intraband pixel correlations, which enhances spectral characterization and spatial structure perception of hyperspectral data. Extensive experiments verify that the proposed approach achieves superior performance against state-of-the-art HSIC methods.