S2CMamba: Structure-to-Contrast Semantic Mamba for Hyperspectral and LiDAR Joint Classification

Hyperspectral imaging (HSI) and light detection and ranging (LiDAR) provide complementary spectral and elevation information, making their fusion an effective solution for multimodal remote sensing (RS) classification. However, existing fusion methods often fail to adequately capture complex spatial topology and long-range contextual dependencies. Moreover, foreground–background semantic entanglement and the lack of semantic-guided fusion mechanisms further degrade feature discrimination. To address these challenges, a novel structure-to-contrast semantic Mamba (S2CMamba) network is proposed for the joint classification of HSI and LiDAR data. Specifically, a structure-aware state space modeling (S2SSM) module is designed for both HSI and LiDAR branches to encode multi-directional spatial structures and capture long-range global spatial dependencies within each modality. Then, the proposed semantic decoupling unit (SDU) explicitly separates foreground and background semantic responses in the fused HSI–LiDAR representations, which effectively suppresses background interference and alleviates semantic ambiguity during multimodal feature fusion. Based on the decoupled semantic priors, a contrast-guided feature aggregation module (CGAM) is developed to enable contrast-modulated local window aggregation over the fused HSI–LiDAR features, thereby improving intraclass compactness and enhancing class-boundary discrimination. Through experiments conducted on three commonly used HSI and LiDAR datasets, the performance of S2CMamba was consistently demonstrated to surpass that of state-of-the-art (SOTA) methods for HSI and LiDAR.