Multi-Stage Deep Sequential Fusion and Dynamic Spectral–Spatial Interaction for HSI and LiDAR Joint Classification

To address the limitations of existing Land Cover Classification (LCC) methods in handling hyperspectral and LiDAR data—such as information loss, data redundancy, noise introduction, and insufficient fusion, this paper proposes a novel deep learning network named Spectral-Spatial-Fourier Cross Mamba (SSFCMamba). Unlike existing approaches, the core innovations of this work lie in: first, the introduction of an Orthogonal Spectral Frequency Hybrid Feature Module, which overcomes the redundancy caused by traditional spectral-spatial feature coupling through deep processing in both spatial and frequency domains, enabling efficient fusion of global and local spectral features. Second, a Multi-Modality Multi-Scale Spatial Feature Interaction Mechanism is proposed, which can adaptively focus on key spatial contextual information while effectively suppressing noise. Furthermore, by fusing the two modalities, the network generates a more discriminative joint representation. Experimental results on the Houston 2013 and Trento datasets show that SSFCMamba outperforms existing methods, achieving Overall Accuracy of 90.37% and 98.79%, respectively, thereby validating its effectiveness in LCC. We release the code in https://github.com/LBY-1895/Spectral-Spatial-Fourier-Cross-Mamba.