MSFFRN: Multimodal Spatial–Frequency Fusion and Refinement Network for Hyperspectral and LiDAR Data Classification

The fusion of hyperspectral image (HSI) and light detection and ranging (LiDAR) data leverages complementary spectral and elevation information to improve land cover classification, yet its effectiveness is hindered by heterogeneous data distributions and a common oversight of frequency-domain information. To address this, we propose a novel multimodal spatial-frequency fusion and refinement network (MSFFRN). The framework is designed to holistically capture and fuse multimodal features. It employs a dual-branch architecture that integrates wavelet-based frequency decomposition with CNN-based spatial encoding for complementary feature extraction. Furthermore, a hierarchical fusion strategy using adaptive gating and cross-attention mechanisms dynamically integrates cross-modality and cross-domain information, while a dedicated feature refinement module enhances robustness against environmental distortions and preserves structural feature. Extensive experiments on the Houston 2013 and MUUFL datasets demonstrate that MSFFRN achieves state-of-the-art performance, increasing the overall accuracy (OA) by 2.66% and 1.32% on the respective datasets.