CMFNet: Cross Mamba Fusion Network for Hyperspectral and LiDAR Data Classification

The joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data, which leverages their complementary information, has emerged as a crucial research direction in remote sensing. However, the significant differences in imaging mechanisms result in high heterogeneity between HSI and LiDAR data in terms of dimensionality and feature distributions, posing substantial challenges for semantic representation and correlation of multimodal data. To address this challenge, this article proposes a Cross Mamba Fusion Network (CMFNet) for hyperspectral and LiDAR data classification. The proposed framework adopts a dual-branch architecture to facilitate multimodal feature extraction and interaction, initially performing multiscale convolutional feature embedding before adaptively fusing global and local features through a dedicated Global-Local Feature Extraction (GLFE) module. To further exploit the complementary advantages of multi-source data, we design a Cross Mamba Fusion (CMF) module that extends the Mamba architecture to enable efficient dual-modal interaction. This module enhances cross-modal feature fusion while reducing computational redundancy and achieves hierarchical feature extraction through a multi-layer structure. Comprehensive experiments on three benchmark remote sensing datasets demonstrate that CMFNet significantly outperforms state-of-the-art methods, with ablation studies thoroughly validating the effectiveness of each component. Habitually, the CMFNet’s source code can be made publicly accessible on my profile page at https://github.com/li-zi-qi/CMFNet.