MHCMamba: Multiscale Hybrid Convolution Mamba Network for Hyperspectral and LiDAR Image Classification

The volume of multisource remote sensing (RS) data has grown significantly, making the integration of such data a critical focus in Earth observation research. Although multisource data fusion contributes to improving image classification accuracy, it still faces multiple challenges in data alignment, efficient feature fusion, and computational resource optimization. To address these issues, this article proposes a Multiscale Hybrid Convolution Mamba (MHCMamba) network for the joint analysis of hyperspectral and light detection and ranging (LiDAR) images. The MHCMamba framework begins with a multiscale feature extraction module designed to capture spectral, spatial, and elevation features from both hyperspectral and LiDAR images, followed by a parallel feature tokenization module that converts these features into feature tokens. To effectively fuse multimodal features, the network introduces a novel deep learning framework based on Mamba and designs a mamba feature fusion encoder (MFFE), which leverages the global modeling capacity of Mamba to extract long-range dependencies within each modality, followed by an adaptive fusion mechanism to integrate intermodal information. Furthermore, to enhance the model’s generalization capability, an Edge Prediction Decoder (EPDecoder) is developed to predict LiDAR edge maps using hyperspectral image features, thereby introducing a new loss function. The experiments were conducted on three benchmark datasets: Houston2013, Trento, and MUUFL. The results demonstrate that the proposed MHCMamba model achieves significant improvements in overall accuracy, outperforming state-of-the-art methods by 1.30%, 0.21%, and 1.42%, respectively. Habitually, the MHCMamba’s source code can be made publicly accessible on my profile page at https://github.com/li-zi-qi/MHCMamba.