MambaTriNet: A Mamba-Based Tribackbone Multimodal Remote Sensing Image Semantic Segmentation Model

Integrating digital surface models (DSM) with remote sensing image has emerged as a pivotal strategy for remote sensing semantic segmentation. While prevailing dual-branch convolutional frameworks independently process DSM and remote sensing image, their inherent limitations in modeling long-range contextual dependencies persist due to convolutional operations’ local receptive fields. Notably, the hierarchical feature in U-model inherently embodies multiscale complementary relationships, yet current multimodal fusion paradigms insufficiently exploit this architectural advantage. In this letter, we propose an efficient three-branch structure encoder to simultaneously extract DSM features and local and global features of remote sensing images. Notably, we leverage the recently introduced Mamba instead of Transformer to capture long-range dependencies, significantly reducing computational complexity while maintaining competitive performance. The extracted features are integrated using the tri-feature complementary fusion (TriFusion) module, which employs a stepwise fusion strategy to learn spatial complementarity between global and local features and channel-wise complementarity between remote sensing images and DSM. Additionally, we introduce a cross-layer feature guidance (CLFG) module within the skip connections to improve segmentation accuracy by facilitating enhanced cross-layer feature propagation. Extensive experiments conducted on two high-resolution remote sensing datasets, ISPRS Vaihingen and Potsdam, demonstrate that the proposed MambaTriNet surpasses existing state-of-the-art methods in performance metrics, achieving 83.84% and 86.05% mIoU, respectively.