MambaUNet: exploiting long-range dependencies for RGB-D semantic segmentation in remote sensing

Semantic segmentation of remote sensing imagery is essential for a wide range of Earth observation applications but remains challenging due to complex backgrounds, scale variations, and spectral ambiguities. To address these issues, we propose MambaUNet, a dual-branch hybrid convolutional neural network (CNN)–Mamba architecture tailored for multimodal semantic segmentation of RGB-D remote sensing data. The framework leverages CNNs for efficient local feature extraction and Mamba modules for capturing long-range dependencies, which are combined through a hierarchical midfusion strategy. In particular, shallow-level interactions are facilitated by the proposed shallow-modality fusion modules, whereas high-level cross-modal semantics are reinforced via a Mamba-modality fusion module. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that MambaUNet consistently outperforms both unimodal and multimodal baselines. Furthermore, ablation studies verify the contribution of each component, and complexity analysis highlights a favorable trade-off between segmentation accuracy and computational efficiency.