Key points are not available for this paper at this time.
Remote-sensing image semantic segmentation plays a crucial role in land resource surveys, urban planning, and environmental monitoring, requiring precise boundary delineation and robust cross-domain generalization capabilities. However, existing methods exhibit significant limitations: Convolutional Neural Networks (CNNs) struggle to establish long-range dependencies due to limited receptive fields, transformer models face quadratic computational complexity and poor cross-domain adaptability, while State Space Models suffer from directional bias when processing anisotropic spatial structures in remote-sensing images. This paper proposes DirecMamba, a novel hybrid expert architecture that integrates CNNs, Transformers, and Mamba models through three key innovations: employing a Directional Mamba Mixture of Experts system (DirecMamba-MoE) composed of four expert groups with independent parameters for specialized processing of different scanning directions; integrating enhanced self-attention mechanisms with SwiGLU gating units to improve nonlinear expression capabilities; deploying MLP-Mixer networks across decoder stages to achieve efficient multi-scale feature fusion with linear complexity. Extensive experiments validate the superior performance of our method: achieving 71.3% mIoU on the UAVid dataset, surpassing UrbanSSF-L (71.0%) and significantly outperforming UMFormer (65.7%); obtaining 85.3% mIoU, 92.3% mF1, and 93.5% OA on the ISPRS Vaihingen dataset. The directional expert modules effectively capture orientation-dependent features, with road segmentation showing particularly notable improvements.
Li et al. (Mon,) studied this question.