MSRS-MambaUNet: A multi-source remote sensing model for landslide detection

Recent advances in artificial intelligence have increasingly empowered landslide detection and geological hazard mitigation. However, existing deep learning architectures still face challenges in global contextual modeling, multi-scale feature extraction, and computational efficiency. To address these challenges, we propose the MSRS-MambaUNet, a novel multi-scale remote sensing model designed for landslide detection. The proposed model integrates the Omnidirectional Selective Scan Module (OSSM) and the Multi-Scale Feed-Forward Network (MS-FFN) as its core components. Specifically, the OSSM block globally models multi-directional long-range dependencies, while the MS-FFN aggregates multi-directional contextual information to facilitate the efficient extraction of multi-scale features. We evaluate the proposed model on two representative clustered landslide events triggered by the 2022 Lushan earthquake in a seismically active mountainous region and the 2024 Shaoguan rainstorm in a rainfall-prone hilly area. Beyond the red-green-blue bands of optical imagery, additional features derived from synthetic aperture radar, spectral indices, and topographic data are incorporated through a progressive feature enrichment strategy to enhance data representation capability. Experimental results show that the proposed model outperforms conventional approaches in both study areas. In the Lushan study area, F 1 -score and intersection over union (IoU) increase by 0.81%–4.06% and 1.33%–6.69%, respectively, while in the Shaoguan study area, the corresponding improvements range from 0.35%–6.87% and 0.56%–10.92%. Multi-source remote sensing data integration also improves detection performance, with topographic features acting as among the most effective auxiliary inputs, yielding improvements of 7.50% ( F 1 -score ) and 12.00% ( IoU ) in Lushan, and 2.63% and 4.05% in Shaoguan.