LHEM-MSegNet: A Landslide Hazard Extraction Model Based on Multi-Source Data Fusion

Accurate detection of loess landslide hazards is critical for disaster prevention, yet remains challenging due to complex terrain, spectral similarity to background regions, and the lack of specialized datasets. To address these challenges, this paper proposes LHEM-MSegNet, a multisource landslide hazard extraction model based on Transformer and attention fusion. A loess landslide hazard region segmentation dataset for Linyou County (LYPLH), Shaanxi Province, China, is constructed using a transfer-learning strategy based on the Segment Anything Model (SAM), followed by iterative manual correction and data augmentation to improve annotation quality. LHEM-MSegNet integrates Transformer blocks with spatial and channel attention modules to enhance feature representation, and adopts a multi-pathway architecture for multisource fusion. Specifically, optical remote sensing imagery is combined with geomorphological information, including DEM, slope, and aspect maps, to improve landslide hazard extraction. Experimental results on the LYPLH dataset demonstrate that the proposed LHEM-MSegNet achieves an mIoU of 89.29% by integrating optical imagery with geomorphological information. Ablation studies first demonstrate the effectiveness of the proposed model structure and further confirm the significant contribution of terrain-related features to landslide hazard extraction. In addition, the supplementary evaluation on the external Landslide4Sense benchmark shows that the proposed method maintains competitive performance under different data distributions. The proposed approach provides an effective solution for early loess landslide hazard identification and supports disaster mitigation in the Loess Plateau region.