<p>Landslides are highly destructive geological hazards commonly triggered by earthquakes, extreme rainfall, and permafrost degradation, posing serious threats to ecosystems and infrastructure. Rapid and reliable post-event landslide mapping remains challenging in complex mountainous environments, where field investigations are inefficient and many automated approaches struggle to jointly achieve accuracy, interpretability, and computational efficiency. To address these challenges, we present a dual-encoder multiscale Transformer fusion network (DMTFNet) for landslide detection using multi-source remote sensing data. Sentinel-2 multispectral imagery and a digital elevation model are integrated through 17 input channels. A dual-branch convolutional encoder decouples spectral and topographic feature learning, while a multiscale Transformer decoder with multi-head attention captures long-range spatial dependencies. An adaptive-weighted fusion module integrates hierarchical representations to enhance boundary delineation. Experiments in the 2022 Luding and Lushan earthquake zones yield mIoU of 92.74% and 82.71%, respectively, outperforming state-of-the-art baselines. Qualitative results indicate improved landslide delineation with reduced omission and false detection. Factor-importance analysis further reveals region-specific feature sensitivities, with vegetation-related cues dominating discrimination in Luding and topographic controls prevailing in Lushan. Overall, DMTFNet provides an accurate, interpretable, and computationally efficient framework for medium-resolution satellite-based landslide mapping, achieving a favorable accuracy–efficiency trade-off and supporting large-area, near-real-time post-disaster response, inventory maintenance, and regional risk assessment.</p>