<p>Regional landslide extraction is the basis of landslide hazard risk management. Acquiring landslide location and size information timely and accurately provides essential information for post-event damage assessment, rapid updating of hazard maps, and identification of potential source areas for secondary hazards, thereby supporting disaster risk reduction strategies. Recently, deep learning for landslide extraction has increased the speed and accuracy of extraction. However, the large amounts of labeled samples required for deep learning models remain a major challenge in the study on the automated extraction of mining-induced landslides. Here, we develop the MRSD-U-Net model, an enhanced deep learning network integrating multi-source remote sensing data and residual learning module. We find that this model enhances the extraction of landslide features by deepening the network and capturing topographic, textural, and spectral information. It achieves high-precision landslide extraction with a limited number of available labeled samples. The performance assessment shows that MRSD-U-Net achieves lower error rates, higher accuracy, stronger generalization ability, and higher consistency between the extracted landslide areas and the ground reference data compared to other advanced deep learning models. Furthermore, MRSD-U-Net also exhibits the applicability in cross-scene landslide extraction. This study provides an effective method for the rapid and precise extraction of mining-induced landslides in data-scarce areas.</p>