Large-scale landslide susceptibility mapping via deep learning: A case study of Pakistan

Landslide hazards are highly abrupt and extremely destructive, posing a severe threat to global geological security. Landslide susceptibility mapping (LSM) is a critical technology for reducing geological hazards, yet current methods still exhibit shortcomings in large-scale spatial modeling and the reliability assessment of results. In this study, an integrated pixel-based framework for large-scale LSM was developed and applied to the entire territory of Pakistan (880,000 km 2 ) using 20 landslide conditioning factors (LCFs) grouped into six categories. We implemented difference-weighted stratified sampling (DWSS) for training-negative selection, and adopted the LandslideNet model integrated with the spatial perception block (SPB) and dynamic channel squeeze-and-excitation (DCSE) module to conduct spatial correlation modeling. The challenge of large-scale computing was addressed via a multi-level block division and combination strategy, and the multi-stride median fusion (MSMF) algorithm was developed to achieve seamless merging of sub-blocks. The proposed framework generated the first 100 m national landslide susceptibility map for Pakistan. Under five-fold spatially contiguous regional hold-out evaluation, LandslideNet achieved the highest ROC-AUC (93.62 ± 3.41%) and PR-AUC (94.22 ± 4.17%) among the benchmark models, together with an F1-score of 89.07 ± 5.67%. The resulting susceptibility surface reveals a pronounced national spatial gradient, with high susceptibility concentrated in the northern mountains, more dispersed susceptible terrain across the western ranges, and predominantly low susceptibility over the southeastern plains. SBAS-InSAR analysis further showed the strongest deformation–susceptibility correspondence for LandslideNet among the evaluated models. SHAP analysis identified elevation and slope as the dominant contributors to model predictions, followed by distance to roads, temperature, and lithology. These results demonstrate the value of integrating established techniques for national-scale landslide susceptibility assessment and provide detailed susceptibility information for regional hazard screening and infrastructure planning. The code is available at https: // github.com/Trifurs/LandslideNet .