Potential Landslide Area Identification Method Based on Deep Generative Adversarial Reinforcement Learning (DGARL-LS)

Landslides pose serious threats to lives and property in mountainous areas, yet large-scale preliminary screening of potential landslide areas remains costly and constrained by vegetation obstruction and data availability. Existing deep learning methods depend heavily on multi-source remote sensing data and dense annotations, limiting their applicability in data-scarce regions. This study proposes DGARL-LS, a potential landslide area identification method, which uses 10 m resolution DEMs as the sole data source and extracts slope and curvature information to construct curvature–slope (CS) stereoscopic maps. A pix2pix conditional GAN serves as the segmentation backbone, integrated with a bidirectional exploratory deep reinforcement learning module in which forward exploration localizes candidate regions and backward trajectory priors alleviate reward sparsity. GAN discriminator scores provide the reward signal, forming a closed-loop localization–segmentation–optimization framework. Training utilizes 15,000 augmented sample pairs annotated via a dual-constraint morphological–structural system in the Three Gorges Reservoir area. DGARL-LS achieves IoU of 76.8%, Dice of 83.5%, precision of 85.6%, and Recall of 80.2% on a 1800 km2 test area, outperforming U-Net, pix2pix, Trans U-Net, and Swin U-Net, with approximately 25% faster convergence than U-Net. Cross-regional tests in Dazhou (Sichuan) adopt a small, unaugmented dataset for full retraining. DGARL-LS achieves 64.98% IoU, exceeding the second-best baseline (53.69%), while its absolute performance, especially IoU, decreases relative to the main research area. These results demonstrate that accurate, large-scale potential landslide screening is achievable using only freely available DEM data, providing a lightweight and transferable technical pathway for geohazard prevention in complex mountainous regions.