Deep learning-augmented crack mapping and SPH-based dynamic simulation for landslide kinematic prediction

On July 19, 2020, the Tanjiawan landslide in Badong County, China, underwent pronounced deformation triggered by extreme rainfall, posing substantial risks to local communities and infrastructure. To elucidate its spatiotemporal evolution and failure mechanisms, a comprehensive field campaign was conducted, supported by multi-source datasets including geological, topographic, lithological, precipitation, satellite remote sensing, and unmanned aerial vehicle (UAV)-based photogrammetry. Deep learning (DL) models (U-Net and ResU-Net) were employed to automatically extract ground cracks from high-resolution UAV-based orthophotos. ResU-Net achieved superior recognition accuracy (89.7%) compared to U-Net (85.9%), with cracks predominantly distributed along the rear scarp and lateral margins, consistent with field observations. Landslide stability was evaluated using the limit equilibrium method, coupled with Monte Carlo simulations to account for parameter uncertainty. Under extreme rainfall conditions (320 mm/d), the factor of safety decreased markedly from 1.079 to 0.822, while the failure probability increased by 72.82%. The post-failure dynamics were simulated using a smoothed particle hydrodynamics (SPH) model, which predicted a total run-out duration of 48 s, a peak velocity of 28 m/s at 19 s, and an average final deposit thickness of approximately 5 m. The integrated framework demonstrates the potential of DL probabilistic stability analysis, and particle-based modeling for quantitative landslide hazard assessment and early-warning applications in rainfall-sensitive mountainous regions.