A deep learning-based model for endorsing predictive accuracies of landslide prediction: insights into soil moisture dynamics

Abstract Introduction and Research Gap This study presents a comprehensive framework for predicting volumetric water content (VWC) to mitigate shallow, rainfall-induced landslides, bridging existing gaps in the literature. Methodology The framework synergistically integrates the empirical strengths of deep learning (DL) with the physical dynamics of the VWC subsurface behavior. Statistical, shallow machine learning (ML), and DL models were investigated with optimization techniques and sensitivity analyses to establish benchmarks for comparison and derive optimal predictions. DL and probability theory enable both point and interval predictions. Findings Validation on the Pa Mei landslide demonstrates strong performance with mean absolute errors (MAE) ranging from 0.35% to 1.22% and Predicted Interval Coverage Probabilities (PICP) from 0.86 to 0.91. Predicted VWC deviations were propagated into Factor of Safety (FOS) calculations, yielding robust performance metrics with R 2 and PICP of 0.89 and 0.85, respectively. Transferability is demonstrated at the Tung Chung landslide, where MAE ranges from 0.36% to 1.25% and PICP from 0.86 to 0.95. Significance This framework demonstrates improved accuracy and introduces a practical data-sharing mechanism to address monitoring challenges such as power consumption and data loss, offering a robust tool for hazard mitigation and decision support.