Physics-Informed Machine Learning Framework for Predicting Rainfall-Induced Shallow Landslides in the Colorado Front Range

Landslides are complex natural hazards that are challenging to predict. This study introduces a physics-informed machine learning (PIML) framework that integrates a differentiable physics-based landslide model into neural networks (NNs). The framework is process-aware, incorporating hydrological processes, such as rainfall infiltration and runoff, to introduce temporal dependencies and reduce the reliance on precise timing of landslide occurrences. Within the PIML framework, the NN estimates key geotechnical parameters for the physics-based model, which predicts landslide occurrence under evolving hydrological conditions based on the infinite slope assumption. The PIML framework is applied to the September 2013 storm-triggered landslides in the Colorado Front Range, using a comprehensive data set that includes a landslide inventory, hourly rainfall, and remote sensing data. Results show that while the physics-based model provides a general assessment of hillslope stability, it struggles to accurately identify failure-prone locations, with an area under the curve (AUC) of 0.84. In contrast, the PIML framework improves landslide prediction performance by 13% (AUC = 0.95). The PIML effectively bridges the gap between data-driven learning and physics-consistent modeling, enabling robust landslide prediction.