Landslide susceptibility mapping using physics-guided machine learning: a case study of a debris flow event in Colorado Front Range

Abstract

Landslides are common geohazards worldwide, resulting in significant losses to economies and human lives. Data-driven approaches, especially machine learning (ML) models, have been widely used recently for landslide susceptibility mapping (LSM) by extracting features from geospatial variables based on their contribution to landslide occurrences using known distributions of landslides as the training dataset. However, challenges remain in applying ML models for LSM models due to the scarcity and uneven spatial distribution of landslide data coupled with the spatial heterogeneity of hillslope conditions. Moreover, ML models developed with limited data often exhibit unexpected behaviors, resulting in poor interpretability and predictions that deviate from intuitive expectations and established domain knowledge. To overcome these challenges, this study proposes a physics-guided machine learning (PGML) framework that integrates landslide domain knowledge into ML models for LSM. The PGML framework was developed and assessed using a detailed debris flow inventory from a storm event in the Colorado Front Range. Based on the infinite slope model, the factor of safety for the study area was first determined and was subsequently used to constrain the prediction of ML models through a modified loss function and measure the physics consistency of model predictions. To evaluate the robustness and generalizability of the models, this study uses geographical sample selections for model performance evaluation, where ML models are trained and tested across heterogeneous ecoregions. The results of this study demonstrated the efficacy of both physics-based and data-driven methods in determining landslide susceptibility in the study area; however, pure data-driven ML models produced physically unrealistic results and poor generalization performance in new ecoregions. With the incorporation of physical constraints, the PGML model demonstrated notable enhancements in physics consistency and generalization capability, along with reduced model uncertainties across various ecoregions, surpassing the performance of benchmark ML models.