Integration of Physics‐Based and Data‐Driven Approaches for Landslide Susceptibility Assessment

Abstract

Rainfall‐triggered landslides pose a significant threat to communities and infrastructure around the world. Various data‐driven and machine learning (ML) based algorithms have been applied to assess landslide susceptibility. However, purely data‐driven methods are affected by issues such as uncertainty in the selection of landslide conditioning factors, potential extrapolation problem, as well as the quantity and quality of historical landslide datasets. On the other hand, physics‐based models require soil properties and initial/boundary conditions which are difficult to obtain. In this work, we develop an enhanced methodology for landslide susceptibility assessment by integrating physics‐based and data‐driven models. For this purpose, Transient Rainfall Infiltration and Grid‐Based Regional Slope‐Stability (TRIGRS) and the shallow slope stability (SHALSTAB) physics‐based models are adopted which provide important insights into the geological and hydrological characteristics of the study area. For the data‐driven approach, XGBoost is implemented due to its demonstrated effectiveness in landslide predictions. We propose two strategies for integrating the physics‐based and data‐driven models: (1) physics‐informed machine learning (PIML) which incorporates outputs of TRIGRS and SHALSTAB into the ML models, and (2) a matrix approach for combining data‐driven susceptibility maps with factor of safety maps. We evaluate the overall performance of models based on three aspects: general prediction capability, data efficiency, and extrapolation. Subsequently, landslide susceptibility maps of California are generated and compared using the PIML model, the data‐driven model, and the matrix approach. The results indicate a substantial enhancement in landslide susceptibility mapping, extrapolation capability and model performance, particularly when limited data is available.