Operational Regional Scale Landslide Forecasts: Physics-Based and Data-Driven Models

Landslides on natural slopes represent a critical concern in disaster risk management due to the escalating frequency of high-intensity rainfall events. Understanding the mechanisms of slope failure triggered by rainfall or snowmelt, and accurately forecasting slope stability, is challenging due to the inherent complexities involved. At a regional scale, forecasting stability across multiple slopes within predefined boundaries necessitates precise data collection from numerous locations within the region. Implementing a physics-based model for such analysis also requires parameter calibration through back-analysis, followed by forward analysis for slope-stability forecasts.
Geotechnical slope stability models at a regional scale calculate Factor of Safety (FS) values using numerical tools, e.g., infinite slope models, incorporating precipitation data, topographical information, and subsurface characteristics. However, integrating these models into operational slope stability forecasts encounters two primary challenges: firstly, physics-based models often operate independently, making their integration into fully automated workflows utilizing cloud computing challenging; secondly, the complex data collection requirements for these models necessitate frequent updates.
This study addresses these challenges by employing a hybrid approach that combines physics-based and data-driven models for regional-scale slope stability forecasts. A data-driven model predicts the probability of landslides across a large area composed of multiple first-order catchments, for a selected study area in Norway. Additionally, a physics-based model, TRIGRS, predicts pixel-wise FS values within each catchment. Both models operate as cloud services, providing forecasts once daily, and the results are accessible through the Norwegian Geotechnical Institute's (NGI) data platform, NGI Live. The results also show the importance of dividing the large area into smaller zones with more representative geotechnical parameters to improve the overall performance of the model. This research illustrates the practical application of integrating data-driven and physics-based methodologies to develop operational landslide forecasts, a crucial component of effective Landslide Early Warning Systems (LEWS).

Keywords: Landslides, Early warning, Data-driven, Dashboards, TRIGRS.