Combining meteorological and soil wetness information in machine learning modelling for landslide early warning

Shallow landslides triggered by intense rainfall events pose a serious threat to people and infrastructure in mountainous areas. Regional landslide early warning systems (LEWS) have proven to be a cost-efficient tool for informing the public about the imminent landslide danger. These LEWS are often based on the statistical relationship between rainfall characteristics and landslide inventory information. Previous studies in Switzerland have demonstrated that periods of increased landslide danger are correlated with relative changes in volumetric water content measured at soil moisture stations across the country. In this study, we combine such soil moisture information (including soil water potential) with meteorological data to establish dynamic thresholds for the prediction of landslide probability in both time and space. We train a random forest classifier to separate between critical and non-critical rainfall events. The models are trained and tested on data measured at 136 locations across the entire country during the period from 2008 to 2023. Our trained algorithm allows us to quantify (1) the importance of different climate and soil wetness variables and (2) the benefits of integrating soil wetness and meteorological information within LEWS. We are confident that this study will improve the accuracy and reliability of landslide forecasting at a national scale and contribute to improved landslide risk management in areas with steep slopes.