Abstract
To prevent damage from landslides, data-driven early warning forecast models have proven to be cost-effective tools. Traditionally, early warning models relied on rainfall exceedance thresholds to differentiate between landslide-triggering and non-triggering conditions. Recent studies have shown that using soil moisture information to assess the moisture state before rainfall events can improve hazard predictions. To explore the possible benefits of using soil moisture measurement data, we compiled a national dataset for Switzerland that combines volumetric water content, soil water potential, and ground temperature data from in situ-based soil moisture measurement stations with meteorological data from nearby stations. The combined dataset was processed by means of two machine learning methodologies to predict landslide hazard at a spatio-temporal scale, herein called “sequential” and “rainfall event” method. The sequential method processes time series data using neural network algorithms, while the rainfall event method predicts landslide probability based on rainfall event characteristics using a random forest classifier. Both methods outperformed traditional intensity-duration rainfall thresholds, with the sequential method overall achieving the highest accuracy in landslide hazard predictions. Including soil moisture data slightly improved the performance metrics compared to purely precipitation-based models and proved particularly valuable in predicting landslides that were triggered by low-intensity rainfall. This study demonstrates the value of in situ soil moisture measurements for national-scale landslide hazard prediction. By combining soil moisture and meteorological data in state-of-the-art machine learning models, we showcased the potential for developing reliable landslide early warning forecast models.