Enhancing rainfall-triggered landslide forecasting in Switzerland using ensemble learning

Shallow landslides, typically occurring on steep slopes, are often triggered by intense, short-duration rainfall or extended periods of lighter rainfall. These events present severe hazards in mountainous regions, causing substantial soil loss, fatalities, and economic damage (Tonini and Cama, 2019). Accurate prediction and early warning systems are essential for mitigating such impacts. To address these challenges, previous studies in Switzerland have examined rainfall thresholds related to landslide triggering by regionalizing landslide occurrences according to geomorphological factors (Leonarduzzi et al., 2017). To enhance the overall accuracy of such predictions, it is essential to utilize datasets with higher temporal and spatial resolution. 

This work adapts a robust deep learning approach initially developed by Mondini et al. (2023) for Italy to the case of Switzerland. Unlike previous studies that relied solely on rain gauge data, which is often highly variable, we use the CombiPrecip product from the Swiss Federal Office of Meteorology. This product integrates radar measurements with rain gauge data to provide a kilometer-scale, hourly precipitation dataset covering the past 20 years. The landslide input dataset comes from the Swiss Federal Institute for Forest, Snow, and Landscape Research (WSL), which has systematically collected data on damage caused by naturally triggered floods, debris flows, and landslides since 1972 (Hilker et al., 2009). 

To compensate for the relative sparsity of landslide events in our training set, we carry out an ensemble approach where we train 24 classifiers, thus resulting in increased robustness and a probabilistic outcome. The ultimate goal of this research is to compare various classification algorithms and evaluate their integration into an early warning system that leverages susceptibility maps and geological factors.

REFERENCES

• Tonini M, Cama M (2019). Spatio-temporal pattern distribution of landslides causing damage in Switzerland. Landslides 16, 2103–2113. 
• Leonarduzzi E, Molnar P, McArdell BW (2017). Predictive performance of rainfall thresholds for shallow landslides in Switzerland from gridded daily data. Water Resources Research, 53(8), 6612‑6625. 
• Mondini AC, Guzzetti F, Melillo M (2023). Deep learning forecast of rainfall-induced shallow landslides. Nature Communications, 14(1), 2466. 
• Hilker N, Badoux A, Hegg C (2009). The Swiss flood and landslide damage database 1972-2007. Nat Hazards Earth Syst Sci 9:913–925.