Demonstration of a deep-learning-based system for rainfall induced shallow landslides forecasting in Italy

A common and largely unresolved problem of national-scale landslide early warning systems is their independent evaluation. In this work, we evaluated the performance of a recently proposed deep-learning-based system for short-term forecasting of rainfall-induced shallow landslides in Italy. For our evaluation, we used hourly rainfall measurements from the same rain gauge network used to construct the forecasting system, and different and independent information on the timing and location of 163 rainfall-induced landslides that occurred in Italy in a period non considered in the construction of the forecasting system, obtained from the FraneItalia catalogue (https://zenodo.org/records/7923683). The independent evaluation confirmed the good predictive performance of the forecasting system and revealed no geographical or temporal bias in the forecasts. The analysis also revealed that the forecasting system was more effective at predicting multiple landslides in the same general area than single landslides. This was a good result, as multiple landslides are potentially more dangerous than single failures. Analysis of the few misclassified landslide cases showed that approximately one-third of the landslides were rockfalls, and for approximately another third there was uncertainty about when or where the landslides occurred. We conclude that, despite the inevitable misclassifications inherent in any probabilistically based national-scale landslide forecasting system, the deep-learning-based system analysed is well suited for short-term operational forecasting of rainfall-induced shallow landslides in Italy.