Prediction of rainfall-induced landslides in the Veneto region using LSTM model

Rainfall-induced landslides threaten nearly 8.2% of the human population globally, making accurate forecasting vital for risk mitigation in vulnerable topographies like the Veneto region of Italy. The existing Landslide Early Warning Systems (LEWS) issue alerts using rainfall thresholds derived from fixed time windows, which struggle to capture complex meteorological histories such as heavy rainfall following prolonged dry spells. This thesis investigates whether Long Short-Term Memory (LSTM) recurrent neural networks, which can process an entire precipitation time series rather than a fixed threshold, can better predict rainfall-induced landslides in the Veneto region. This study utilized a dataset made up of 835 landslide events spanning 2000 to 2024, compiled from the Italica and FraneItalia inventories and matched with thirty-day hourly precipitation time series from 152 ARPAV rain gauges using ArcGIS Pro spatial tools and Thiessen polygons. A random sample of 10,000 non-landslide thirty-day precipitation time series was also prepared. Because the inventories vary in trigger type and spatial/temporal accuracy, four train, validation, and test configurations were built, ranging from full unfiltered dataset (configuration 1) to progressively stricter filters for rainfall-only triggers and high geographical accuracy (configuration 2-4), each tested under both balanced and unbalanced landslide-to-non-landslide ratios. The LSTM model performed very well. Results demonstrate that configuration 2 (utilizing events strictly triggered by rainfall or combination of rainfall with other events) achieved the most reliable predictive performance, with an Area Under Curve (AUC) of 0.899 and an overall accuracy of 84.2% in the unbalanced scenario. Thereafter, a test using only seven-day precipitation time series on configuration 2 resulted in a performance matching that of the thirty-day model. This suggests that a majority of the recorded landslides behave as shallow, short-response events rather than deep-seated ones. The results indicate that data quantity and trigger-label consistency are the most important factors of this study. A critical caveat for any operational warning system built on this approach could be the degradation of the recall value, which governs missed landslides, in realistic unbalanced class ratios. Future improvements could involve the application of XAI (explainable AI) to identify which parts of the rainfall series derive predictions, and replacement of observed rainfall with rainfall forecast reanalysis data to predict slope failures before they occur in the real world.