Transformer, more than meets the eye: A deep learning approach to integrate rainfall time-series in multi-type landslide probability modelling

Mountainous and hilly regions are frequently affected by landslides, leading to severe damage and loss of life. Different failure mechanisms have distinct geomorphological characteristics and impacts, requiring type-specific treatments. Moreover, incorporating triggering factors, particularly rainfall, is essential to shift from static susceptibility assessments to dynamic, event-specific probability models. While conventional data-driven methods typically use scalar representations of rainfall (e.g., cumulative values over fixed windows), We propose a more comprehensive approach using rainfall as a continuous time series. In this study, we model landslide-type-specific probabilities in the Emilia-Romagna region, which experienced two extreme rainfall events in May 2023, 14 days apart, with a combined return period exceeding 500 years. These events triggered over 80,000 mapped landslides, classified into five types. The 11,670 km² study area was divided into slope units (SUs), with landslide presence labelled by initiation area-SU intersections. Rainfall time series were generated for each SU via spatial interpolation using a linear radial basis function of 188 rain gauges at daily (31 days) and hourly (744 h) intervals. We embedded the rainfall sequences into a Transformer Neural Network (TNN), coupled with a Dense Neural Network (DNN) for static predictors, and tested various model configurations. Performance evaluation on a 30% test set and stratified cross-validation showed strong results (AUC > 0.90; F1 from 0.75 to 0.20, depending on landslide type), with daily time series consistently performing best. To interpret model behavior, we applied the SHAP-based Expected Gradients algorithm, revealing the temporal and spatial influence of rainfall. Though event-specific, the study sets the foundation for generalized, spatiotemporal landslide forecasting models.