Emilia-Romagna, Italy: could the landslide phenomena of May 2023 have been predicted?

In May 2023, the Emilia-Romagna region (Italy) was hit by intense rainfall, causing significant economic damage (8.6 billion euros) and 17 fatalities. After a prolonged dry period, with only occasional light precipitation, two major rainfall events occurred on May 1 st - 3 rd and May 16 th -17 th , bringing cumulative precipitation exceeding 500 millimetres in various locations. The Emilia-Romagna Region reported 80,946 landslides triggered by these rainfall events, many of which damaged roads in approximately 2,000 cases and numerous houses. Among the fan of landslide typologies recognized, mud/earth/debris flows, which are the focus of this research, largely prevail on subordinated slides, mainly affecting Miocene sandy clay/sandstone alternations. The main question of this research is: could these landslide phenomena have been predicted? Focusing on the small basin of the Senio River (180 km²), we first collected a pre-event and post-event inventory, identifying flow landslides through remote mapping from 2002 to 2022 (11,597 cases, seasonal rainfall trigger), and after May 2023 (8,323 cases, linked to extreme events). Using the Multivariate Adaptive Regression Splines (MARS) statistical method and a set of eleven predictor variables, we analysed the relationships between presence/absence of landslides and spatial distribution of the geo-environmental variables. In particular, two model-building procedures - the PRE2023 and POST2023 models - were developed by calibrating with 75% of the pre/post-event inventory (a balanced subset of the main inventory), respectively. Four validation procedures were then performed: SELFPRE2023 and SELFPOST2023, for which the above-mentioned models were validated using the remaining specific 25%, FRW (forward validation) where the SELFPRE2023 model was used to predict the 25% of the post-2023 inventory, and BKW (backward validation) where the SELFPOST2023 model was used to predict the 25% of the pre-2023 inventory. Even if the PRE2023 model achieves exceptional results in predicting the coeval blinded 25% (AUC= 0.95, Sensitivity =0.89, Specificity = 0.89), its prediction skill significantly decreases in FRW validation (AUC= 0.8, Sensitivity= 0.56, Specificity =0.89), demonstrating a marked inability to predict the landslides triggered by the extreme events. On the other hand, the POST2023 model delivers very good results both in predicting the coeval inventory (AUC= 0.87, Sensitivity= 0.71, Specificity= 0.88) and in BKW validation (AUC= 0.9, Sensitivity= 0.78, Specificity= 0.88). However, the slightly high sensitivity value of the POST2023 model suggests that it is not particularly effective at discriminating landslides triggered by the extreme event (coeval), as evidenced by the increase in sensitivity in BKW validation. The results obtained led us to a detailed analysis of the variables used by the two models, as well as an extensive study of the precipitation trends before the 2023 events. The latter was conducted by examining the records from six rain gauges located within or near the basin, covering the period from 2006 to 2023. The results show that not only did the intensity of the 2023 events exceed that of any other extreme event recorded during the considered period, but also that the spatial distribution of these precipitations was anomalous, triggering areas of the basin that had historically been stable. This study was carried out within the RETURN Extended Partnership and received funding from the European Union Next-GenerationEU (National Recovery and Resilience Plan – NRRP, Mission 4, Component 2, Investment 1.3 – D.D. 1243 2/8/2022, PE0000005).