Shaping shallow landslide susceptibility as a function of rainfall events

Abstract. This paper tests a multivariate statistical model to simulate rainfall-dependent susceptibility scenarios of shallow landslides. To this end, extreme rainfall events spanning from 1977 to 2021 in the Orba basin (a study area of 595 km2 located in Piedmont, northern Italy) have been considered. First of all, the role of conditioning and triggering factors on the spatial pattern of shallow landslides in areas with complex geological conditions is analysed by comparing their spatial distribution and their influence within logistic regression models, with results showing that rainfall and specific lithological and geomorphological conditions exert the strongest control on the spatial pattern of landslides. Different rainfall-based scenarios were then modelled using logistic regression models trained on different combinations of past events and evaluated using an ensemble of performance metrics. Models calibrated on multiple events outperform the ones based on a single event, since they are capable of compensating for local misleading effects that can arise from the use of a single rainfall event. The best-performing developed model considers all the landslide-triggering rainfall scenarios and two non-triggering intense rainfall events, with a score of 0.90 out of 1 on the multi-criteria TOPSIS-based1 performance index. Finally, a new approach based on misclassification costs is proposed to account for false negatives and false positives in the predicted susceptibility maps. Overall, this approach based on a multi-event calibration and on a misclassification cost analysis shows promise in producing rainfall-dependent shallow landslide susceptibility scenarios that could be used for hazard analyses and early warning systems and could assist decision-makers in developing risk mitigation strategies.