Comparative analysis of conventional and machine learning techniques for rainfall threshold evaluation under complex geological conditions

Abstract This research focuses on the essential task of defining rainfall thresholds in regions with complex geological features, specifically at a regional scale. It examines a variety of methodologies, from traditional empirical-statistical methods to cutting-edge machine learning (ML) techniques, for establishing these thresholds. The Emilia-Romagna region in Italy, known for its intricate geological structure and prevalence of weak rocks that often lead to large and deep-seated landslides, serves as the study area. The region’s complex interplay between rainfall and landslide incidences poses a significant challenge in accurately determining rainfall thresholds. The effectiveness of ML methods is compared against conventional empirical-statistical approaches, evaluating factors such as prediction accuracy, model complexity, and the interpretability of results for use by regional landslide warning system operators. The findings indicate that machine learning techniques have an edge over traditional methods, yielding higher performance scores and fewer false positives. Nevertheless, these advancements are modest when considering the increased complexity of ML methods and the incorporation of additional rainfall parameters. This underlines the continued need for improvements in data quality and volume. The study stresses the importance of enhancing data collection and analysis techniques, especially in an era where advanced AI tools are increasingly available, to improve the accuracy of predicting rainfall thresholds for effective landslide warning systems.