Modelling landslide susceptibility through a glass-box machine learning

The field of landslide susceptibility modelling has seen the adoption of many different data-driven approaches, spanning from linear models to the most recent deep-learning solutions. In short, simpler models offer greater interpretability, while predictions derived from complex architectures are more difficult to explain. For this reason, complex algorithms are often referred to as black-box models. However, in the context of landslide susceptibility mapping, the ability to provide highly accurate results along with interpretable predictions is highly valuable. In light of these considerations, this study presents a landslide susceptibility mapping by exploring the capabilities of a new generation of interpretable models, namely Explainable Boosting Machines (EBMs). Unlike the majority of explainable approaches that unveil the decisions of a complex model in a post-processing phase, EBMs offer direct interpretability and full transparency. As a consequence, EBMs fall into the category of glass-box models. Notably, the incorporation of these models within studies focusing on the relationship between landslide occurrence and extreme rainfall events raises considerable interest and represents the aim of this work. Therefore, this contribution focuses on landslides triggered by a heavy rainfall event on September 15, 2022, in Central Italy. To analyze the interaction between landslide occurrence and the event, a novel rainfall variable is introduced among the set of predictors, capturing the event’s intensity relative to historical rainfall patterns. Specifically, this rainfall variable is computed as the percentage of precipitation attributed to the event compared to the mean annual rainfall. The rainfall variable also introduces a dynamic component to the proposed modelling, since it may vary at every future rainfall event. As a consequence, by combining the dynamic nature of the rainfall variable with the exact intelligibility of EBMs, the study also presents a landslide susceptibility mapping under potentially different rainfall scenarios with respect to the September 15, 2022 event.