TY  - JOUR
PY  - 2024//
TI  - Shifting from traditional landslide occurrence modeling to scenario estimation with a "glass-box" machine learning
JO  - Science of the total environment
A1  - Caleca, Francesco
A1  - Confuorto, Pierluigi
A1  - Raspini, Federico
A1  - Segoni, Samuele
A1  - Tofani, Veronica
A1  - Casagli, Nicola
A1  - Moretti, Sandro
SP  - ePub
EP  - ePub
VL  - ePub
IS  - ePub
N2  - Extreme rainfall events represent one of the main triggers of landslides. As climate change continues to reshape global weather patterns, the frequency and intensity of such events are increasing, amplifying landslide occurrences and associated threats to communities. In this contribution, we analyze relationships between landslide occurrence and extreme rainfall events by using a "glass-box" machine learning model, namely Explainable Boosting Machine. What sets these models as a "glass-box" technique is their exact intelligibility, offering transparent explanations for their predictions. We leverage these capabilities to model the landslide occurrence induced by an extreme rainfall event in the form of spatial probability (i.e., susceptibility). In doing so, we use the heavy rainfall event in the Misa River Basin (Central Italy) on September 15, 2022. Notably, we introduce a rainfall anomaly among our set of predictors to express the intensity of the event compared to past rainfall patterns. Spatial variable selection and model evaluation through random and spatial routines are incorporated into our protocol. Our findings highlight the critical role of the rainfall anomaly as the most important variable in modeling landslide susceptibility. Furthermore, we leverage the dynamic nature of such a variable to estimate landslide occurrence under different rainfall scenarios.<p />  <p>Language: en</p>
LA  - en
SN  - 0048-9697
UR  - http://dx.doi.org/10.1016/j.scitotenv.2024.175277
ID  - ref1
ER  -