Spatial susceptibility to flash floods through comparative assessment of bivariate statistical and machine learning techniques in the upper Camiña Basin, northern Chile

Flash floods often occur with little warning, and can severely disrupt daily life, affecting local communities, economies and the environment. Their sudden nature makes preparation difficult, which increases the associated risks. In arid regions, long dry periods followed by abrupt rainfall events add complexity to this challenge. Due to these circumstances, it is essential to have spatial susceptibility data to guide the design of effective preventive measures and to help mitigate associated risks. In this study, we evaluated the spatial susceptibility to flash floods associated with precipitation events using three models––frequency ratio (FR), maximum entropy (MaxEnt) and weights of evidence (WoE). All models were integrated in GIS for the upper Camiña Basin in northern Chile. An inventory of 103 recorded flash-flood events were created from historical records and participatory mapping. Of these, 70% were used for model training and 30% for validation of the results. Each model incorporated 11 conditioning factors, and flash-flood susceptibility maps were generated, divided into five categories, from very low to very high. The performance of the models was evaluated using receiver operating characteristic analysis with area under the curve (AUC) metrics. The highest accuracy was obtained by the MaxEnt model, with an AUC value of 93.33%, followed by the WoE model with 90% and FR model with 83.33%. This work expands our understanding of previously unexplored regions using wellestablished flood-susceptibility approaches. Finally, the findings provide valuable information for decision-makers and planners involved in flash-flood risk management.