Maximum Entropy Modeling for Multi-Hazard Spatial Distribution: A Case Study of Flood-Triggered Sinkholes

Recent decades have seen a growing availability of detailed geo-environmental data, coupled with powerful open-access software and machine-learning algorithms, driving significant advancements in natural hazard forecasting. Exploring cutting-edge machine-learning techniques is essential to understanding their strengths and limitations, which vary with factors such as data quality, hazard types, and the complexity of variable relationships. In this study, we extend the application of the Maximum Entropy model (MaxEnt) initially applied to ecological research to a novel context by characterising a common multi-hazard scenario in karst settings (i.e., flood-triggered sinkholes). While MaxEnt has been widely used by ecologists to model species distributions, its application in natural hazard modelling, particularly for hidden hazards like sinkholes in karst regions, remains underexplored.

Here, we applied MaxEnt to forecast the spatial probability distribution of flood-triggered sinkholes. Model inputs included past sinkhole occurrence data and geo-environmental factors such as topography, local geology, hydrology, and flood hazard. The model was validated using 70% of the sinkhole inventory for training and the remaining 30% for testing, with performance assessed using the Area Under the Curve of the Receiver Operating Characteristic (AUC-ROC).

The resulting susceptibility map highlights areas up to 1 km south of the Loire River and low-elevation zones as most vulnerable to flood-triggered sinkholes. Our findings demonstrate that this multi-hazard scenario mapping approach is a valuable tool for identifying flood-triggered sinkholes in Val d’Orléans and other karst regions worldwide, supporting effective land-use planning. By applying MaxEnt at different spatial scales, we also identified limitations affecting the model’s final accuracy, which provide insights for future improvements.