Abstract
Sinkholes cause persistent damage to infrastructure, property, and groundwater resources across Florida. Statewide sinkhole hazard mitigation has been constrained by incomplete inventories and the absence of a high-resolution susceptibility map. This study presents a statewide sinkhole susceptibility assessment using an explainable GeoAI framework that integrates the citizen-reported Subsidence Incident Reports (SIR) database and a geologist-confirmed collapse-type sinkhole inventory from the Florida Geological Survey (FGS). Random Forest (RF) is used as the primary learner with ten geologic, hydrologic, topographic, and anthropogenic predictors. The influence of negative-sample buffer geometry on model behavior is evaluated across five annular buffer conditions (100-1000 m through 1000-10000 m). Two Explainable Artificial Intelligence XAI methods, Gini-based feature importance, and accumulated local effects (ALE), identify dominant controls and nonlinear predictor responses. The statewide model achieved receiver operating characteristic area under the curve (ROC-AUC) of 0.859 under the selected 500-5000 m buffer. Land use/land cover (LULC), overburden thickness, and distance to karst features are the leading predictors, and their nonlinear thresholds align with well-established karst dissolution and collapse processes.
Rights Information
DOI
https://doi.org/10.5038/9781967518012.1015
Altarawneh-Fig.2.jpg (312 kB)
Altarawneh-Fig.3.jpg (167 kB)
Altarawneh-Fig.4.jpg (796 kB)
Altarawneh-Fig.5.jpg (446 kB)
Altarawneh-Table.1.jpg (367 kB)
Altarawneh-Table.2.jpg (203 kB)
Altarawneh-Table.3.jpg (242 kB)
Statewide Sinkhole Susceptibility Mapping in Florida Using Explainable GeoAI
Sinkholes cause persistent damage to infrastructure, property, and groundwater resources across Florida. Statewide sinkhole hazard mitigation has been constrained by incomplete inventories and the absence of a high-resolution susceptibility map. This study presents a statewide sinkhole susceptibility assessment using an explainable GeoAI framework that integrates the citizen-reported Subsidence Incident Reports (SIR) database and a geologist-confirmed collapse-type sinkhole inventory from the Florida Geological Survey (FGS). Random Forest (RF) is used as the primary learner with ten geologic, hydrologic, topographic, and anthropogenic predictors. The influence of negative-sample buffer geometry on model behavior is evaluated across five annular buffer conditions (100-1000 m through 1000-10000 m). Two Explainable Artificial Intelligence XAI methods, Gini-based feature importance, and accumulated local effects (ALE), identify dominant controls and nonlinear predictor responses. The statewide model achieved receiver operating characteristic area under the curve (ROC-AUC) of 0.859 under the selected 500-5000 m buffer. Land use/land cover (LULC), overburden thickness, and distance to karst features are the leading predictors, and their nonlinear thresholds align with well-established karst dissolution and collapse processes.