Sinkhole susceptibility analysis using machine learning for west central Florida

This study examined the feasibility and accuracy of applying machine learning for sinkhole classification and prediction and using the results in automated sinkhole susceptibility mapping for west central Florida. A two-stage processing pipeline was developed. In the first stage, we assessed the predictive power of five exemplary machine learning algorithms: random forest (RF), logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), and multilayer perceptron (MLP), and select the best-performing model. The top-performed model was then chosen to develop a sinkhole susceptibility map (SSM) in the second step of the process. Nine feature layers were derived from the collected geospatial data and utilized as conditional variables. Several statistical metrics and receiver operating characteristic curves were utilized to evaluate the accuracy of the models. The results showed that the RF model, with a ROC of 0.984, had the highest prediction capability in the research area.

We generated a susceptibility map using the RF model, and the study area was classified into high susceptibility (H) and low susceptibility (L) areas. Confusion Matrix (CM) and Matthews Correlation Coefficient (MCC) were used to confirm the results of the sinkhole susceptibility map's classification. We present a model that predicts sinkhole distribution in the study area, and the output of our model is consistent with the sinkhole hazard map that the Florida Division of Emergency Management had previously created. This work can assist the government, community, and land managers in creating plans for mitigating hazards and land degradation.