MACHINE LEARNING IN IDENTIFYING KARST SINKHOLES IN THE INNER BLUEGRASS REGION OF KENTUCKY

Karst sinkholes are a major natural hazard, and approximately half of Kentucky is prone to sinkhole development. Information on existing sinkholes is critical in evaluating sinkhole hazard and understanding mechanisms leading to sinkholes, especially catastrophic cover-collapse sinkholes. LiDAR provides accurate and high-resolution topographic information and has been used to improve delineation of sinkholes in many karst regions. However, sinkholes often occur in large numbers, and mapping them from LiDAR through manual visual inspection of each topographic depression can be slow and laborious. To improve the efficiency of LiDAR sinkhole mapping, we applied four machine learning methods (RusBoost, Naïve Bayes, Support Vector Machine, and Neural Network) that use characteristics of LiDAR-derived topographic depressions to identify sinkholes. Sinkhole data from three counties (Bourbon, Woodford, and Jessamine) in the Inner Bluegrass Region of Kentucky were used to derive a dataset for training and testing the machine learning methods. This dataset consisted of 22,884 records with 10 variables for each record. All variables were derived from morphological characteristics of LiDAR-derived topographic depressions. For each method, a random subset of 80 percent of the records was used for training and the remaining 20 percent were used for testing. The test ROCs (receiver operating characteristic curves) showed all four methods achieved high accuracy, as demonstrated with all the AUCs (Area Under the Curve) greater than 0.88. Neural Network appeared to be the most accurate method with an AUC of 0.95. This study suggests that machine learning is promising to help identify sinkholes in karst areas with available high-resolution topographic information.