Mapping of karst sinkholes from LIDAR data using machine-learning methods in the Trieste area

The distribution and characteristics of karst sinkholes is critical for the understanding and evaluation of geohazards. A two-step process, involving computer vision and machine learning methods, has been developed to map and classify depressions as sinkholes. Every depression has been mapped from a LIDAR derived DTM first and later a machine learning random forest binary classifier was then used on the extracted depressions to identify karst sinkholes. The study shows that the two-step method can accurately map karst sinkholes from high resolution DTM even if a visual check must be done on the non-karst sinkhole dataset to improve the classification.