DEVELOPING A FREE AND OPEN-SOURCE GRAPHICAL USER INTERFACE FOR CNN-BASED GEOMORPHIC FEATURE DETECTION: CAROLINA BAYS AS A CASE STUDY

Throughout the Atlantic Coastal Plain of the United States (ACP), evidence of extensive aeolian activity throughout the Pleistocene is revealed through LiDAR in the form of Carolina bays. These shallow, circular to ovate, and sandy-rimmed depressions are found from southern New Jersey to southern Georgia, with total count estimates ranging between 10,000 and 500,000. With such a large population size and with such uncertainty around the actual population size, mapping these enigmatic features is a problem that requires an automated detection scheme. Using LiDAR-derived digital elevation models (DEMs) of the ACP as training images, two types of convolutional neural networks (CNNs), Faster R-CNN and Mask R-CNN, were trained to detect Carolina bays. Implementation of the Faster R-CNN model provides bounding boxes with confidence levels for each Carolina bay detection while the Mask R-CNN model provides segmentation masks with confidence levels for each detection. With GPU acceleration, these trained networks can scan through thousands of images per hour and can provide a near-complete catalog of Carolina bays. False positives are filtered using existing geologic and land-use datasets, while edge effects and double counting are mitigated with appropriate DEM tiling schemes. The detection outputs can then be used to subset and analyze regions of the DEMs for statistics on Carolina bay morphology. This method for detecting geomorphic features is a highly efficient process that will provide better means for mapping various types of abundant geomorphic features in the future. Here we present the Carolina bay detection results as well as a free and open-source graphical user interface for object detection on geographic datasets.