IMPROVED MAPPING OF GEOMORPHIC FEATURES THROUGH MACHINE LEARNING IN THE BUFFALO RIVER WATERSHED, ARKANSAS

Fluvial terrace deposits in the Buffalo National Scenic River watershed (BNR) have been the focus of recent research to understand landscape evolution in the catchment. Although geologic mapping has been ongoing for more than twenty years by the U.S. Geological Survey and the Arkansas Geological Survey, terrace deposits were not initially identified and mapped. As new quadrangles are mapped and older maps revised, these critical geomorphic features are being added. The remote and steep terrain of the BNR means this is not a trivial endeavor; therefore, developing automated methods for predicting diagnostic geomorphic features, such as terraces, has the potential to improve field mapping efforts. Working with a high-resolution LiDAR dataset, this research applies machine learning to the problem of predicting geomorphic features in the intricate landscape of the BNR. As the use of machine learning in geomorphology is relatively novel, this work focused on basic comparisons of the relative performance of several learning algorithms and the effects of using different descriptive variables. Since geomorphic features lack a one-to-one relationship between physical form and geomorphic class and are frequently not well-bounded, making geomorphic feature recognition a challenging task. Local terrain derivatives, such as slope and curvature, cannot sufficiently represent these complex classes; therefore, regional land surface parameters (LSPs) were selected to describe features in a broader hydrological and topographic context. Models were trained for each of the five learners, both with and without the addition of regional LSPs, and their performance was tested with a dataset of field-mapped terraces and manually-delineated landforms. Model complexity was not significantly related to accuracy; however, distance-based learners were significantly less accurate than others. The addition of regional LSPs was crucial to achieving acceptable accuracy, highlighting the importance of variable selection in building useful models. Overall, this research demonstrates the potential for using machine learning as a tool for exploiting high resolution data and supporting efficient field mapping.