Using LiDAR Topographic Data and Machine Learning Techniques to Identify Near-Surface Soil Saturation for Improved Environmental Planning-Scale Wetland Mapping

Wetlands are important ecosystems that are threatened by agricultural and development repurposing, pollutant runoff, and climate change. Accurate and widely-available wetland inventories have the potential to support wetland conservation and environmental planning entities. The growing collection of remote sensing data has created new opportunities for wetland identification, and Light Detection and Ranging (LiDAR) data, specifically, has been widely embraced by the wetland science community. This dissertation aims to develop an open source wetland identification tool that leverages LiDAR elevation data and machine learning techniques to identify likely wetland areas at an environmental planning scale. The designed wetland identification tool is implemented and evaluated across four study areas in Virginia that encompass a range of ecoregion, built environment, and topographic characteristics. Key components of the wetland tool are developed and refined through three studies. The first study focuses on the identification and evaluation of LiDAR topographic metrics as indicators of near-surface soil moisture, using a Random Forest model. The second study focuses on the effects of alternative hydrologic terrain processing methods on wetland predictions and the Random Forest model used to generate them. The third study evaluates the potential for using deep learning for identification of wetlands from images that represent LiDAR-derived geomorphic characteristics, with relatively limited training data resources. Key research findings are as follows. i) The topographic wetness index, curvature, and cartographic depth-to-water index are successful wetland indicators for a range of landscapes, but there is potential to improve their abilities to distinguish wetted areas from dry uplands through site-specific modifications. ii) By applying a sophisticated LiDAR DEM preprocessing workflow, these topographic indices are better able to model wetlands and, therefore, considerably improve prediction accuracy for all sites studied. iii) Accurate wetland predictions can be produced using a basic deep learning architecture and imagery composed of the topographic indices and complementary vegetative information, despite being limited to training data sets that are significantly smaller relative to most deep learning applications. With the completion of model refinement through this research, the wetland tool has the potential to benefit the broader environmental planning and conservation community.