Challenges in Approaching the Detection Limits for Hillslope Erosion Using Terrestrial Laser Scanning

Accurately quantifying soil loss due to water erosion is a critical step in managing soils. Terrestrial LiDAR (Light Detection and Ranging) presents a potential alternative to traditional soil loss measurement by estimating soil erosion and deposition through detecting surface changes. Terrestrial LiDAR can also provide spatial distribution information without disturbing the observed surface. While erosion estimation through terrestrial LiDAR detects large magnitude erosion well, the finer temporal/spatial scale erosion experienced on the hillslope in sheet and rill erosion has remained a challenge to detect. This research addresses two of the challenges in using terrestrial LiDAR on fine scales in two manuscripts. Chapter One presents Las2DoD, a new method for quantifying uncertainty in surface change analysis by operating directly on point clouds produced by terrestrial LiDAR. The theory of Las2DoD is given and supported by a case study of two erosion plots in comparison to two existing uncertainty analysis methods. The methods are compared with collected sediment delivery data to evaluate the effects of the uncertainty analysis methods in the context of measuring soil erosion. Las2DoD generated the most accurate estimate on both plots, capturing 90% and 65% of the measured sediment delivery, with the second-best performing method measuring 70% and 48%. Las2DoD also managed to preserve more low-magnitude changes relative to the other methods, which is particularly important when measuring small spatiotemporal scale changes. In the second chapter, terrestrial LiDAR scans of an erosion plot while bare and vegetated are differenced to produce a very dense testing dataset. This dataset is used to assess the performance of three point cloud filtering algorithms, namely, a cloth simulation algorithm, a slope-based filter, and a random forest classifier. The methods are tested against a subset of the testing dataset using different sets of input parameters to identify optimum parameters. The highest-performing implementations of the filtering methods are then applied to the entire testing dataset and compared with one another. The cloth simulation filter achieved the best classification, though the classification was highly influenced by parameter values that differed from established recommendations