Object Detection of Tree Uprooting Signatures in Gridded LiDAR Elevation Data

Abstract Tree uprooting (or tree throw) is a natural forest disturbance with consequences for ecology, pedology, hydrology, geomorphology, and atmospheric fluxes. These events leave topographic signatures in the form of pit‐mound couplets (PMCs) that persist for decades to centuries. Therefore, the land surface serves as an archive of tree uprooting events. High resolution topographic data is capable of resolving many of these pit‐mound couplets, allowing us to access this archive. We have trained an object detection model, YOLOv5 (You Only Look Once), to automatically map PMCs in USGS 3D Elevation Program topographic data. The model achieves performance metrics of 0.76 for F1, 0.72 for precision, and 0.8 for recall. We field‐validate positive inferences and compare the predicted PMCs to those that we map in 0.25 m resolution LiDAR topographic data. We also provide a brief analysis of PMCs in Brown County, IN that demonstrates the spatial density of PMCs as a function of hillslope aspect and slope.