LEVERAGING MACHINE LEARNING FOR REMOTE CLAST SIZE IDENTIFICATION IN POST-WILDFIRE DEBRIS-FLOW DEPOSIT MATERIALS

Post-wildfire debris flows are a growing hazard in the Western U.S. due to the increased annual areal extent of wildfire linked to climate change. There are relatively few observations of debris-flow grain-size distributions. This limits modeling efforts and hazard assessments that rely on properties related to the grain-size distribution (such as friction angle). Additionally, these properties are difficult to measure in a lab environment due to the large fraction of cobble-sized (6.4 cm) and larger particles. To improve our understanding, we collected a sub-centimeter resolution terrestrial lidar scan of a uniquely well-preserved debris flow deposit associated with the Grizzly Creek Fire in Glenwood Canyon, Colorado. These remotely sensed topographic data are leveraged to develop a semi-automated workflow for clast identification and grain-size distribution calculation. First, a training and testing dataset is created by performing point-level classification of matrix material or clasts larger than cobble size in the lidar point cloud. This analysis is a preliminary step towards automatically delineating the grain size distribution of a deposit.