LAUSR

Abstract Point cloud classifiers have the potential to rapidly perform landscape characterization for a variety of applications. The generalization (i.e., transferability to new sites) of such classifiers could improve their accessibility and usefulness for both engineers and researchers alike, but guidelines for classifier generalization are lacking in the literature. This study develops and applies a Random Forest machine learning classifier for Terrestrial Laser Scanning (TLS) point clouds, and generalizes the classifier to point clouds from several different locations. The classifier is trained to identify basic hillslope topographic features, including vegetation, soil, talus, and bedrock using multi-scale geometric features of the point cloud. Four rock and soil slopes in western Colorado were scanned using TLS. Generalization experiments were performed testing point density, occlusion, and between-site domain variance factors, and all factors showed a significant influence on generalization accuracy. Several methods for improving classifier generalization accuracy were tested and compared, including combining training data from multiple sites, imposing probability thresholds, and a Domain Adaptation methodology known as Active Learning. It was found that incorporating data from multiple sites resulted in improved generalization accuracy, but in most cases the largest improvements in accuracy were associated with adding new training data from the target site. In this case, using Active Learning resulted in significant accuracy improvements with an over 90% reduction in the number of added training points. The results suggest that scanning characteristics are important factors in classifier generalization accuracy, but their effects can be mitigated by using the techniques described herein.