Generalized machine-learning-based point cloud classification for natural and cut slopes

Abstract

Processing and interpretation of 3D point cloud datasets is often a limiting factor for their use in geohazard engineering projects, prompting a growing interest in supervised Machine Learning (ML) algorithms to automatically extract objects of interest. Objects might include rock outcrops, vegetation, components of vegetation, or other natural features. However, ML methods are subject to several well-known limitations that require a significant degree of expertise on the part of the user to address. Two key issues include the “generalization gap”, which describes the often significant performance difference between validation and testing, and the danger of ascribing a degree of infallibility and superior performance to black-box ML algorithms where none exists in reality. Geological engineering studies to date have not considered these issues in any detail, despite their critical importance for successful, justified models where lives could be at stake. This thesis provides a comprehensive critical evaluation of using ML-based techniques to interpret point clouds of natural and cut slopes consisting of a variety of materials, focused on the two issues mentioned above. A first-of-its-kind database of 12 manually annotated lidar and photogrammetry datasets is compiled, and ML models are developed to accurately classify lidar and photogrammetry derived point clouds. Next, a variety of tests are performed to empirically quantify the generalization gap. Finally, the framework is applied to a landslide monitoring case study, where ML is used to automatically extract tree trunks. Generalization test results show that many factors contribute to the generalization gap, including feature engineering, lighting conditions, geologic setting, geomorphology, point density, and occlusions. Due to these factors, in some cases up to 50\% or more of possible model configurations produce unacceptable classification results. However, it is shown that making informed choices while building and interpreting a classifier can greatly improve the likelihood of success. Thus ML is not a panacea, but it is undoubtedly a valuable addition to the geological engineer’s toolbox.