[Back]

N. Li:
"Contextual Semantic Classification of ALS Point Clouds in Urban Environment";
Supervisor, Reviewer: N. Pfeifer, G. Vosselman, N. Haala; Department of Geodesy and Geoinformation, TU Wien, 2021; oral examination: 2021-12-20.

ALS (Airborne Laser Scanning)/Airborne LiDAR (Light Detection And Ranging) is characterized by its capability of providing highly accurate and dense 3D point clouds at a city-wide scale. Therefore, ALS has been widely used for a variety of applications in urban environments, which makes the automatic classification of 3D point clouds a crucial task. Urban environments are a complex combination of both built-up and natural objects. Consequently, automatically producing highly accurate classification results from ALS point clouds is challenging. The commonly used machine learning methods such as Random Forest and Support Vector Machine often lead to noisy classification results, since they rely heavily on the input of handcrafted features and lack the consideration of spatial context. In order to improve classification performance, the aim of this dissertation is to incorporate context into the classification of ALS point clouds. The work in this dissertation comprises methodological developments presented in publications I-III, an investigation of the cutting-edge deep learning methods, in which contextual features can be directly learned in the training phase, presented in publication IV, and an effort on extending training data that is presented in publication V.Publications I-II propose a high-dimensional tensor-based sparse representation for the classification of ALS point clouds. This novel data structure is introduced to keep the handcrafted features in their original geometric 3D space, such that the spatial distribution and handcrafted features can be considered at the same time. Using only a few training samples, promising classification results can be obtained. Publication III develops a label smoothing strategy to refine classification results. Without the aid of additional training data, the proposed label smoothing strategy can directly learn context information from initial classification results by estimating an adaptive neighborhood and a probabilistic label relaxation. Experiments exhibit its strength in improving classification accuracy. Publication IV conducts a comprehensive comparison between three state-of-art deep learning models, namely PointNet++, KPConv, and SparseCNN, w.r.t. classification accuracy, computation time, generalization ability, and the sensitivity to the choices of hyper-parameters. Publication V proposes a method to extend training data by selecting the most informative samples from the neighborhood of the initial training samples.After an overview of the publications, a comparison between all investigated methods is carried out on two separate ALS datasets in urban areas. Based on the results, the superior performance of the selected deep learning models, especially SparseCNN, is further confirmed. The proposed label smoothing also turns out to be advantageous, regardless of the diversity of scenes and point densities involved. More importantly, it is independent of different training efforts, which have a profound impact on deep learning methods. The works conducted in this dissertation provide practical examples and valuable insights into the contextual classification of ALS point clouds. The presented studies also show that supervised machine learning, especially deep learning, heavily depends on training data. Therefore, the automated generation of training data by active learning could be a way to further facilitate the classification of point clouds.

https://doi.org/10.34726/hss.2022.36642