| Author: | Jiang, Ge |
| Title: | Maximum entropy-based airborne LiDAR point cloud classification |
| Advisors: | Yan, Wai Yeung (LSGI) Lichti, Derek (LSGI) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Remote sensing -- Data processing Optical radar Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Land Surveying and Geo-Informatics |
| Pages: | xx, 150 pages : color illustrations |
| Language: | English |
| Abstract: | Airborne Light Detection And Ranging (LiDAR) systems demonstrate the capability to collect 3-dimensional (3D) point clouds and facilitate various topographic and earth observation applications. To support these applications, high-quality point cloud and effective and accurate 3D point cloud classification become essential and critical. This thesis contributes to the algorithmic development of airborne LiDAR point cloud classification in the following aspects, including point cloud denoising, homogeneous neighbor selection for traditional point cloud classification, and optimal neighbor selection and novel graph pooling for deep learning point cloud classification, all built upon the principle of maximum entropy. Airborne LiDAR data inevitably suffer from noisy returns floating in the empty space between the system and ground. Though various point cloud denoising methods are proposed based on sparsity assumption and elevation, they are more likely unable to simultaneously remove both clustered and scattered noisy points and suffer from thresholds adjustment, which limits the generalization to other point cloud scenarios with noise. Accordingly, we propose a Maximum Entropy-based Outlier Removal (MEOR) method to produce a global-local strategy considering the terrain fluctuation and search for elevation thresholds to distinguish noisy points and valid points, which is a threshold-insensitive strategy. The comprehensive experimental results demonstrate that MEOR significantly outperforms the other existing denoising methods in terms of removing noisy points and preserving valid points. In airborne LiDAR point cloud classification, existing methods often extract contextual features of point cloud by querying a fixed scale/number of neighbor points or selecting a variable size neighborhood based on certain optimality criterion. These approaches tend to introduce heterogeneity from the local neighborhood or select insufficient contextual features, hindering the performance of classification. Therefore, we propose an homogeneous neighbor selection method based on the Maximum Entropy (MaxEnt), which determines the homogeneity of the local neighborhood for each point based on MaxEnt and constructs geometric and radiometric features from selected homogeneous neighbor point. The generated contextual features then serve as the input of various machine learning classifiers for point cloud classification. Extensive experimental results demonstrate that MaxEnt shows the priority to achieve better classification results compared with other neighbor selection methods and makes full use of the merit of multispectral intensity data. Although traditional feature extraction and machine learning-based methods are effective in classifying airborne LiDAR point cloud, it is indeed not flexible to be applied to all kinds of topographic scenes. Therefore, a novel deep learning-based point cloud classification method by exploiting Maximum Entropy is presented in this research. The architecture directly extracts sufficient optimal Homogeneous neighbor points for individual points and designs a novel designed graph Pooling layer to encapsulate the selected neighbor points into small-size but informative graphs to build hierarchical features (MEHPool). The model includes a Homogeneous Neighbor Selection (HNS) network, two Graph Pooling (GP) layers, and three graph neural networks. The experimental results indicate that our proposed method consistently outperforms the other related deep learning methods. Both the HNS and GP modules make contribution to the classification performance and the combination of two modules outperforms either of them exploited alone. The proposed three maximum entropy-based algorithms, i.e., MEOR, MaxEnt, and MEHPool, presented in this thesis are effective in processing and analyzing airborne LiDAR point cloud data and outperform other previously published methods, which validates the effectiveness of the maximum entropy in point cloud analysis. Also, all the derived research certainly contributes to the on-going development of AI-based point cloud processing software bundle, e.g., preprocessing, information extraction functions, etc. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13381