本研究提出基於深度學習的方法對點雲進行分類並應用於數值高程模型的產製流程當中。透過訓練以幾何特徵為依據的點雲分類器，從空載光達（LiDAR）點雲中自動而準確地擷取點雲中的地面點，大幅減少傳統數位高程模型（DEM）生成流程中所需之點雲後處理之人力與時間成本。
本研究的主要貢獻體現於兩方面: 首先，本研究提出了一種針對點雲資料的前處理採樣流程，旨在確保提取地面點特徵的準確性和完整性，在此過程中儘可能保留了點雲資料中地面點和非地面點的空間幾何特徵；其次，本研究採用基於集成學習的策略提升了地面點分類器的分類能力。鑑於台灣空載光達點雲資料中，每一幅點雲所包含的地形特徵類型較多，本研究強調了針對不同地形類型建立特化模型的重要性，透過使用集成學習技術結合這些特化模型，從而建構了一個端對端(end-to-end)的分類器。該策略提高了地面點提取的準確性，同時增進了DEM生成流程的整體效能。在點雲地面點分類上，整體精確度(accuracy)在不同的地形上皆可達到至少80%。在DEM產製方面，本論文針對參考DEM與基於地面點分類器產製之DEM進行高程差比較，在不同地形的DEM模型中，有80%以上的圖幅面積之高程差落在± 20公分內。
透過實驗與性能分析，本研究驗證了所提出的自動化方法的正確性。進一步與基於傳統空載光達點雲數據生成的DEM進行比較，更突顯出基於深度學習方法的優勢，特別是在分類準確性和減少人力需求方面。本研究對於利用空載光達點雲數據生成DEM的技術進步做出了貢獻，並為應對台灣獨特的地形樣態提出了有效的解決方案。

This thesis presents a deep-learning-based approach to generate digital elevation models (DEMs) from LiDAR point clouds while reducing labor-intensive editing required in postprocessing. The proposed method utilizes a point-based deep learning classifier to extract ground points from Taiwan's Airborne Laser Scanning (ALS) data and generate DEMs based on the separated ground points. This research contributes in two primary ways. Firstly, a sampling workflow was proposed to ensure the accuracy and integrity of the extracted ground points while preserving the unique features of the point cloud data. Secondly, the research emphasizes the significance of specialized models for different terrain types. By leveraging ensemble learning techniques, the proposed method combines the strengths of these specialized models, enabling the construction of an end-to-end classifier. This approach improves the accuracy of ground point extraction and contributes to the overall effectiveness of the DEM generation process. The traditional workflow of DEM generation is time-consuming, which may take days to generate a DEM from a point cloud. With this approach, same quality DEMs with accuracy of greater than 80% can be generated within less than a day. Comparing the DEM generated by proposed method with the reference DEM, over 80% of coverage in point cloud from different dataset exhibits an elevation difference within the error standard of ± 20 centimeters.
Through comprehensive experimentation and performance analysis, the proposed method's effectiveness and efficiency are demonstrated. A comparative analysis with traditional LiDAR DEM generation methods highlights the advantages of the deep learning-based approach, particularly in terms of classification accuracy and reduced labor requirements. This research contributes to the advancement of LiDAR DEM generation techniques and provides valuable insights for optimizing the process in Taiwan's unique terrain conditions.

[1] J. R. E. Villanueva, L. I. Martínez and J. I. P. Montiel, "DEM Generation from Fixed-Wing UAV Imaging and LiDAR-Derived Ground Control Points for Flood Estimations," Sensors, p. 14, 2019.
[2] G. Sammartano and A. Spano', "DEM Generation based on UAV Photogrammetry Data in Critical Areas.," in Geographical Information Systems Theory, Applications and Management, 2016.
[3] M. Soilán, B. Riveiro, J. Balado and P. Arias, "Comparison of heuristic and deep learning-based methods for ground classification from aerial point clouds," International Journal of Digital Earth, pp. 1115-1134, 2019.
[4] C. R. Qi, H. Su, K. Mo and L. J. Guibas, "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation.," Fourth International Conference on 3D Vision, pp. 601-610, 2016.
[5] A. Nurunnabi, J. L. F. N. Teferle1, R. C. Lindenbergh and A. Hunegnaw, "An Efficient Deep Learning Approach for Ground Point Filtering in Aerial Laser Scanning Point Clouds," Remote Sensing and Spatial Information Sciences, pp. 31-38, 2021.
[6] F.-D. W. G.-S. X. Wuzhao Lia, "A geometry-attentional network for ALS point cloud classification.," ISPRS Journal of Photogrammetry and Remote Sensing, pp. 26-40, 2020.
[7] A. Diab, R. Kashef and A. Shaker, "Deep Learning for LiDAR Point Cloud Classification in Remote Sensing.," Sensors, p. 20, 2022.
[8] Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein and J. M. Solomon, "Dynamic graph CNN for learning on point clouds," ACM Transactions on Graphics, 2019.
[9] V. Svetnik, T. Wang, C. Tong, A. Liaw, R. P. Sheridan and Q. Song, "Boosting: An Ensemble Learning Tool for Compound Classification and QSAR Modeling," Journal of Chemical Information and Modeling, pp. 786-799, 2003.
[10] C. R. Qi, H. Su, K. Mo and L. J. Guibas, "PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space," in Computer Vision and Pattern Recognition, 2017.
[11] J. Niemeyer, F. Rottensteiner, U. Soergel and C. Heipke, "Hierarchical higher order crf for the classification of airborne lidar point clouds in urban areas.," Int. Arch. Photogramm. Remote Sens. Spatial Inform. Sci. 41, 655–662., 2016.
[12] Y. X. U. S. R. Huang, "GraNet: Global relation-aware attentional network for semantic segmentation of ALS point clouds," ISPRS Journal of Photogrammetry and Remote Sensing., pp. 1-20, 7 2021.
[13] X. Wang and S. Ji, "Roof Plane Segmentation From LiDAR Point Cloud Data Using Region Expansion Based L0 Gradient Minimization and Graph Cut," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, pp. pp. 10101-10116, 2021.
[14] J. Niemeyer, F. Rottensteiner and U. Soergel, "Contextual classification of lidar data and building object detection in urban areas," ISPRS Journal of Photogrammetry and Remote Sensing, pp. 152-165, 2014.
[15] D. P. Kingma and J. L. Ba, "Adam: A Method for Stochastic Optimization.," in International Conference on Learning Representations, 2015.
[16] T.-Y. Lin, P. Goyal, R. Girshick, K. He and P. Dollár, "Focal Loss for Dense Object Detection," in IEEE International Conference on Computer Vision, 2017.
[17] Huang, Faliang, Xie, Guoqing and R. Xiao, "Research on Ensemble Learning," in International Conference on Artificial Intelligence and Computational Intelligence, Shanghai, China, 2009.
[18] G. Aman, R. Rohan, S. Jun and K. S. Sathiya, "Adam vs. SGD: Closing the generalization gap on image classification," in 13th Annual Workshop on Optimization for Machine Learning (OPT2021), 2021.
[19] A. C. Wilson, R. Roelofs, M. Stern, N. Srebro and B. Recht, "The Marginal Value of Adaptive Gradient Methods," arXiv:1705.08292, 2018.
[20] L. Luo, Y. Xiong and Y. Liu, "Adaptive Gradient Method With Dynamic Bound of Learning Rate," in International Conference on Learning Representations (ICLR), 2019.
[21] P. Zhou, J. Feng, C. Ma, C. Xiong, S. Hoi and W. E, "Towards Theoretically Understanding Why SGD Generalizes Better Than ADAM in Deep Learning," arXiv:2010.05627, 2020.
[22] 林緯程, 王驥魁, 林昭宏, 勞宏斌, 許育維, 王敏雄 及 湯凱佩, “基於幾何特徵以 UNet 分類空載光達地面點,” 航測及遙測學刊, 2023.