摄影测量学
稳健性(进化)
点云
计算机科学
遥感
激光雷达
斜格
分割
树(集合论)
人工智能
计算机视觉
激光扫描
图像分割
工作流程
杠杆(统计)
地质学
随机森林
数据挖掘
Boosting(机器学习)
航空影像
变更检测
全球定位系统
空间分析
作者
Xuan Wang,Xianfeng Huang,Hanyu Xiang,Fan Zhang
标识
DOI:10.1109/tgrs.2025.3608289
摘要
Individual tree crown delineation (ITCD) in urban areas is essential for various downstream applications, such as refined urban modeling, urban forest inventory, and individual tree management. Thanks to the soaring development of the Unmanned Aerial Vehicle (UAV)-based photogrammetry technique, swift 3D data collection is plausible in the city scale. However, a fully automatic and precise ITCD method is desired for the photogrammetry point cloud data. Previous ITCD approaches have faced significant scale inadaptability and low robustness problems, especially for the scenario in this paper. We introduce a novel Normal Direction Accumulation Tree Center Detection (NDACD) method to fully leverage the rich side information from the oblique photogrammetry data, extending beyond just the top-of-canopy perspective. A multiscale segmentation workflow is proposed to automatically adapt to the diversity of crown scales across different tree species. Additionally, the multiscale segmentation workflow is guided by high-level shape priors, such as convexity and tightness, to further enhance the robustness to complex tree crown structures. We tested our method on a dataset of more than 700 trees collected by oblique photogrammetry in Shenzhen and Wuhan, China, demonstrating its effectiveness and robustness with low-quality real-world data. Compared to other methods, our approach achieves a higher ITCD accuracy, with an Acc of 91. 2% and an average IoU of 86.9%. Furthermore, it shows good adaptability to alternative data sources, including drone and mobile LiDAR data.
科研通智能强力驱动
Strongly Powered by AbleSci AI