机载激光雷达人工林单木分割方法比较和精度分析

Comparisons and Accuracy Assessments of LiDAR-Based Tree Segmentation Approaches in Planted Forests

【目的】研究分水岭算法、四次多项式拟合法和基于点云的距离判别聚类法对人工林单木分割的适用性,分析3种方法对人工林单木分割的精度,探索进行单木分割时3种方法关键参数的最优选择。【方法】结合地面实测数据和目视解译方法,计算单木探测率、准确率和F得分,比较分水岭算法、四次多项式拟合法和基于点云的距离判别聚类法的单木分割精度,并通过改变栅格化冠层高度模型(CHM)的分辨率及调整基于点云的距离判别聚类法的距离阈值,分别对3种方法进行单木提取效果的敏感性分析。【结果】1)分水岭算法、四次多项式拟合法和基于点云的距离判别聚类法对人工林单木总体分割精度较高(F=0.76～0.83); 2)对于"复杂林型"样地,基于点云的距离判别聚类法的分割精度最高(F=0.78),优于分水岭算法(F=0.74)和四次多项式拟合法(F=0.53);对于"中等复杂林型"样地,基于点云的距离判别聚类法的分割精度最高(F=0.89),优于分水岭算法(F=0.84)和四次多项式拟合法(F=0.75);对于"简单林型"样地,基于点云的距离判别聚类法(F=0.89)、分水岭算法(F=0.89)和四次多项式拟合法(F=0.93)的分割精度都较高; 3)敏感性分析结果表明,当CHM分辨率为0.5 m×0.5 m时,分水岭算法和四次多项式拟合法的分割精度最高;当基于点云的距离判别聚类法的距离阈值近似样地平均冠幅半径时,其分割精度最高。【结论】对多种类型样地进行单木分割,体现了分水岭算法、四次多项式拟合法和基于点云的距离判别聚类法对人工林单木分割的适用性;结合多种类型样地充分评估并比较了3种方法对人工林单木分割的精度;通过对3种方法进行敏感性分析,阐述了进行单木分割时关键参数的最优选择。

【Objective】This paper studies the applicability of the watershed algorithm,polynomial fitting method and Point cloud-based cluster segmentation for individual tree segmentation,analyzes the accuracy and explores the optimal selection of the key parameters of the three methods for individual tree segmentation.【Method】The field measured and visual interpretation data were combined to calculate the individual tree detection rate,precision of detected trees and overall accuracy index.In addition,the grid canopy height model(CHM,canopy height model)resolution of the watershed algorithm and polynomial fitting was changed and the distance threshold of point cloud-based cluster segmentation was adjusted to perform the sensitivity analysis of individual tree extraction.【Result】The result showed that:1)The three segmentation methods used to segment individual trees in planted forests have relatively high overall accuracy(overall accuracy F=0.76-0.83).2)For“complex forest type”samples,point cloud-based cluster segmentation has a higher extracting accuracy(overall accuracy F=0.78)than the watershed algorithm(overall accuracy F=0.74)and polynomial fitting(overall accuracy F=0.53);for the“moderately complex forest type”samples,point cloud-based cluster segmentation has a higher extracting accuracy(overall accuracy F=0.89)than the watershed algorithm(overall accuracy F=0.84)and polynomial fitting(overall accuracy F=0.75);for the“simple forest type”samples,point cloud-based cluster segmentation(overall accuracy F=0.89),the watershed algorithm(overall accuracy F=0.89)and polynomial fitting(overall accuracy F=0.93)have similar precisions.3)Sensitivity analysis result showed that when the CHM resolution is0.5m×0.5m,the watershed algorithm and the polynomial fitting segmentation accuracy has the highest accuracy,whereas when the threshold approximately equals to the average of the crown projection radius,the point cloud-based cluster segmentation reaches the highest precision.【Conclusion】The individual tree segmentation of multiple types of plots reflects the applicability of the three methods to the planted forests.The accuracy of individual tree segmentation of planted forest by three methods is fully evaluated and compared with many types of plots.The sensitivities of the three methods were analyzed,and the optimal choice of key parameters during individual tree segmentation was described.