欢迎访问林业科学,今天是

林业科学 ›› 2022, Vol. 58 ›› Issue (9): 48-59.doi: 10.11707/j.1001-7488.20220905

•   • 上一篇    下一篇

高郁闭度人工林无人机激光雷达单木分割方法优化

朱泊东,罗洪斌,金京,岳彩荣*   

  1. 西南林业大学林学院 昆明 650224
  • 收稿日期:2021-06-17 出版日期:2022-09-25 发布日期:2023-01-18
  • 通讯作者: 岳彩荣
  • 基金资助:
    国家自然科学基金项目(42061072);云南省科技厅重大科技专项(202002AA100007-015);云南省教育厅项目(2018JS330)

Optimization of Individual Tree Segmentation Methods for High Canopy Density Plantation Based on UAV LiDAR

Bodong Zhu,Hongbin Luo,Jing Jin,Cairong Yue*   

  1. Forestry College, Southwest Forestry University Kunming 650224
  • Received:2021-06-17 Online:2022-09-25 Published:2023-01-18
  • Contact: Cairong Yue

摘要:

目的: 针对高郁闭度林分条件下基于LiDAR点云数据单木分割林木提取困难、总体精度较低等问题,提出一种基于冠层起伏率结合分水岭算法和基于点云的局部最大值聚类算法的分层分割法,为开展无人机LiDAR技术森林资源调查提供技术参考,为提高单木分割总体精度提供新策略。方法: 利用无人机激光雷达数据,采用分水岭算法、基于点云的局部最大值聚类算法和基于冠层起伏率结合分水岭算法和基于点云的局部最大值聚类算法的分层分割法对高郁闭度思茅松人工林进行单木分割,并分析分水岭算法中4种CHM空间分辨率和3种DSM插值方法对单木分割效果的影响,与无人机高分辨率影像单木树冠目视解译结果进行比较,以探测率r、准确率pF得分为指标对单木分割精度进行验证和评价。结果: 在幼龄林中,冠层起伏率较大,分水岭算法对单层林的分割效果优于基于点云的局部最大值聚类算法;在中龄林和近熟林中,冠层起伏率较小,分水岭算法易将思茅松树枝识别为树冠,基于点云的局部最大值聚类算法的分割效果优于分水岭算法;基于冠层起伏率结合分水岭算法和基于点云的局部最大值聚类算法的分层分割法充分考虑不同龄组的林分结构差异,精度最高(F = 0.75),优于分水岭算法(F = 0.71)和基于点云的局部最大值聚类算法(F = 0.68);在分水岭算法中,当分辨率为0.5 m×0.5 m时采用反距离权重法(IDW)插值得到的CHM单木分割精度最高(r = 0.70,p = 0.94,F = 0.81)。结论: 林分结构存在差异,单一单木分割方法效果欠佳,通过提取样地冠层起伏率确定分水岭算法和基于点云的局部最大值聚类算法单木分割适用的林分条件,可拓宽单一单木分割方法在不同林分条件下的优势,提升单木分割精度。

关键词: LiDAR, 单木分割, 分水岭算法, 基于点云的局部最大值聚类算法, 分层分割

Abstract:

Objective: UAV LiDAR technology has become an important technical means to obtain refined forest parameters. Individual tree segmentation using LiDAR point cloud data is the basis for extracting refined forest parameters, and the accuracy of forest parameters acquisition depends on whether the individual tree segmentation is accurate or not. To address the problems that existing individual tree segmentation algorithms is difficult to extract trees under high canopy cover stand conditions with low overall accuracy, this study proposes a layered segmentation method based on CRR (canopy relief ratio), combined with watershed algorithm and the local maximum clustering algorithm based on point cloud to provide a new strategy to improve the accuracy of individual tree segmentation. Method: The ULS(unmanned aerial vehicle laser scanning) data were used to segment the individual tree of high canopy cover Pinus kesiya var. langbianensis plantation using the watershed algorithm, the local maximum clustering algorithm and the stratified segmentation method based on CRR proposed in this paper, and the effects of the watershed algorithm on the individual tree segmentation result under 4 CHM spatial resolutions and 3 DSM interpolation method were analyzed separately for the watershed algorithm, and the result of the visual interpretation of individual tree crown from UAV high-resolution images were used as validation, with detection rate r, accuracy p and F-score as evaluation indexes. Result: In young forest, the CCR is higher, and the segmentation accuracy of the watershed algorithm is better than that of the local maximum clustering algorithm for single-layer forest. In middle-aged and near-mature forest, the CCR is lower, and the watershed algorithm can easily identify Simao Pine branches as canopy, and the segmentation effect of the local maximum clustering algorithm is better than that of the watershed algorithm. The segmentation method based on CCR combined with the watershed algorithm and the local maximum clustering algorithm has the highest accuracy (F=0.75). The layered-segmentation method based on CCR fully considers the differences in stand structure among different age groups, and the layered-segmentation method had the highest accuracy (F=0.75), which is better than watershed algorithm (F=0.71) and local maximum clustering algorithm based on point cloud (F=0.68). Secondly, in the watershed algorithm, when the resolution is 0.5 m × 0.5 m, the CHM individual tree segmentation accuracy obtained by interpolation using the inverse distance weighting (IDW) is the highest (r = 0.70, p = 0.94 and F = 0.81). Conclusion: The accuracy of using only one individual tree segmentation algorithm is often unsatisfactory due to the differences in stand structure. The combination of the watershed algorithm and the local maximum clustering algorithm can expand the applicability of a single segmentation method under different stand and thus improve the accuracy of individual tree segmentation by determining the respective applicability range through CCR.

Key words: LiDAR, individual tree segmentation, watershed algorithm, local maximum clustering algorithm based on point cloud, hierarchical segmentation

中图分类号: