关键词: LiDAR, 人工林, 单木分割, 分水岭算法, 四次多项式拟合法, 基于点云的距离判别聚类法

Abstract: [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 is 0.5 m×0.5 m, 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.

Key words: LiDAR, planted forests, individual tree segmentation, the watershed algorithm, polynomial fitting method, point cloud-based cluster segmentation