基于无人机激光雷达点云数据的单木分割研究

Research on single tree segmentation based on UAV LiDAR point cloud data

【目的】探索不同树种在样地和单木尺度上无人机激光雷达点云数据的单木分割效果,选取哈尔滨城市林业示范基地阔叶林(水曲柳)和针叶林(樟子松)两块样地为研究对象,对样地内树木点云进行单木分割并评价其分割效果,为后续单木结构参数的提取提供数据支持,同时丰富森林资源信息的调查手段。【方法】通过无人机激光雷达获得样地树木点云数据,然后分别采用改进的K均值聚类算法和基于相对间距的阈值分割算法对水曲柳和樟子松样地进行单木点云数据分割。其中,水曲柳样地点云数据处理采用改进的K均值聚类算法,通过树干点云位置推算初始聚类中心,减少因树冠重叠导致的错误分割;樟子松样地点云数据处理则采用基于相对间距的阈值分割算法,通过设定多条阈值规则并利用动态最大值滤波器对树顶进行精确探测,提高算法的分割精度。最后,基于样地和单木点云完整度两个方面对点云数据分割效果进行评价。【结果】1)从样地尺度来看,水曲柳和樟子松样地单木识别率分别为0.91和0.87,相应的调和值(F)分别为0.91和0.88,结果显示单木分割的整体效果较好。水曲柳样地召回率(r)为0.87,精确率(p)为0.95,算法分割过程中产生的错误分割较少。樟子松样地算法分割的单木也多为正确分割(r=0.82、p=0.94),其分割误差主要来自于单木欠分割,过分割现象相对较少。2)从单木点云分割的完整程度来看,水曲柳样地的单木平均正确分割率为75.6%,点云平均欠分割率为24.3%,平均过分割率为18.5%,单木点云的最大错误分割率为31.8%,不同单木之间分割精度有较大差异;樟子松样地的单木分割精度较稳定,平均正确分割率达84.1%,平均欠分割和过分割率分别为16.3%和9.0%,表明单木点云不存在大量错误分割的情况。【结论】基于树木形态结构特征改进了两种优化单木点云数据分割算法,两种算法下分割的森林样地单木精度在不同评价尺度中表现均较好,对林中树木出现的树冠重叠、遮挡、偏移等现象均有一定的辨别能力,实现了森林样地树木点云数据的单木精确分割。

【Objective】In order to explore the single tree segmentation effect of different tree species on the plot and single tree scale of UAV LiDAR point cloud data,two plots of broad-leaved forest(Fraxinus mandshurica)and coniferous forest(Pinus sylvestris)in Harbin Urban Forestry Demonstration Base were selected as the research objects.Single tree segmentation on the tree point cloud in the sample plot was performed and segmentation effect was then evaluated,providing data support for the subsequent extraction of single timber structure parameters as well as enriching the investigation methods of forest resources information.【Method】The point cloud data of the trees in the sample plot was obtained by UAV LiDAR,and then two different improved algorithms were used to segment the single tree point cloud data of the plots of Fraxinus mandshurica and Pinus sylvestris.Among them,an improved k-means algorithm was adopted to process the cloud data of Fraxinus mandshurica plot,which calculated the initial clustering center through the position of the tree trunk point cloud,reducing the false split due to crown overlap.A threshold segmentation algorithm based on relative distance was used to process the cloud data of the Pinus sylvestris plot.By setting multiple threshold rules and using dynamic maximum filter to detect the tree top accurately,the segmentation accuracy of the algorithm was improved.Finally,the effect of point cloud data segmentation was evaluated based on the sample plot and the integrity of the single tree point cloud.【Result】(1)In terms of plot scale,the single tree recognition rates of Fraxinus mandshurica and Pinus sylvestris were 0.91 and 0.87,and the corresponding F values were 0.91 and 0.88,respectively.The results showed that the overall effect of single tree segmentation was better.Among them,the recall rate r of the Fraxinus mandshurica plot was 0.87,and the accuracy rate p was 0.95.There were fewer errors in the segmentation process.While the single trees segmented by the algorithm in the Pinus sylvestris plot were mostly correctly segmented(r=0.82,p=0.94),and the segmentation errors mainly came from the under-segmentation of the single tree,and the phenomenon of over-segmentation was relatively rare.(2)From the perspective of the integrity of single tree point cloud segmentation,the average correct segmentation rate of single trees in the Fraxinus mandshurica plot was 75.6%,and the average under-segmentation rate of the point cloud was 24.3%,and the average over-segmentation rate was 18.5%.The maximum wrong segmentation rate was 31.8%,and the segmentation accuracy varied greatly among different single trees;the single tree segmentation accuracy of the Pinus sylvestris plot was relatively stable.The average correct segmentation rate was 84.1%,and the average under-segmentation and over-segmentation rates were 16.3%and 9.0%,respectively,indicating that there was no large number of wrong segmentations in the single tree point cloud.【Conclusion】In this study,two optimized single tree point cloud data segmentation algorithms were improved based on the morphological and structural characteristics of trees.The accuracy of the single tree segmentation of sample plots under the two algorithms performed well in different evaluation scales and had a certain ability to distinguish the phenomenon of tree crown overlap,shelter,offset and so on in the forest,which achieved the precise segmentation of single tree point cloud data in forest plots.