基于UAV可见光遥感的单木冠幅提取研究

Individual Tree Crown Extraction based on UAV Visible Light Remote Sensing Technology

为优选出云南松单木树冠的提取方法,以无人机获取的典型天然云南松纯林可见光影像为基础数据源,采用多尺度分割、标记分水岭分割和基于均值漂移的超像素分割3种方法,提取单木长冠幅、短冠幅及树冠面积,与地面实测数据对比,分析不同分割方法对单木冠幅信息提取精度的影响。结果显示:在单木尺度和样地尺度上,单木分割效果最优者均为标记分水岭算法,其准确率分别为82.23%和83.86%;基于均值漂移的超像素分割效果次之,其准确率分别为77.17%和77.11%;多尺度分割效果最差,其准确率分别为66.45%和67.66%。在单木尺度上,标记分水岭算法提取的单木冠幅信息精度最高,长冠幅、短冠幅和树冠面积提取的R^(2)分别为0.84、0.71、0.94,平均相对误差分别为9.20%、8.28%、14.05%。在样地尺度上,标记分水岭的单木树冠面积提取精度最高,平均相对误差为20.57%。对于重叠、粘连的树冠,标记分水岭算法能有效解决单木分割中合并和过度分割的问题,且能够精确提取单木冠幅信息。

For optimizing the extraction method of individual tree crown of Pinus yunnanensis,canopy is the most direct part that could be observed by forestry low altitude UAV remote sensing.Based on the visible light image of typical natural Pinus yunnanensis pure forest obtained by UAV,three methods,multi-scale segmentation,marked watershed segmentation and super-pixel segmentation based on mean shift,were used to extract the long crown,short crown and crown area of individual tree.Compared with the measured data on the ground,the effects of different segmentation methods on the extraction accuracy of individual tree crown information were analyzed.The results showed that the marked watershed algorithm was the best on individual tree scale and plot scale,with an accuracy rate of 82.23% and 83.86%,the super-pixel segmentation based on mean shift was the second,with an accuracy rate of 77.17% and 77.11%,and the multi-scale segmentation was the worst,with an accuracy rate of 66.45% and 67.66%.On individual tree scale the R^(2) of Long crown,short crown and crown area were 0.84,0.71 and 0.94,and the average relative errors were 9.20%,8.28% and 14.05%,respectively.On plot scale,the extraction accuracy of individual tree crown area marked watershed was the highest,and the average relative error was 20.57%.For overlapped and conglutinated tree crowns,marker watershed algorithm can effectively solve the problems of merging and over segmentation in individual tree segmentation,and can accurately extract individual tree crown information.