机载激光雷达单木识别研究进展

Review on Individual Tree Detection Based on Airborne LiDAR

随着激光雷达的发展,基于机载激光雷达提取单木及林分参数是目前的研究热点之一。准确的单木识别是后续林木参数提取的重要基础。机载激光雷达单木识别方法可以分为基于冠层高度模型(CHM)的单木识别法和基于点云分布的单木识别法两类。基于CHM的单木识别方法通过CHM分割确定树冠边界或通过局部最大值识别树冠顶点并且进行区域生长或图像分割。基于点云分布的单木识别法在三维空间上采用区域生长或聚类算法识别树冠。分析不同方法在单木识别中的优缺点,对比不同单木识别法对单木识别精度、欠分割误差、过分割误差的影响。分析数据类型、点云密度、季节和林木生长状况等多个影响识别精度的因素,分析可得全波形数据优于离散回波识别精度,点云数据密度10pt/m2即可满足单木识别要求,冬季识别精度优于夏季识别精度。探讨机载激光雷达数据的局限性及其在单木识别中的缺陷,从数据获取时间、获取方式及类型、数据组织管理、多源数据融合、多种识别算法综合应用、机器学习增加训练集寻找最优模型等方面展望了未来单木识别的发展方向,拓宽我国森林资源调查及相关领域的研究思路。

As light detection and ranging（LiDAR）develops,the extraction of forest structure parameters has been one of hot topics in related fields in the past years.However,the accuracy of detection is the key factor in obtaining the forest individual tree parameters.The individual tree detection methods can be divided to two types：one is based on the canopy height model（CHM）and the other is based on the point cloud distribution.We can identify an individual tree by using the method of the crown boundary segmentation.Also,we can identify the tree top by local maximum algorithms and then perform the regional growth or image segmentation.Based on the point cloud distribution,the canopy is identified by region growing or clustering algorithms in three-dimensional space.We analyze the advantages and disadvantages of different individual tree detection methods in terms of precision of individual tree detection,and compare their effects on omission errors and commission errors in different regions.The factors influencing the precision of data such as data type,point cloud density,season and tree growth status are discussed.It is found that the accuracy of the full-waveform data is higher than that of discrete-echo data.Theobtained in winter is higher than that in summer.The limitation of airborne LiDAR data and its shortcomings in individual tree detection are discussed.In the end,the future directions of individual tree detection are described,from the aspects of data acquisition type,data acquisition time,data organization and management,multi-source data fusion,comprehensive application of multi-detection algorithms,and machine learning increasing the training set to find the optimal model,to help with the research and management of forest and related fields.