面向激光雷达点云数据的多结构树种识别

Multi-structured tree species recognition for LiDAR point cloud data

针对由于树木种间相似性和种内差异性带来的识别困难,以及由于采集环境及设备的多样性导致的点云质量差异,提出面向激光雷达点云数据的多结构树种识别方法(MSTSR)。首先借助改进的组合采样策略,在有效降低数据冗余的同时,保留单木的主体枝干结构;其次通过内建的近邻感知与增强模块(NAE)层次化聚合点云属性,以形成高阶的语义描述;最后通过融合树冠、主干以及整树的多结构信息,生成跨尺度的树木点云表征。在地面激光雷达采集的树种点云数据集上验证该方法的有效性,该数据集由7个树种共690棵树组成的。结果表明:该方法的总体准确率达到94.2%。相比主流的PointNet和PointNet++深度点云分类网络,分别提升13.04和9.42个百分点;相比基于点云的多视图2D投影方法,提升8.19个百分点;相比基于多个测树因子的随机森林方法,提升24.63个百分点,从而证实采用深度网络直接进行树种点云识别的潜力。

Aiming at the difficulty of tree identification caused by the similarity between species and the difference between species,as well as the difference of point cloud quality caused by the diversity of collecting environment and equipment,a multi-structured tree species recognition method(MSTSR)based on LiDAR point cloud data was proposed.Firstly,a combined sampling strategy was designed to effectively reduce data redundancy while preserving the trunk and main branches of a single tree.Then,a built-in neighborhood awareness and enhancement(NAE)module was devised to hierarchically aggregate point cloud attributes into high-level semantic descriptions.Finally,three types of information extracted from the crown,trunk and entire tree were fused to generate the cross-scale representation.The effectiveness of the method was verified on a point cloud dataset consisting of 690 trees of seven tree species acquired by terrestrial LiDAR.The results demonstrated that the method s overall accuracy(OA)reached 94.2%.Compared with mainstream deep learning methods for point cloud classification,such as PointNet and Point Net++,the improvement was 13.04 and 9.42 percentage points,respectively.In addition,the proposed method was improved by 8.19 percentage points compared with the multi-view 2D projection method,and improved by 24.63 percentage points compared with the random forest method using multiple tree measurement factors.These results confirmed the potential of the deep point cloud network for tree species recognition.