Multi-structured tree species recognition for LiDAR point cloud data
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.
tree species recognitionLiDARpoint clouddeep learning
万方数据
维普
