Single Tree Extraction Of Airborne Forest Point Cloud Based On Deep Learning Of Candidate Regions

Forest resources are one of the most important resources in the world,and investigating them is an important basic work in forest management and protection.Therefore,accurate single tree extraction plays an important role in forest surveys.At present,single tree extraction methods based on airborne Li DAR point clouds mainly use manual design rules to find seed points in the rasterized canopy height model(CHM)or point cloud,and perform subsequent segmentation based on the found seed points.However,this type of method There are problems such as insufficient rule universality leading to low feature performance,design rules incapable of making full use of point cloud spatial information,and difficulty in extracting the lower tree due to the difficulty of obtaining the lower tree due to the obscuration of the upper tree.In order to solve the limitations of the current method,this paper proposes the idea of automatically extracting single tree candidate frames for single tree extraction.By drawing lessons from the deep learning target detection method,design a single tree extraction deep learning model to achieve accurate single tree extraction results.This paper uses public data sets to conduct theoretical verification and experimental analysis on the single tree extraction network.Aiming at the traditional CHM-based single tree extraction accuracy limited by the selection of seed points,this paper applies Faster RCNN to CHM for single tree extraction.The network automatically establishes a single tree candidate frame to establish an outsourcing frame for each location where a tree may exist.,And extract the outsourcing box containing the tree to complete the single tree extraction,which solves the problem of single tree extraction relying on manual rules to design seed points.When CHM is established,the rasterization causes the lower tree to be covered by the upper tree,so it is more conducive to the single tree extraction to establish a model that directly extracts the single tree from the point cloud.Because the forests are all similar tree point clouds and the trees are relatively dense,the current existing 3D target detection deep learning network cannot be directly applied.This paper draws on the idea of Faster RCNN and proposes a deep learning model for single-tree extraction directly from the point cloud.Firstly,various shape candidate regions are generated by gridding the point cloud.Secondly,use the Point Net++ network to extract the features of the candidate area to obtain the feature of the candidate area.Then,according to the characteristics of the candidate area,it is judged whether it is a single tree,and the classification loss function is constructed.Finally,the frame regression is performed on the candidate area containing the single tree,so that the candidate area is close to the real outer frame,and the single tree extraction is realized.Experimental results show that Faster RCNN’s single-tree extraction accuracy on CHM reaches 70.93%,which is more than 4% higher than the overall accuracy of other traditional CHM-based methods.This shows that the method of automatically establishing a single-tree candidate frame is relative to the manual design rule single-tree extraction method Advantages.In addition,the single tree extraction accuracy of the single tree extraction network based on the point cloud of the three-dimensional candidate area is72.75%,which is more than 1% higher than the overall accuracy of the single tree extraction network based on CHM.This model solves the problem that the lower trees cannot be effectively segmented in the single tree extraction network based on CHM,and proves the rationality and effectiveness of the single tree extraction network based on the point cloud of the threedimensional candidate area.Using different regional data sets for testing and verification,the final average accuracy is 66.65%,which is higher than the single tree extraction network method based on CHM.It proves that the single tree extraction network based on deep learning proposed in this paper has a certain generalization ability and universality,and provides new research ideas and theoretical support for the research of single tree extraction in forest areas.