Research On Plantation Canopy Model Segmentation Method Based On UAV Image Matching Point Cloud Data

Forest resources as a very important natural resource has an irreplaceable important position,however,with a large number of published national measures to protect natural forest,only relying on natural forest resources fail to meet the growing market demand of forest resources nowadays.So,it is inevitable for China’s forestry industry development to cultivate efficient and high-yield artificial forest to replace natural forest resources.China,as the largest country in the world in the total amount of plantation,invests countless human and material resources in the investigation of plantation resources every year,then it is of great significance to find a more real-time,cheap,convenient and accurate investigation scheme of plantation resources.Nowadays,the methods of forest resources investigation can be divided into artificial,satellite remote sensing,lidar and UAV remote sensing.However,the data collection efficiency of manual survey is low,and delayed;the amount of satellite remote sensing information is large and the acquisition cycle is long,which is only suitable for large-scale and long-term forestry survey;the lidar has large information and the visual effect is good,but the cost is expensive.UAV,as an emerging technology,has been quite mature in canopy surface information acquisition,but it lacks of three-dimensional information.It is difficult to obtain accurate information of tree height and canopy morphology by ordinary ground survey in that the intense competition of plantation growth,which is also a major difficulty in the investigation of plantation resources.It has no potential to be popularized widely because of its low popularity and weak practicability,although airborne lidar can solve this problem well.Good characteristics of low cost,portability and accuracy leads to be a trend to use UAV multi view matching algorithm instead of expensive laser equipment to produce point cloud data for forestry resources investigation,as the development of UAV tilt photography technology.The development of an efficient and accurate point cloud segmentation algorithm from UAV matching point cloud data for single tree crown can solve the problem that other survey methods can not accurately obtain the height and canopy shape information of single tree in today’s environment,which open up a new idea for the investigation of plantation resources.To solve the above problems,this paper obtains high-definition remote sensing images of artificial forest land which based on UAV tilt photogrammetry technology,and using multi view image matching method to obtain the corresponding dense matching point cloud data.But large amount of image matching point cloud data existed in it and contained a large number of isolated points and noise points,we use a new point cloud noise reduction filtering algorithm to preprocess the data.The point cloud single tree segmentation and tree height extraction are carried out through deep learning algorithm and the accuracy is compared,after obtaining the point cloud data.There are some main research results:(1)Study the point cloud matching algorithm.Through the combination of SIFT algorithm,SFM algorithm and PMVs algorithm,the dense point cloud data of experimental forest land is obtained by multi view matching from high-definition images of multi angle UAV,which solving the problem that the point cloud data obtained by traditional feature point matching is not dense.It can more accurately reflect the forest information in good degree.(2)Propose a novel point cloud filter.While a single filter fail to achieve the desired effect,a new filter is proposed.The new filter combines the advantages of multiple filters,meanwhile it has the functions of fast segmentation,thinning and ground point segmentation.After filtering the initial point cloud data,the initial 638301 and 385735 points are reduced to 410219 and 262059 points,reducing about 35.5%of the invalid point cloud data.It not only removes the outliers and noise points,but also highlights the morphological structure,which is convenient for subsequent segmentation and extraction.(3)Propose a deep learning point cloud segmentation algorithm based on local feature enhancement of PointNet network.In point cloud segmentation,we improve the structure of the classic deep learning point cloud segmentation algorithm PointNet,use R nearest neighbor algorithm to enhance the ability of local feature extraction,and increase the convolution kernel to make it more suitable for local segmentation.The segmentation results are compared with the segmentation method based on seed point region growth and mean shift algorithm the recall rate and accuracy of deep learning algorithm are 89.5%and 91.4%respectively,which are significantly higher than other algorithms.Afterwards,we extract the tree height of the point cloud model extracted by the three algorithms and compare it with the manual measurement data.The results also help to prove that deep learning algorithm in favor of canopy point cloud model segmentation.