Research On Extraction Method Of Tree Measurement Factors Based On UAV Image And LiDAR

Forest resources are an important part of terrestrial ecosystem,which provide oxygen and abundant wood materials for human production and life.They are renewable natural resources.Therefore,the basic work for forest resources investigation and management is quickly and accurately obtaining the parameters of individual tree(such as tree height,DBH,crown width,etc.).Among the main methods of forest resources survey at present,the cycle of manual investigation is long and inefficient,usually takes the measure of years.Satellite remote sensing has a large amount of information,but is suitable for large-scale surveys.Li DAR not only has high precision but also has good visual effect,however,it is expensive and price limits its application.With the development of computer vision,the use of multi-view images for 3D reconstruction to generate point cloud data has gradually become a new way to obtain point cloud data.The use of point cloud data can effectively solve a series of problems related to forest resources investigation,reduce the time and steps of manual data processing,speed up the extraction of forest information,and reduce the required human resources and hardware costs.Focusing on the above problems and backgrounds,this paper selects the Mengjiagang Forest Farm in Jiamusi City as the research area,obtains remote sensing data based on UAV and backpack Li DAR,and extracts the parameters of individual tree in the forest land.The specific work content of the paper is summarized as follows:(1)According to the principle of camera imaging,UAV oblique photography is used for matching to obtain 3D point cloud instead of airborne Li DAR.The structure will be recovered from motion through feature point extraction and feature point matching.The process of generating point cloud from tilt photography is preliminarily completed.Dense matching of visual images realizes the expansion of sparse point cloud to dense point cloud,and the generated point cloud has good effect,which can clearly see the woodland features and canopy structure,so the problem that the point cloud is not dense due to the sparse characteristic of feature points has been solved.(2)The U-Net convolutional neural network is used to complete the research on the segmentation of individual tree crown in UAV remote sensing image,and further count the number of trees.At the same time,compared with the watershed algorithm and K-means clustering segmentation of traditional digital image processing,the segmentation results are evaluated by accuracy and recall.It is proved that the deep learning algorithm effectively reduces the over segmentation phenomenon when segmenting tree crown compared with the traditional image processing method.The crown contour can be extracted accurately.(3)An improved point cloud segmentation algorithm based on Point Net is proposed,which has better local feature extraction ability by modifying the original network structure and increasing convolution kernel.At the same time,the advanced and representative point cloud instance segmentation algorithm is deeply studied and applied to forestry data set.Comparing the segmentation results of three different deep learning point cloud segmentation algorithms based on improved Point Net,SGPN and 3D-Bo Net for multi-source point cloud data,the accuracy of improved Point Net is 86.88% and the recall of improved Point Net is 86.55%,which are higher than the other two methods.Therefore,the tree measurement factors are extracted based on its segmentation results.This result also proves that the deep learning algorithm can be used for forestry information extraction and has excellent performance.It has a good application prospect.(4)Furtherly,the parameters of individual tree will be extracted from the results of individual tree segmentation.Tree height is extracted through the difference between the highest point and the lowest point within the individual tree range.The DBH is extracted by fitting the circle with the least square method,and the error between the extracted value and the measured value is compared through linear regression.The results show that the error between the extracted value and the measured value is small and the correlation coefficient is high.The main technical methods proposed in this paper can be used to accomplish the investigation or rapid update of forest resources information.