Research On Urban Single Tree Detection And Parameters Extraction Based On UAV Image And Mask R-CNN

As one of the most important land surface resources,trees play an important role in the ecosystem,such as air purification,wind prevention and sand fixation,and climate regulation.Timely and effectively detecting trees and obtaining information on various parameters of trees play a vital role in monitoring the growth of trees.The various parameters of a single tree include crown area,crown width,tree height and diameter at breast height,etc.Traditional methods of obtaining tree information are usually time-consuming and laborious.But the emergence and development of UAV platforms provide a new data source channel for tree parameter extraction in city.This research is based on the remote sensing image of ginkgo biloba by UAV,using the Mask R-CNN algorithm in deep learning combined with digital orthophoto maps to detect and segment the ginkgo biloba tree crown in different scenes in the city,and automatically obtain the crown width,the crown area,and the diameter at breast height.This research mainly carried out the following aspects:(1)Obtain remote sensing images of ginkgo biloba through UAV and generate digital orthophoto maps and 3D point cloud data.Based on the digital orthophoto map,the measured values of the crown width and crown area of the ginkgo biloba were extracted by visual interpretation,and the tree height were extracted based on the 3D point cloud data.(2)In this research,the UAV tree crown remote sensing image is preprocessed and data labeled,and the Mask R-CNN network model is trained through migration learning,so that the network model is suitable for detecting and segmenting ginkgo canopy in different scenes in the city.And in order to explore the impact of different image types and numbers of dataset training models on the effect of tree crown detection and segmentation,this research produced seven different training datasets.The research results show that the OBL-90 dataset training model had the best detection effect,and the overall F1-score reached 91.66%.According to the detection and segmentation results,the crown width and crown area parameters were extracted,and the extracted crown parameters were compared with the actual measured values obtained through visual interpretation.The average relative error of crown width and crown area is 7.5% and 11.15%.It shows that the combination of UAV image and Mask R-CNN algorithm can effectively and automatically extract ginkgo biloba tree crown parameters in different scenes in the city.(3)Substituting the crown width and crown area parameters extracted by the above method into the three binary DBH inversion models(crown width & crown area-diameter at breast height)made by the tree parameters in the training area to predict tree diameter at breast height.And compared with the corresponding actual measurement of DBH,the binary linear model U2 DBH has the best DBH prediction accuracy,with an average relative error of 9.37% and a comprehensive error of 0.107,achieving a good DBH prediction effect.Add the tree height parameters extracted from 3D point cloud data to make a three-dimensional diameter at breast inversion model(crown width & canopy area &tree height-diameter at breast height).Substituting the tree crown parameters into the inversion model to predict tree diameter at breast height and doing comparative analysis.The ternary power function model P3 DBH has the best DBH prediction effect with an average relative error of 8.24% and a comprehensive error of 0.092.In addition,the ternary inversion model generally has higher prediction accuracy than the binary inversion model,and can predict the DBH parameters with higher accuracy.The research results show that based on automatically extracted tree crown parameters and the DBH inversion model,the DBH of ginkgo biloba can be obtained accurately.