Urban Cinnamomum Camphora Crown Detection Research Using DOM-DSM Images And Deep Learning

As an important part of urban ecosystem,urban trees are of significant ecological value.It is indispensable for the management and operation of urban trees to trees’ species and spatial distribution.In order to quickly and accurately get these,this study combined remote sensing technology and deep learning to detect individual tree crowns,which providing technical support for the management and operation of urban trees.At present,in the research of individual tree crown detection based on deep learning,researchers relied on the spectral information of ground objects to detect tree crowns,while the elevation information was ignored.Therefore,this thesis combined spectral information and elevation information of ground objects to detect individual tree crowns in urban,so as to explore whether the combination of these two kinds of information can bring gain to the task performance.This thesis mainly carried out the following parts of work.Firstly,two Cinnamomum camphora crown data sets were constructed: DOM(Digital Orthophoto Map)data set and DSM(Digital Surface Model)data set,which respectively represent the spectral information and elevation information of ground objects.Secondly,FPN-Faster-R-CNN and Yolov3 were respectively trained and tested in two data sets.In both networks,DOM data set achieved higher average accuracy than DSM data set.The result indicate that DOM has richer feature information than DSM for individual tree crown detection task.Thirdly,band combination and dual-source detection network architecture were used to combine spectral information and elevation information to perform individual tree crown detection task.In both networks,the dual-source detection network architecture achieved the best results,while two of the three schemes in the band combination achieved higher average accuracy than the DOM scheme.The result indicate that proper combination and utilization of two kinds of information can improve the performance of urban individual tree crown detection task.For dual-source detection network architecture,three weight strategies were used respectively,and the average accuracy only fluctuated very little,which indicate that the three weight strategies have similar effects.Fourthly,based on the principle of dual-source detection network architecture,a individual tree crown detection system combining DOM and DSM was constructed to improve the task performance and realize automation.