Detection And Identification For The Urban Forest At Tree Level Based On Convolutional Neural Network And Google Earth Images

Urban forest occupies an important position in the urban ecosystem.It can not only beautify the urban environment,but also help protect biodiversity and maintain ecological balance.Effective urban forest management is the basic requirement to ensure its sustainable development.Traditional urban forest management usually requires a large amount of materials and labor to conduct field surveys,or use hyperspectral remote sensing data to detect surface vegetation for data statistics.These methods are often limited by data source acquisition and data quality.Google Earth displays global geospatial data with an intuitive three-dimensional perspective,which can effectively monitor the coverage of urban forests and collect tree information.The method of acquiring data is faster and more efficient,which can largely avoid the waste of manpower and time,and thus can substitute the traditional urban forest data acquisition method to a certain extent.The Mask RCNN(Mask Regions with Convolutional Neural Network)is an advanced algorithm in Convolutional Neural Network series.It performs well in image recognition,especially in target detection of high-resolution images,and has good stability and detection accuracy.Therefore,this research uses Mask RCNN for modeling and testing to detect and recognize tree crowns on Google Earth images.This method applies the target detection method in deep learning to urban forest research,and proposes new ideas and supervision models for the statistics and management of trees in urban forestry,which can provide necessary data for biomass estimation and carbon sequestration research in the urban ecological environment.This method has great practical significance for urban forest management.This study uses Central Park in New York of the United States as the study area,which is a typical representative of urban forests.Using the Google Earth image data set of Central Park,the Mask RCNN network model can independently learn the characteristics of the urban forest canopy,and realize the automatic identification of the single canopy of the urban forest and the statistics of the number of trees.The accuracy of the model detection is verified by comparing the samples with artificial naked eyes and Google Street View images.The research results show that based on the Mask RCNN method,combined with Google Earth images,it is possible to build an urban forest canopy recognition model and successfully detect the number of urban forest canopies and the area of a single canopy.The main research conclusions are as follows:1.Based on the Mask RCNN,using selected parts(about 10%)of Central Park Google Earth sub-images as the training set,the model automatically learns the urban forest canopy features to realize the automatic detection and recognition of the urban forest canopy.Put all the Google Earth sub-images of Central Park as a test set into the model,and output all the canopy information of Central Park,including the position,boundary and area of the canopy.The overall accuracy of the model in identifying tree categories was 89.7%.A total of 5491 independent trees were identified,and the total canopy area detected was 975,252 square meters.2.This method can detect the canopy information of urban forests accurately.In the randomly selected sub-image samples,the detection error of the model is obtained based on the number of tree crowns recognized by the naked eye and the tree crown area calculated manually,and the overall crown missing rate and the crown area error of the model are estimated to be 15% and 3.27%,respectively.Based on this error correction,the number of trees and the canopy area of the Central Park in New York of the United States are estimated to be 6,443 and 996,745 square meters,respectively.Overall,this model can provide a new method for tree information statistics in urban forest management.3.Google Street View can effectively assist this research in conducting a simulated field survey of canopy area.Taking the tree canopy area in a certain sub-image(No.130)as an example,based on the sample observation of Google Street View,we concluded that the average error between the model detection value of the tree canopy area and the street view observation value is about 27.76%,of which about 80% of the model estimates are slightly higher than the observed values of Street View.Overall,this model provides a new and cost-effective method for tree survey and statistics in urban forestry,provides a reference for the further application of deep learning in the forest field in the future,and provides a scientific basis for the study of urban ecosystem service functions.