Study On Object-oriented Classification And Individual Tree Crown Extraction Based On High Resolution Imagery

Remote sensing technology has become a new vegetation information monitoring means,with its some advantages such as rapid,scientific,and small outside workload.Forestry workers combine interpretation results of remote sensing images with field survey data to complete a variety of tasks for forest resource monitoring,including estimation of stock volume,dynamic detection of land use changes,tree species identification,estimation of canopy density,pest and disease monitoring,crown recognition and so on.In recent years,the field of forestry remote sensing has developed rapidly.Not only do data source become more diversified and higher spatial resolution,but the interpretation method has been innovated too.A lot of achievements have been made in the research field and the actual production process.Based on QuickBird image of Yanqing in Beijing,this paper explores the object-oriented classification method based on the optimal scale.And based on this method,the individual tree crowns extraction research were carry out with QuickBird image and GF-2 image of study area.The research strive to provide reference for the project of forest resources monitoring.Firstly,the optimal scale range is obtained by analyzing the maximum area of the object,and then combined with the segmentation quality evaluation model to determine the most optimal segmentation level.Specific research work is as follows:(1)Study of object-oriented classification method based on optimal segmentation scale:The segmentation experiment is carried out by combining the maximum area method and the segmentation quality evaluation model.Firstly,the optimal scale range is obtained by analyzing the maximum area of the object,and then,and then,combining the results of the segmentation quality evaluation model,the optimal segmentation level is determined.On the basis of this,according to the comprehensive analysis of the features of spectrum and texture of samples,the classification rules are formulated and object-oriented land cover type classification is finally completed.The result shows that the classification method based on the optimal multi-scale hierarchical and rules has obtained better classification results,its overall accuracy is 88.8%,and the Kappa coefficient is 0.861.While the overall accuracy of the nearest neighbor method based on single scale is 81.4%and the overall accuracy of the method based on rules and single scale is 83.2%.(2)Individual tree crown extraction based on optimal scale:Taking the QuickBird image and the GF-2 image as the research object,the object-oriented multi-scale segmentation method is used to extract the individual tree crowns of some classes images in different canopy density level,and the whole and local extraction results are compared with those crown delineation map of seed point area growth method and watershed segmentation algorithm.According to the visual results and analysis,it is found that the tree crown extraction results based on object-oriented multi-scale segmentation method and is better than the other two methods.According to field survey data and artificial interpretation results,we verify the accuracy of GF-2 image extraction based on multi-scale segmentation method,including individual tree level and stand level.Through the verification of 10 plots,the following conclusions are obtained:①There is a linear relationship between the number of automatic extraction trees and the number of detected trees in the field survey,and the average number of trees was satisfied with the formula "auto=0.9402manual-5.1476" and the correlation coefficient is 0.8483;②The validation results of individual tree level shows that overall classification accuracy is good,and the plots with 0.6 canopy density obtain the highest overall accuracy,which reach 85%.And the phenomenon of leakage and misjudgment will rise with the canopy closure;③After the extraction accuracy analysis of the average crown diameter,it is found that the crown diameter from automatic extraction is slightly smaller than that of from field measurement.Besides,because of the miscarriage of justice,there is a phenomenon that the extraction crown is greater than the actual crown in some plots with large canopy.