Object-oriented Building Information Extraction Based On High Resolution Image

With the rapid development of high-resolution satellites,high resolution image provides richer information,and becomes important data sources for information extraction.The process of urbanization is getting faster and faster in our country,buildings as a core feature of the city are constantly changing,so the timely update of urban spatial databases is more and more important.How to use high-resolution images to automatically,efficiently and accurately extract building information becomes one of the hot spots in remote sensing applications.Traditional pixel-based image classification has been unable to meet the needs of high-resolution image information extraction due to its limitation,such as simple classification based on spectral information,low classification accuracy,and the existence of "pepper and salt"phenomena.The object-oriented classification based on eCognition has gradually become the mainstream method of information extraction.This paper takes the WorldView-II image of Hangzhou as an example to study the application of object-oriented classification method in high-resolution image building information extraction.The main research content is as follows:(1)Combining ESP optimal scale evaluation method with PSE-NSR-ED segmentation quality assessment method to determine the optimal segmentation scale.The study first determines the optimal homogeneity factor,and then initially selects the optimal segmentation scale through the ESP tool.Finally,determines the optimal segmentation scale of different class by the segmented quality assessment method.The optimal segmentation scale is 420,300,260,220,the shape factor is 0.5,and the compactness is 0.5.(2)Paper adopts mRMR feature optimization algorithm to select the optimal classification feature subset.Research selects 47 features such as spectrum,shape,texture and so on for image objects,then uses the mRMR algorithm for maximum correlation and minimum redundancy optimization,finally identifies 10 optimal classification features.(3)Analyzing optimal segmentation scales and optimal classification features,then establishing classification hierarchy and determining the classification rules for different ground objects.Considering the characteristics of high-resolution images with multiple shadows,this paper sets up contrast experiments based on the presence or absence of shadows and evaluates accuracy through the confusion matrix method and inconsistency accuracy evaluation method.Finally,the paper optimizes the results of building extraction by mathematical morphology.(4)Accuracy evaluation shows that both experiments performed well in building extraction.In Experiment one,the classification overall accuracy is 86.5%,Kappa is 0.798,the building extraction accuracy is 85%,the integrity is 74.76%,and the detection rate is 84.67%.In Experiment two,the classification overall accuracy is 91.01%,Kappa is 0.866,the building extraction accuracy is 95%,the integrity is 86.98%,and the detection rate is 92.07%.The experimental results show that based on the spatial relationship between shadows and buildings,the building extraction accuracy is higher.The research of building information extraction technology based on high-resolution image provides important ideas for information extraction of urban buildings.It also provides reference for object-oriented image classification method.