Research On Extracting Urban Buildings From VHR Aerial Images And Detecting Changes

With the development of remote sensing technology and deep learning technology,machine learning with deep learning as the mainstream has gradually begun to replace the method of manual visual interpretation to provide the possibility of automatic extraction of ground features in remote sensing images;At the same time,with the rapid development of urbanization and the rapid growth of urban buildings,the extraction and change detection of buildings in aerial images provide great convenience for the automatic detection of land supervision,reconstruction or demolition of buildings after natural disasters such as earthquake and fire.High spatial resolution(HR)and ultra-high spatial resolution(VHR)remote sensing images have more feature information and smaller objects that can be identified,which is beneficial to deep learning algorithm models to extract deep-level structural features and improve the accuracy of feature detection.However,at present,there are few two-phase aerial image building data sets based on HR and VHR,which increases the workload of building change detection in remote sensing images;At the same time,the traditional methods have low accuracy and inaccurate edges when detecting remote sensing images Question,this paper has the following research content:(1)Summarize the public data sets related to HR and VHR remote sensing image building change detection summarized in the research process of this paper,to provide convenience for the relevant research scholars of urban building change detection.(2)VHR aerial image building extraction based on DeepLab v3+ model.Because buildings are the main part of urban land types,with the largest area and the most active changes,this paper applies the hollow convolutional image segmentation model(DeepLab v3+)based on the structure of coding and decoding to large-scale VHR WHU building data set to achieve the integrated extraction of the segmentation and classification of urban buildings,the experiment shows that the DeepLab v3+ model can achieve a building extraction accuracy of 89%;At the same time,the Sobel edge detection algorithm is used to further optimize the edge of the building extraction.,the experiment shows that this paper uses the method of edge detection algorithm to optimize building extraction accuracy can retain part of building edge information.(3)VHR aerial image building change detection based on DeepLab v3+ model.The buildings in the two periods after the earthquake were extracted separately,and the extraction results of the two-phase buildings were superimposed on the Sobel edge detection results and further processed to obtain the final building change detection result map.The experiment shows that based on the DeepLab v3+ model and Sobel edge detection method are feasible in VHR aerial image building change detection.