| As an active element of urban development,buildings are one of the focuses in the fields of land resource management,urban planning,geographic database updating,urban and rural construction and urban security.Efficient and accurate building information extraction and change detection have important theoretical significance for these businesses.In recent years,with its excellent generalization ability,deep learning is particularly suitable for the analysis and expression of massive big data,and has been widely used in various fields.The method based on deep learning,provided new ideas for extracting and detecting specific information in high-resolution remote sensing images.Aiming at the problems of missing detection and rough boundary recognition in the process of building change information extraction and detection by deep learning algorithm,this paper established an AUNet-CRFs building extraction model to extract all buildings in a remote sensing image.And the Siam-UNet++ building change detection model was established to detect the changed buildings in two remote sensing images with different time sequence,so as to facilitate the management of staff.The detailed research contents were as follows:(1)Preprocessing of the remote sensing image data..This paper selected WHU building aerial remote sensing data,Google building data and self-collected Google data in the Qinling area of Xi'an as the input of the deep learning network model.Firstly,the self-collected Google data in Qinling area of Xi'an city was processed with histogram matching,image registration,image denoising,normalization and annotation to supplement Google data source and enhance its data diversity.Then the three kinds of data were sampled and augmented to get the data set needed by this task.2)In order to solve the problems of complex mapping relationship of deep features in building information extraction by deep learning algorithm,such as false detection,missing detection and rough building edge extraction,a deep learning network model based on AUNetCRFs was proposed for high-resolution remote sensing images.Inspired by U-Net,the network model used the expansion path of the atrous convolutional layer structure to reshape the deepest receptive field of the U-Net network.By expanding the receptive field,the correlation between building features was improved,and then the feature extraction accuracy was improved,and the rough extraction of buildings was realized.The fully connected random field was applied to improve the edge segmentation of buildings,so as to realize the fine extraction of buildings Fine extraction.(3)In view of the fact that the input mode of early fusion(EF)of dual time sequence image was easy to ignore the feature information of image change,a Siam-UNet++high-resolution remote sensing building change detection model was proposed.Using UNet++as the backbone extraction network,the Siam-diff(Siamese-Difference)structure was used in the encoder part to extract the change characteristics of the two sequential images,and the attention mechanism was introduced after the up sampling and horizontal jump path connection in the decoding stage.so as to highlight the change features of buildings and inhibit the learning of other change features by the network;Meanwhile,the Multiple Side-Output Fusion(MSOF)strategy was used to weight and fuse feature information of different semantic levels,which improved the accuracy of building change detection;finally,a sliding window method was adopted to predict large-scale remote sensing images,reducing the hole pattern generated by the change result map during the splicing process.The experimental results demonstrate that our algorithm shows better performance than other models.The experimental verification results on a large-scale building data set show that the proposed deep learning network model has good performance for building extraction.And change detection,and improves the accuracy of building recognition to a certain extent. |
PDF Full Text Request