Research On Building Extraction And Change Detection In High Resolution Remote Sensing Images Based On Deep Learning

?High-resolution remote sensing images can provide rich feature information,which is an important tool for acquisition building information.Among them,the extraction of buildings can quickly understand the distribution characteristics of urban space,and the change detection of buildings also provides analysis conditions for the planning and layout of social economy.Scientific,accurate and timely study of building information from images is of great significance to urban planning and development,dynamic changes of cities and disaster prevention.In the traditional methods of building extraction from high-resolution remote sensing images,most of them rely on a priori knowledge and manual feature extraction,and the methods of manual recording and field investigation require personnel with relevant experience and knowledge,which have the problems of high cost and low efficiency.The building extraction model based on deep learning technology can not only quickly obtain the information of different time periods in the whole city area,but also adapt to the development and demand of the digital era.Therefore,this paper conducts research experiments on building extraction and change detection based on deep learning of high-resolution remote sensing images,and the main work is as follows:(1)For the current problem that the quality of deep learning training building dataset is not high and the dataset is not fully utilized in the experimental process,this paper carries out the enhancement processing of building dataset.The utilization of the dataset is improved mainly by cropping,filling,panning,rotating and eliminating the data with large blank areas.(2)To address the problems of blurred boundaries and insufficient integrity in the current building extraction results,a building extraction method based on context fusion and edge preservation is proposed.The method extracts building features at different scales by fusing parallel roads that do not interfere with each other in the convolutional neural network,and adopts an attention mechanism to weight the features extracted by the convolutional neural network to obtain a feature map with both detail information and global information to achieve accurate building extraction.Based on the open source dataset Massachustetts building dataset and WHU building dataset,this study achieves 94.22% OA score and 94.04%OA score,respectively.Compared with the mainstream U-Net network extraction results improved by 1% and 6% on average.(3)The above building extraction method is used to achieve building change detection for individual detection error results that occur in building groups discontinuously in building change detection results.The proposed method achieves an Io U score of 78.63% on the LEVIR-CD bi-temporal remote sensing image dataset,which improves the detection results compared to commonly used change detection networks,and also validates the accuracy and applicability of the architecture extraction method based on context fusion and edge preservation.