Study On Urban Built-up Area Information Extraction Using High Resolution Remote Sensing Image Based On Deep Learning

With the development of China’s social economy and the continuous expansion of the scale of urbanization,we need to have an objective and comprehensive grasp of the comprehensive level of urban land use and regional development.The urban built-up area is important basic information for urban regional research,and it is also the premise of the implementation of regional planning and the layout of urban functional space.In recent years,with the development of earth observation technology and the improvement of remote sensing images resolution,it has become possible to more accurately and effectively extract the information of urban built-up areas.High spatial resolution remote sensing image is an important research direction in the field of earth observation because of its large amount of information,wide coverage and clear image.The environment of the urban built-up area with high-resolution remote sensing images is complex,and there are many differences in the location and development scale of the urban built-up areas,so it is difficult to extract the information.Aiming at the information extraction of urban built-up areas from high-resolution remote sensing images,this thesis applied deep learning technology to GF-2 remote sensing image,and made intensively research and analysis.The specific research contents are as follows:(1)This thesis introduced the related theory of earth observation technology and deep learning,including the features of high-resolution remote sensing images,the development and current status of deep learning technology,and the current status of research on information extraction from remote sensing images.The feasibility of applying deep learning technology applied to high-resolution remote sensing images was analyzed in detail,and the advantages of deep learning in information extraction were pointed out.(2)For high-resolution remote sensing images,a series of processes from original image preprocessing to automatic information extraction were realized.Radiometric calibration,atmospheric correction,ortho-rectification,image fusion and standardized image cutting were performed on the original remote sensing image to obtain an image that is as close as possible to the real situation of the earth’s surface.The processed image data was used in the subsequent algorithm research,and also better reflected the features of the urban built-up area.(3)This thesis implemented the information extraction of high-resolution remote sensing images of urban built-up areas based on convolutional neural networks.In order to solve the problems of using deep learning algorithms that are prone to over-fitting and low robustness,data augmentation and enhancement technology was adopted to further enhance the network model capabilities.In the high-resolution remote sensing image of 10000?10000 size,mainly in the Qiongshan District,Haikou City,Hainan Province,the accuracy rate of the trained deep learning algorithm is more than 88% for the information extraction of urban built-up areas.However,this method had limited ability to extract the deep features of remote sensing images,and the detailed information of the features of the extraction results was insufficient.(4)To achieve accurate extraction of GF-2 remote sensing image information based on deep learning,the expression ability of convolutional neural networks needed to be strengthened.By using the refinement module for the feature map and the attention module in the channel domain,this thesis improved the fully convolutional neural network,so that the network model could effectively learn and select features that contribute significantly to the information of the urban built-up area,and improve the accuracy of extraction of the remote sensing image information.In this thesis,sliding window prediction and fully connected conditional random field post-processing methods were used to optimize the results of the improved convolutional neural network to more accurately achieve the extraction of urban built-up areas.The results show that it is feasible to extract feature information from high-resolution remote sensing images using deep convolutional neural networks for image semantic segmentation.Compared with traditional classification method maximum likelihood and machine learning method random forest method,deep learning technology can get better results,but its feature extraction ability can be improved.By enhancing the information extraction capabilities of feature maps and channel domains,using sliding window prediction and fully connected conditional random field post-processing methods,the improved convolutional neural network more accurately extracted the information of urban built-up areas in high-resolution remote sensing images.In general,the thesis focuses on the application of extensive deep learning algorithms,which is widely used in the field of computer vision,to achieve the extraction of information on urban built-up areas on large-scale high-resolution remote sensing images.The ultimate goal is combining deep learning technology and remote sensing information processing to provide an efficient method of extracting information from urban built-up areas for the government and other research institution to better serve national well-being,people’s livelihood and social development.