| City is the place that owns the highest density of people throughout the country,and it is the area with the most frequent human activities.The way to develop and predict the future of the city,and make a long-term urban planning blueprint is one of the most important topics for the city builders and managers.Nowadays,the research on the classification of urban features and the prediction of future development has become the core problem of urban planning and decision-making,management and construction.Satellite remote sensing image can provide a clear urban area,and has many advantages,such as convenient,fast,low cost,multi time and so on.Convolutional neural network has the accuracy and adaptability for different data that other algorithms do not have in image processing,while instance segmentation network has higher recognition accuracy by adding target detection into traditional pixel-based convolutional neural network.In this paper,case segmentation network is used to classify urban features.After literature research,this paper studied the feasibility of deep learning case segmentation network for feature classification,and determined the scheme and technical framework of this research work.Firstly,the classification of surface features was established,and the corresponding data sets were obtained after processing the 10-meter resolution remote sensing image of sentinel-2 and landsat-8 image,and the sample labels were obtained by manual visual interpretation.In order to get more training samples,landsat-8 and other images were enhanced.Then,this paper carried out three groups of control variables contrast experiments,respectively,explored the best network for urban feature classification through the network's own parameter adjustment;explored the training results of different resolution image data sets;and explored the impact of multi band image on the training results.The results showed that the Mask R-CNN network with Res Net101-FPN skeleton has the best classification effect on sentinel-2 remote sensing image,and the multi band remote sensing image composed of normalized vegetation index and normalized architecture index could achieve the best classification results.In the process of result prediction,aiming at the problems of boundary fault and incomplete semantic information in the results,this paper proposed an optimization algorithm for the defects of convolution neural network prediction results to improve the recognition accuracy and improve the semantic information.By adding overlap algorithm,the IOU of buildings and vegetation could increase by about 2% and 7% respectively,and the accuracy of predicted results was improved.The gap repair algorithm proposed in this paper filled the gap through the correlation function in the Post GIS database,so that the small tiles could be truly spliced into a complete semantic information vector map,which could not only ensure that the quality of the image in the prediction process is not affected,but also completed the task in a short time when the amount of data is very large.Experiments showed that the algorithm had a significant optimization effect in large amount of data image prediction.The urban feature classification method based on case segmentation network proposed in this study can accurately extract the main feature types of most cities in China,and generate vector results with complete semantic information,which has broad practical application prospects. |
PDF Full Text Request