The Extraction Of Urban Public Space Information From GF-1 Image Based On Deep Learning

In the past 10 years,UN-Habitat has promoted the use of public space as a strategy for the implementation and execution of projects such as urban planning,housing and slum reconstruction,governance and urban security,basic services and even post-conflict reconstruction.Urban public space is an important part of the direction of urban development towards sustainability,tolerance and safety.Whether it is the reconstruction of the old urban areas of mega-cities or the construction planning of small and medium-sized cities,they are faced with the problem of lack of sufficient information support.The traditional method is difficult to count urban public space information and consumes a lot of manpower and material resources.With the continuous maturity of remote sensing technology,higher resolution remote sensing images provide a reliable data source for surface feature information.Due to the complex types of urban features and the influence of factors such as illumination and satellite shooting angles,information extraction from urban remote sensing images has always been a research difficulty in this field.Faced with the massive high-resolution remote sensing image information extraction work,researchers are constrained by heavy workload,low extraction accuracy,and high difficulty in practical applications.In recent years,the rapid development of deep learning technology has provided new development ideas for related research in the field of computer vision.This paper takes urban public space information extraction as the main research task,and whole country provincial capital cities GF-1 satellite images as the main data source.First,based on remote sensing images,urban spatial features are divided into buildings,roads,playgrounds,water bodies,etc.According to the above classification,use ENVI5.3 software to manually mark the area of interest.Due to the huge workload of manual labeling,the open source map OSM and other open source vectors are used to assist in completing the labeling of the samples.Secondly,use the SegNet network and DeepLabv3+network based on deep learning for semantic segmentation network training.The accuracy is evaluated according to the accuracy evaluation index of the semantic segmentation network-mIoU(Mean Intersection over Union).The results show that the DeepLabv3+network based on the data set in this paper has better segmentation effect than the SegNet network.In view of the low segmentation accuracy of the network and insufficient edge features of small targets and some dense buildings,the original samples of the data set are sharpened.The backbone network of DeepLabv3+is replaced with ResNet101,and the model of Xception71 feature extraction network is compared and analyzed,and the channel attention mechanism is introduced to further optimize the model.In the end,the average intersection ratio of the DeepLabv3+model with ResNet101 as the backbone network increased by about 2.7%.Use the trained DeepLabv3+network to segment and extract information from the features in the built-up area of Changchun City.According to statistics,the proportions of roads,playgrounds,squares,vegetation,and building gaps in public space in Changchun are 6.29%,1.57%,1.80%,20.55%,and 18.25%,respectively,and public space accounts for 48.46%.