Road Extraction From High Resolution SAR Images Based On Deep Learning

As the main part of modern transportation network,road not only has important geographical,political,economic and military significance,but also can provide very rich and valuable information for the extraction of other objects.From the application point of view,roads can provide rich information in the monitoring of urban development.When extracting other objects such as vehicles,pedestrians and buildings that rely on roads,the extraction of the road is even more indispensable.From a technical point of view,it is still difficult to extract roads from high-resolution SAR images.First,there are many interferences and occlusions in the SAR image,which cause the structural damage of the road,and the scattering of the road is not strong,so that the distinction between the road and the surrounding environment is not high.Second,roads in SAR images account for a small proportion and are slender,which often cannot be extracted completely and accurately.Third,the shape of the road is changeable,there are small winding roads and there are two-way lanes isolated by the isolation belt,so general extraction can’t get a good result.These problems are the key to road extraction from SAR images.Therefore,for a series of problems at the application level and technical level of high-resolution SAR images,the main work of this thesis is as follows:（1）In view of the lack of relevant SAR image road datasets,we select appropriate image data from many SAR images containing roads to establish a SAR image road dataset in this thesis.Due to the changeable road shapes,it is too difficult to extract them in a general way,therefore,the roads are classified according to their attributes:urban roads,rural roads,and then labeled with the “Labelme” labeling tool.Finally,an annotated dataset for road extraction from SAR images was produced: 83 urban roads and 218 rural roads.（2）Aiming at the problems of many interferences and occlusions and poor road scattering in SAR images,we proposes an improved MAU-net network based on a mixed attention mechanism in this thesis.The network pays more attention to road targets in SAR images,and can extract road targets more effectively when the contrast between the road and the surrounding environment is low.Experiments show that compared with the existing U-net network and Res U-net network,this network has higher integrity and reliability for road extraction from high-resolution SAR images.（3）Aiming at two problems in the existing research: first,the overall proportion of roads in SAR images is small and has slender features.Second,traditional road extraction methods generally use downsampling and pooling operations to extract feature information,which will lead to the loss of target information and information reconstruction cannot be performed.We proposes a concatenated atrous spatial pyramid pooling module in this thesis,which enables the network to expand the receptive field during up/down sampling,and can more completely acquire multi-scale feature information at different levels of the road without causing information loss.