Road Bone Extraction And Automaticcompletion Based On Remote Sensing Satelliteimage Extraction Results

Road extraction from high-resolution remote sensing images is a key issue in the field of remote sensing image recognition.The accurate extraction of roads in remote sensing satellite images can provide practical references for road planning and regional architectural design,military surveying and mapping,disaster assessment and other fields.At present,the use of image processing,machine learning and deep learning methods to automatically extract roads in remote sensing images is currently a hot research field.Since2018,the use of deep convolutional neural networks for road extraction from remote sensing satellite images has been greatly ahead of other methods.However,even the most advanced neural networks cannot eliminate the problems of identifying road interruptions and rough road edges.These problems largely affect the direct use of road extraction results for spatial decision-making and analysis,which needs human resources to repair and complement roads.Nowadays,with the advancement of basic semantic segmentation technology,it can provide more refined edges for target recognition,which in turn clears obstacles for the previous morphological road bone extraction.In this paper,the progress in the field of basic semantic segmentation is tried on the road recognition problem,and a set of automatic road extraction and road bone extraction algorithms are designed.At the same time,the road automatic repair and completion algorithm for road interruption and other problems is designed for an end-to-end automated road extraction and road bone extraction process program.The work of this article mainly focuses on three aspects.First,it demonstrates the effectiveness of new technologies,such as expanded convolution,deep separable convolution in semantic recognition in road extraction from remote sensing satellite imagery,analyzes of the performance of Deep Lab V3+ on relatively simple problems and improvements of architecture for better processing Road extraction problem.Secondly,for Cas Net,the neural network framework that uses cascaded networks to extract roads and road centerlines in turn is improved,and the effects of possible dual network connecting methods,such as cascade parallel and shared parameters,are explored.Cascade is finally selected as the final network.The structural design further improves the road recognition quality and the road centerline recognition effect.Finally,for the interruption and rough edges of the road,morphological methods are used to repair.The final repair results and recognition results are used to generate a vector road network.Although a certain quality reduction price is paid,it provides more connected road networks,which is convenient and smooth.The road extraction data is used for further road planning and other decision-making.In order to verify the effectiveness of the algorithm,we compare the algorithm in this paper with other popular algorithms on the 2018 CVPR Deepglobe Road Extraction competition data set.It shows that this method greatly improves the accuracy of road extraction.The m IOU index of road extraction is also the road.The extraction quality reached more than 85%,shows better performance on the edge m IOU index.And the road bones extracted by the neural network produce fewer burrs and smoother road centerline curves than other methods.After further post-processing,the generated vector road network can bring effective guidance for the subsequent surveying and mapping applications,which greatly saves manpower and material resources.