Road Extraction From Remote Sensing Image Based On Deep Learning

Road information is one of the basic elements in remote sensing images.Its spatial orientation and trend direction are of great significance for the updating of road network information and the navigation of satellite positioning.With the popularity of high-resolution remote sensing images in recent years,extraction of road information has gradually become one of the key research directions for image interpretation.Limited by the complexity of the surface environment and the limitations of the algorithm,a perfect road extraction method has not yet emerged.In this paper,based on previous studies,the road extraction in high-resolution remote sensing imagery is studied.Based on the deep learning algorithm,two kinds of road extraction methods are proposed to improve the accuracy of road extraction.We provides a new direction for road extraction work.The main work of this paper is as follows:(1)In this article,we classify the road categories,analyze the characteristics of roads,and summarize the characteristics,advantages,and disadvantages of various features.At the same time,we introduce the relevant theories of deep learning algorithms and the deficiencies of existing road extraction methods.We also analyze the feasibility of road extraction with deep learning.(2)We also use the migration learning method to introduce the full convolutional neural network of deep learning to the road extraction of remote sensing images,and extract the information from the network intermediate layer combined with refinement algorithm to extract roads.The experimental results in this paper show that the first-layer convolution results of the network have good road surface denoising characteristics.Therefore,combining this feature with the Zhang-Suen refinement algorithm to get road information which is more accurate than before.(3)We improved the structure of the original neural network,making the network more suitable for extracting road information from remote sensing images,reducing the risk of overfitting,improving training efficiency,and reducing hardware requirements.At the same time,a road shape feature index is improved,the discrimination degree of the network interference information is improved,and the threshold adaptation is implemented to improve the automation degree of the extraction method.The centerline is extracted by a multivariate adaptive spline regression algorithm with the help of road edge information.Compared with using the entire road surface information to return to the centerline,this method can not only avoid the interference caused by road surface voids,but also accelerate the calculation speed.