Research On Automatic Road Extraction From High Resolution Remote Sensing Image Based On Deep Learning

Road information is an indispensable part in the development of modern society.It is of great scientific significance to study the methods of road extraction from remote sensing images.Road extraction involves multiple tasks such as vehicle navigation,urban planning,intelligent transportation,image registration,geographic information system update,and so on.At present,the accuracy of road extraction from low-resolution remote sensing images is far from meeting the requirements of data updating.Therefore,how to use the image features and road features of high-resolution remote sensing images to automatically extract roads has become one of the focuses of current road extraction,and it is an urgent problem in related applications.Earlier methods of road extraction are mostly based on low-layer features and prior assumptions.The accuracy of the extracted roads is not very high,and the noise of extracted road is more.Moreover,these methods are relatively cumbersome and inefficient,especially not suitable for complicated scenes.In recent years,deep learning is increasingly used in remote sensing research.Although the noise and sensor error have been improved,road extraction still exists the inconsistent problems in edges of extracted road due to interference which includes the shadows of trees and vehicles.In addition,the large difference between the number of road pixels and non-road pixels on high-resolution remote sensing images leads to the poor performance of road extraction model.In view of the above problems,the methods and innovations proposed in this paper are as follows:1)An automatic road extraction method based on refining road topology information is proposed and a road segmentation framework with dilated module and information module is built.The dilated module can increase the receptive field and extract multi-scale features without losing the feature resolution.And the information module can effectively obtain the spatial relationship between the pixels in the rows and columns of the image.In addition,training the model using the combined lossfunction constraint.The results show that the accuracy of the proposed method on the Deep Globe Road dataset is 2% ～ 6% higher than that of other segmentation methods,and the accuracy on the Massachusetts Road dataset is 2% ～ 14% higher than other segmentation methods.2)An automatic road extraction method based on context feature and stripe feature enhancement is proposed,and an end-to-end road segmentation framework with attention mechanism and refinement network is built.The attention mechanism is used to select and emphasize useful contextual information at different scales of the feature layer.And the refinement network with parallel operations of stripe pooling and space pooling is used to enhance strip shape characteristics of the road.In addition,the frequency of road pixels is used as the weight in the loss function to constrain model training.The method proposed made up 0.23% ～ 5.31% than other methods on the accuracy in Deep Globe Road dataset,and made up 1.24% ～ 4.75%than comparative methods on the accuracy in the Massachusetts Road.In this paper,the automatic segmentation and extraction of complex road with high-resolution remote sensing image is studied,and two methods of road topology information refinement and feature enhancement are proposed.The experimental results and analysis on two very challenging datasets of Deep Globe Road and Massachusetts Road proved that the effectiveness of proposed methods.