Research On Automatic Road Extraction Method From Domestic Gaofen-2 Satellite Images Based On Deep Learnin

As the infrastructure and bearing body of transportation industry,road plays a very important role.The development of remote sensing technology makes it possible to obtain the distribution of land road network by using satellite remote sensing images.However,because the road is a slender target and is similar to the spectral reflectance of bare land,buildings and other ground objects,it has always been a research difficulty to extract a continuous and clear road network from remote sensing images.Taking the domestic GF-2 satellite image as the data source,this thesis discusses three high-precision automatic extraction methods of road remote sensing by using the technologies of data mining and deep learning.The main research contents are as follows:(1)Using support vector machine technology,an automatic road extraction method of domestic GF-2 satellite image based on spectral features and template matching post-processing algorithm is proposed.Firstly,the road is preliminarily extracted by combining spectral features and texture features.By designing a template matching algorithm,the preliminary road extraction results are post processed to remove the non road part,and then the automatic road extraction is realized.The results show that when the complexity of the surrounding environment is not high,this method can effectively remove the noise in the preliminary road extraction results,and is suitable for identifying wider trunk roads.(2)Aiming at the problem of complex ground objects around roads in remote sensing images,an automatic road extraction method based on SM-Unet is proposed in this thesis.In order to capture the long strip feature information of isolated road area,the stripe pooling module is added before the down sampling of network encoder;In order to enhance the network’s ability to obtain the context information of the road area in the complex scene and make the road feature representation more discriminative,the encoder finally rolls up the layer and adds the hybrid pooling module.The results show that compared with the obvious leakage and fracture in the road extraction results of the existing network,the network proposed in this thesis can extract more complete and continuous roads,which is suitable for the road extraction task of domestic GF-2 satellite image with complex ground object spectrum.(3)Considering that multiple down sampling of Res Net-101 network will cause the loss of edge detail information when extracting roads,based on Res Net network,this thesis reduces the down sampling times,introduces void convolution in the residual block,introduces enhanced void space pyramid module and channel attention module in the network,and designs a Supr-Res Net network structure.Then an automatic road extraction method based on Supr-Res Net is proposed.The results show that the road accuracy extracted by this method is better than the existing networks such as U-Net and Res Net-101.Through the three road extraction methods studied and improved in this thesis,we can extract road targets with high precision in remote sensing images,and adapt to road target extraction in different complex range target scenes,different road shapes and different ranges.It can be effectively applied to traffic transportation,road planning,map navigation and military reconnaissance,and has great application value.