Research On Road Extraction Method In Remote Sensing Images Based On Deep Learning

The road extraction task of remote sensing image is of great value and significance in the applications of geographic information services such as automatic driving,intelligent transportation,rescue and disaster relief and urban digital governance.With the rapid progress of Chinese aerospace science and technology,high-resolution remote sensing image data provides rich data sources and information support for remote sensing image road extraction tasks.Meanwhile,the development of semantic segmentation based on deep learning also provides the algorithm foundation and technical support for remote sensing image road extraction.However,in remote sensing images,roads have the characteristics of complex and diverse backgrounds,some images are blocked by surrounding trees,interfered by buildings and trees’ shadows,and the texture features of roads and other similar ground objects are easily confused,which causes the optical and physical information of some road areas in the road extraction to be weakened,and causes obvious interference to the road target recognition and classification in the local area.Based on the residual network structure,encoder-decoder network structure,attention mechanism,differential convolution network and other methods in deep learning semantic segmentation,this thesis takes road pixel-level feature recognition enhancement and road edge connectivity feature recognition as the main path,and carries out the following main research work:(1)A road extraction method based on improved residual network is proposed in this dissertation.Aiming at the problem that the road background of high-resolution remote sensing images is complex,the road features are not obvious,the data is unbalanced and the road edge is still discontinuous in the semantic segmentation results of road pixels.Based on the deep learning method and the Res Net101 network,through the atrous spatial pyramid pooling operation,the same image is convoluted with multiple dilated convolution kernels of different sizes,multiple results are obtained and then fused into output,so as to extract the relationship characteristics between the features and improve the segmentation accuracy of the network model.By replacing the bilinear interpolation method with the upsampling method based on sub-pixel convolution,the mixing of noise signals in the upsampling process is effectively suppressed without increasing the computational cost.The new training image is expanded by data enhancement method,which effectively suppresses the occurrence of over-fitting problem.Finally the road skeleton is extracted from the semantic segmentation results,then the road edge is automatically extracted based on the road center line to realize the repair of locally discontinuous roads.Through experimental tests,it is verified that the proposed method improves the accuracy and completeness of road extraction.(2)A road extraction method based on multi-feature fusion and attention mechanism is proposed in this dissertation.Aiming at the problems of feature recognition at different scales in road information of high-resolution remote sensing images,the road target data has positive and negative samples,the difficult and easy samples are unbalanced.Taking the encoder-decoder structure of U-Net network as the basic network,the features of different spatial scales are extracted by the atrous spatial pyramid pooling method;Multi-channel feature extraction is realized through residual-based compression excitation network.The features extracted from the two dimensions of space and spectrum are fused to achieve multi-feature fusion.Through the feature recognition method based on attention mechanism,the attention gate network module is designed in the decoding stage and the accuracy index is improved without significantly increasing the calculation and time cost.Using the loss function based on the gradient coordination mechanism the contribution between the samples is balanced.the model training is more efficient and stable.Through experimental tests,it is verified that this method can achieve effective road extraction by experimental tests.(3)A deep learning road extraction method based on road connectivity topology enhancement is proposed in this dissertation.Aiming at the problem that the road has trees and building shadows,the feature information of the road edge is easy to be ignored,resulting in poor road topology connectivity,on the basis of U-shaped image segmentation network model,the multi-enhancement encoder is formed by using the atrous spatial pyramid pooling and compression excitation network in the encoder part,the attention gate module is connected in parallel in the decoder part to form the attention enhancement decoder,the encoding and decoding part of the U-Net network is optimized to improve the semantic segmentation performance of the road.Based on pixel segmentation,a method based on pixel differential convolution is studied to construct a road connectivity topology information extraction module,and the residual network,channel attention and atrous spatial pyramid pooling are used to realize road topology feature extraction.Finally,a hybrid loss function is designed to fuse the road feature recognition results based on semantic segmentation and the road connectivity topology feature recognition results to realize the edge refinement of the road extraction results.Through experimental tests,it is verified that this method further improves the accuracy,integrity and connectivity of road extraction,and also proves that the topological knowledge of road is used to supplement and improve the accuracy of road extraction.