The Methods Research Of Remote Sensing Image Road Extraction Based On Deep Learning Semantic Segmentation

Road extraction based on remote sensing images plays an important role in basic map data processing and services,e.g.automatic driving,emergency command and smart city construction.The development of remote sensing technology and deep learning semantic segmentation method provides data assurance and technical support for the road extraction based on remote sensing images.However,road extraction from remote sensing images that based on deep learning semantic segmentation is still faced with challenges.The road object in remote sensing image is characterized by long and narrow span,complex background,geometric texture features easily confused with the background,unbalanced sample,being easily obscured,difficulty in topological connection,etc.,which make the road extraction based on remote sensing images a challenging job in the field of deep learning semantic segmentation.How to improve the modeling capability of remote dependency,especially the dynamic modeling of adaptive samples is important for the extraction of road object with unique features,which remains a major problem for deep learning semantic segmentation.At present,the problem is mainly solved by means of dilated convolution,pyramid pooling and self-attention mechanism.,but their application effect in road extraction based on remote sensing images is affected as being restricted by the spatial information loss,the difficulty in capturing anisotropic information flexibly or information redundancy,etc.This dissertation aims to improve the road extraction accuracy via enhancing the modeling capability of remote dependency and perfecting road connectivity on the basis of attention mechanism,graph convolutional network,lightweight and efficient encoder-decoder network and other technologies in deep learning semantic segmentation.In view of the above–mentioned problems,we have studied:(1)the semantic segmentation model based on global second-order information aggregation and the road extraction method based on this model;(2)the semantic segmentation model for adaptive capture reasoning of anisotropic information that based on dual dynamic graph convolutional network(D2GCNet)and the road extraction method based on this model;(3)the iteration reinforcement(Ite R)post-processing model for improving road connectivity by location information fusion and the road extraction method based on this model;(4)the data preprocessing methods for remote sensing image and GNSS in road extraction;(5)the road extraction integration method based on a variety of semantic segmentation models.Finally,the following innovative approaches and models are proposed after experiment,testing,analyzing,evaluation and application of various research methods on the basis of public data sets such as Space Net,Deep Globe,and Bei Jing Data Set as well as practical application data.(1)This dissertation proposes a semantic segmentation road extraction method based on global second-order information aggregation.In view of spatial information loss and global long-distance information underutilization caused by dilated convolution,this dissertation proposes a semantic segmentation road extraction method based on global second-order information aggregation.The semantic segmentation method based on global second-order information aggregation first calculate the attention weight coefficient,and aggregate the second-order information of different feature maps via the attention mechanism based on bilinear pooling to generate low-dimensional global feature resources associated with space and feature channel dimensions;next,the feature resources are distributed dynamically according to the attention weight coefficient and adaptively select complementary features according to the needs of each pixel,so that the dependency on global long-distance information can be obtained for the local position;then,the network superimposes its own features through residual operation to make up for the ignored features and re-projects to restore the original input size.Finally,a new road extraction method is formed by embedding the semantic segmentation network based on global second-order information aggregation into the advanced semantic information layer in a lightweight and efficient encoder-decoder network.Experiments show that the result is efficient for road extraction while ensuring the reliability of the model.The proposed method is proved to be effective in information de-redundancy through t-SNE-based cluster and feature matrix visualization analysis.(2)This dissertation proposes a dual dynamic graph convolutional network and road extraction method based on D2 GCNet.In view of the dynamic adaptive modeling of remote dependency,this dissertation proposes a semantic segmentation named dual dynamic graph convolutional network(D2GCNet),designs the network architecture of the method and defines its model abstractly.The proposed model first projects the pixel clusters in the feature space to the nodes in the graph space,and preliminarily filter anisotropic information by setting the graph node parameters;next,the model dynamically constructs the adjacency matrix of graphs by using the KNN to filter global anisotropic information thoroughly again,makes use of graph convolution to spread long-distance information,and designs a dual-branch structure to achieve dynamic reasoning aggregation of information on two dimensions: Node and State;then,the model superposes own features information through residual operation,and re-projects the nodes in the graph space to the pixels in the feature space.Finally,a new road extraction method is formed by embedding the D2 GCNet into the advanced semantic information layer in an encoder-decoder network.Experiments show that the proposed model is small and low complexity,which can effectively improve the accuracy of road extraction and remove redundant information.(3)This dissertation proposes an iteration reinforcement(Ite R)post-processing model based on location information fusion and the road extraction method inserted the model.In view of the difficulty in topological connectivity,this dissertation proposes an iteration reinforcement(Ite R)post-processing method based on location information fusion,designs the network architecture of this method and defines its model abstractly.The proposed model first splice the prediction output map,remote sensing image after histogram equalization,road label and GNSS position data after deviation correction,through which the multi-dimensional,multi-level and multi-resolution contextual information are integrated;then,the model takes the spliced data as the new input data,uses the self-defined reinforcement loss function based on BCE + Dice-coefficient loss for supervision model training,and iterates training until convergence;finally,a new road extraction method based on location information fusion is formed by inserting the Ite R post-processing model into the end of lightweight and efficient encoder-decoder network.Further experiments show that the performance of Ite R post-processing model will increase to the optimal value and then remain stable with the growth of the number of basic blocks of Ite R post-processing model,and the performance is related to the data preprocessing operation of histogram equalization.Ite R post-processing model is effective in supplementing the spatial information details and against the obscure boundary of output results due to information loss.Experiments show that the proposed method can effectively improve the situation of disconnected road extraction caused by shielding from buildings,trees,clouds,shadows,etc.,and then improve the accuracy of road extraction as a whole.(4)This dissertation presents a road extraction integration method based on multiple semantic segmentation models.In order to better support the transformation and application of the research results in this dissertation,a road extraction integration method based on multiple semantic segmentation models is proposed.The integrated methods are selected and combined according to the actual data and the constraints of each model,and the weight coefficients are adjusted according to the actual contribution of each model and the respective prediction output results are automatically superimposed.The application shows that the three methods proposed have strong generalization ability,and the integration method is effective.