Road Extraction From Remote Sensed Imagery Based On Deep Learning Considering Geometric And Topological Properties

Automatic extraction of important geographic information from remote sensing images is of great significance for social development.Road information is the key element of geographic information,which has a wide range of applications.Automatic road extraction can greatly reduce the cost of labor and time,which has been researched for a long time in photogrammetry and remote sensing.The high resolution satellite images provide high-quality data source for road information extraction.It is a significant and challenging task to automatically extract road networks from high-resolution remote sensing images.By analyzing the geometric and topological properties of road network,we proposed two deep learning based road extraction methods.(1)We proposed a novel regression-based method to extract centerlines and edge lines via predicting road confidence map and estimating road width.First,a regression model based on convolutional neural network(CNN)is designed to learn road confidence maps and width maps based on input images.Then,after the CNN predicts the confidence map,non-maximum suppression and road tracking are followed to attain accurate road centerlines.Road edge lines are generated based on the extracted road centerlines and road width estimated by the CNN.Finally,tensor voting is applied to detect road intersections and improve the connectivity of road network,wherethe intersections are used as guidance for the overcome of discontinuities to attain higher quality road networks.(2)We proposed an approach for road topology extraction to improve the connective performance,keeping the integrity and reducing the complex post-processing..By locating road nodes and estimating the connectivity between nodes on images,the road network is represented by an undirected graph.The graph is denoted as G=(V,E),where V and E represent the road nodes and edges between nodes.A convolutional neural network is jointly trained to predict the nodes and connectivity map for nodes.After the network predicts the nodes,the connectivity between different nodes is measured based on the predicted connectivity map.Then the edges between nodes are inferred from the measured connectivity and road topology is reconstructed.Finally,,we apply road topology refinement to the extractedroad network to attain high quality road topology.Experiments show that the proposed methods in this thesis can extract high quality road topology in different scenes and achieve good performance in terms of topological precision and recall.