Road Extraction From High Resolution Remote Sensing Images Based On Road Intersection

Road is the most basic and importrant geographic information data which makes road extraction based on remote sensing images are of great significance and research value.As a node of the road network,road intersections have stable and reliable characteristics and can provide information such as road connection direction and topological relationship.The obtaining of road intersections plays an important role in road extraction.High-resolution remote sensing images are currently the most widely used image data.This paper mainly studies the extraction of road intersections and road from high-resolution remote sensing images.The main work of this article is as follows:(1)Taking road intersections as independent object targets,road intersection detection based on deep learning target detection algorithm is studied.A sample set was constructed for different types of road intersections,and a deep learning object detection model,Faster R-CNN,was trained to automatically identify and locate road intersections from high-resolution remote sensing images.The experimental results show that the road intersection detection based on deep learning can detect the planar intersections formed by the trunk roads in the images from the remote sensing images,and can not only locate the geometrical positions of the road intersection targets,but also can accurately identify the roads.The type of intersection geometry is better than traditional methods.The algorithm is suitable for high resolution remote sensing images with spatial resolution higher than 2m and road width greater than 5 pixels.(2)An image segmentation algorithm based on graph cut is used to extract homogeneity area of road intersections.For the characteristics of road intersections,the GLCM texture feature is introduced,along with the shape index feature and spectral features of the pixel to form the feature space.Then the energy function of the graph cut is constructed.The maximum flow algorithm is used to minimize the energy function,and the overall energy is minimized through iteration to extract road intersection homogenous area.Experiments show that the introduction of multi-features can be used to extract road intersection regions from road intersection image slices.Compared with the marker watershed algorithm and ordinary map algorithm,the results are better.(3)The road extraction combining road intersection extraction results with semi-automatic road extraction method was studied.Firstly,the tensor voting algorithm is applied to obtain the structural information such as the coordinates of the center point of the road intersection and the road branch width and direction from the road intersection binary map.The road intersection structure information is a priori information,and the parameters of the semi-automatic road extraction algorithm are initialized to extract the road,and the human-machine interaction process in semi-automatic road extraction is reduced to some extent.Through the experimental analysis,the research done in this paper can better extract road intersection information from remote sensing images.Combining extracted road intersection information with semi-automatic road extraction algorithms to extract roads from high-resolution remote sensing images can reduce fully or partially the human-machine interaction of semi-automatic road extraction.This study has certain practicality.