Correction Method Of Urban Road Vector Data Under Deep Learning Constraint

High quality road vector data plays an important role in smart city planning,intelligent transportation,public health and other fields.However,due to the lag of the time of vector data acquisition and the inconsistency of standards,there are some geometric deviations between the road vector data and the actual road scene.With the development of Remote Sensing Science and technology,the ground feature information depicted in high-resolution satellite images is more detailed,which provides reliable and sufficient judgment basis for the correction of road vector data.Therefore,this paper proposes a method to correct urban road vector data under the constraint of deep learning from the perspective of matching the forward and vector images.The method can correct the urban road vector data through three stages: the U-Net road extraction,the line segment sequence detection and the vector data correction.The details include:(1)Road extraction method based on convolution neural network U-Net.In view of the current situation that the urban road noise is large and the interference situation is complex.This paper uses a U-Net network structure including down-sampling and up-sampling to extract roads.Firstly,four down samples are used to obtain the context information of the image and extract the target features;Secondly,four up-sampling processes are used,and each layer of up-sampling information is combined to perform precise positioning and restore detailed information;Finally,the convolutional layer is used to classify the pixels in the image one by one,and the road area is extracted.The experimental results show that the average accuracy of road extraction in complex urban scenes by this method can reach more than 60%,and the average recall can reach more than80%.(2)Aiming at the characteristics of long and narrow geometric information of roads,a line segment sequence detection method is proposed.Aiming at the problem that the extraction results of line segments in high-resolution remote sensing images are independent of each other,and it is difficult to adapt to changes in the curvature of the edge of the feature,the method adopts the orderly arrangement of line segments at different angles to achieve the complete expression of the edge of the feature.Based on the connectivity analysis of edge points,the method includes four parts: edge grouping,line segment extraction,line segment optimization and line segment sequence generation.Through the experimental quantitative and qualitative analysis,it shows that the method has the ability to express the edge of irregular objects accurately and completely,and achieves the purpose of expressing the curve by line segment,which is used as the geometric structure information of road edge,which is helpful to the correction of vector data.(3)Vector data correction.According to the fact that the road in the image has a certain width,a curve fitting method of random sampling least square method is proposed.Firstly,in view of the broken line mutation of urban road vector line,it is difficult to describe the vector line with given function.This paper decomposes the vector data by analyzing the geometric characteristics of vector intersection and inflection point,that is,a vector line is decomposed into multiple vector lines;Secondly,based on the u-net road extraction results and the existing vector data,through the matching between the U-Net road extraction results and the line segment sequence,and the geometric information constraints of the existing vector data on the initial line segment sequence set,the line segment sequence in the road region is screened,and the candidate line segment sequence set is extracted and grouped;Finally,in order to meet the needs of curvature change of sub vector line,the dynamic curve model is used,and the vector line is connected with the random sampling least square fitting method to correct the vector data.In order to verify the reliability of the proposed method,this paper uses high-resolution optical remote sensing images of different resolutions and different regions as experimental data to compare with other existing vector data correction methods.Experimental results show that the correction accuracy and recall rate of this method can reach more than 90%,which is much better than other methods.