Road Network Updating Based On Multi-conditional Adversarial Network

Road network,as an important part of GIS,is widely used in people's lives.Therefore,how to quickly and effectively obtain complete road network information has become a hot research issue.At the same time,the development of high resolution remote sensing image technology also provides a new channel for road research.In that case,research on road network updating is carried out in this paper,the main research work is organized as follows.Firstly,a road extraction algorithm based on Aperiodic Directional Structure Measurement(ADSM)is implemented.This algorithm extracts the potential probability model of Road area.Subsequently,under the guidance of the potential probability model,Line Segment Detection(LSD)method is used to extract road primitives from remote sensing images,and the primitive segments are fused and connected to form road network mask.After that,combined with some morphological filtering operations,the final binary road map is obtained.We validate the effectiveness of the algorithm on Shaoshan data set,and obtained a detection accuracy of 0.714.Secondly,aiming at the disruption of road network topology caused by shadows and occlusions and combining with the existing historical road network vector data of business departments,we first propose a road network matching algorithm based on iteratively extended segments.The candidate matching roads for each reference target are determined by a simple Buffer Growing(BG)method.Then,the final matching pairs are selected by extending both the reference and candidate targets iteratively.This method replaces line segment matching with topological network matching.And context connection information of road is taken into account in matching.Finally,87.24%matching accuracy is achieved.Finally,in order to obtain a complete network structure and update the network,an end-to-end Multi-conditional Generation Adversarial Network(McGAN)is proposed.This network consists of two discriminators,a generator and a pre-trained VGG network.One discriminator is used to capture road spectral information for assisting road network generation,and the other one is used to guide road topology reconstruction.The pre-trained VGG network is designed to obtain the fine features of roads.The generator is trained jointly by these two discriminators and a pre-trained VGG network.The input of the network includes three types of samples:original remote sensing image,initial road map and the ground truth.The network considers both the spectral and topological characteristics of the road,so the final road network generated by McGAN is more complete than the ordinary method.Experiments on Massachusetts data set and Shaoshan data set were carried out,and the accuracy of 0.841 and 0.862 were obtained respectively,which proved the validity of the method.