Network Topology Constrained Road Extraction From High-Resolution Remote Sensing Images With Geometrical And Spectral Features

With the improvement of remote sensing technology,spatial resolution of remote sensing image can reach sub meter level,which provides clearer local geometry features and global network topology features for road extraction from remote sensing images.However,the accompanying problems are that vehicles,canopy shadows,traffic signs and other ground objects in the road area are becoming clearer and occupy a considerable area,and the non road areas on both sides of the road area have become two mixtures of buildings,vegetation,vehicles and other ground objects,which greatly weakens the homogeneity of spectral measurements of pixels covered by road area and non road area,as well as the heterogeneity between spectral measurements of pixels covered by both areas.The problems result in omission and disconnection in road extraction results.To improve the quality of road extraction from high-spatial-resolution(abbreviated as high-resolution)remote sensing images,this paper proposes a network topology constrained road extraction algorithm with geometrical and spectral features.The main tasks are summarized as follows.(1)Network topology model and geometry model of road are constructed.Based on the stochastic geometry theory,stochastic segment process is defined to model the center lines of road sections.According to the network topology feature of road,interaction relationship constraining model between segments and state constraining model of segment itself are constructed.A network topology model of road is constructed by combining the segment process model and the constraining models above,which aims to improve the connectivity,etc.of road extraction;Besides,according to the local geometrical feature of road,a geometry model characterizing road area is constructed by using rectangle marked segment process and a geometry model characterizing non road area is constructed by buffering the rectangle marks with certain widths.The geometry models provide spatial constraint for spectral measurement modelling of road.(2)Spectral measurement model of road is constructed.Based on the geometry models characterizing road area and non road area,and according to the mixed homogeneity features of spectral measurements of pixels covered by the road area and non road area,respectively,the Gaussian mixture model is used to construct homogeneity models of spectral measurements of pixels covered by two areas above.Moreover,according to the mixed heterogeneity feature between spectral measurements of pixels covered by two areas above,Kullback-Leibler(K-L)divergence is used to construct a heterogeneity model between two areas.In this way,a spectral measurement model of road is constructed by combining the homogeneity model and heterogeneity model,which can improve the accuracy and completeness of road extraction.(3)Simulation optimization algorithm of road extraction model is designed.According to the Bayes theorem,a posterior probability model characterizing road extraction is constructed by combining the above models.According to characteristics of the road extraction model designing,Reversible Jump Markov Chain Monte Carlo(RJMCMC)simulation algorithm is designed,of which the jump transfer operations are updating of parameters,simple transformations of segments,fusion or separation of endpoints,and addition or deletion of segments.The simulation criterion,maximizing a posteriori probability,is used to obtain the optimal road extraction result.(4)Verification and evaluation of the proposed road extraction algorithm.Comparative experiment of network topology structure simulations,local road segment extraction experiment,road network extraction experiment and corresponding comparative experiments are designed.Experimental results are qualitatively evaluated by using visual interpretation method and are quantitatively evaluated by using buffer evaluation method,quality evaluation method and evaluation method of matching degree between connection nodes.Evaluation results show that the proposed algorithm can extract the roads from high-resolution remote sensing images more effectively and accurately.There are 57 figures,11 tables and 147 references in this paper.