Research On Key Technologies Of Interlligent Road Network Extraction From Remote Sensing Image

As a basic geographic information data,road network has huge importance in many fields.In recently years,a flourishing in both quantity and quality of high resolution remote sensing satellite images of China,provides a rich data source for the study of road network extraction.However,there still exit some problems,such as the phenomenon of "homologous dissimilarity",the unapplicability for different geomorphological conditions or different data source image.Though,many road extraction methods x,like artificial interactive interpretation,edge detection,template matching,etc,have been proposed,but some shorcomings,such as excessive manual interventions,low automation and extraction rules relying on the operator’s professional knowledges and abilities of feature analysis,are not yet setettled.In order to resolve these problems,the study firstly analyzes the advantages and disadvantages of C4.5 algorithm and DBN for road extraction by theory and experiments.Next,propose the combined road extraction method based on C4.5+DBN,and the performance of the new method is verified from different geomorphological environments and different data source images.Finally,study the postprocessing and network vectorization method.The main conclusions can be concluded as:(1)Proved by theoretical analysis and experiment,the complexity and classification accuracy of C4.5 algorithm and DBN are closely related to the number of categories.For the target of road network etraction,the classification categories are defined as road and non-road,which can effectively reduce the complexity of C4.5 and DBN.In the training process,landforms which are esay confused with road need to be fouced on.(2)When use the C4.5+DBN method,the road extraction rules can be automatically established by C4.5 to construct the road candidate area.DBN is used as the intelligent classifier to extract the road information in the candidate area.Manual interventions are mainly on sample selecting in the whole process,which menas the degree of automation is greatly improved.Combining DBN’s strong road feature expression abilities has made up for the shortcomings of linear and single-time feature utilization in C4.5 algorithm.Using the advantages of small training samples set of C4.5 has made up for the disadvantage of DBN’s high requirement for training samples and network depth.With the same precision conditions,the DBN depth can be reduced from 10 to 4,while the sample demand is reduced by 93% and the training time reduced by 82%(3)In terms of road extraction accuracy: the C4.5+DBN method has a 62.8% and 34% lower road false detection rate than the single C4.5 algorithm and DBN with the same depth(4-layers)respectively;The great accuracy in the experiments of different geomorphological conditions and different data source images demonstrats the effectiveness and robustness of the new method.(4)The core algorithms of road network vectorization such as road refinement,node extraction,vector curve fitting,etc.,have been studied and now can be automatically implemented;Road informations are stored in the form of shapefile polyline,which reduces the storage and facilitates subsequent researches.The radius of the curve can be automatically calculated with the error within 5%.And the curve whose radius is less than 12 m can be detected.