Road Extraction Method Using Multi-source Remote Sensing Data

Automatic (semi-automatic) road extraction from remote sensing is challengingstudy and with great attention. Currently, the existed studies involved theories andtechniques in fields of remote sensing, computer vision, and pattern recognition. Lots ofapproaches have been proposed.Feature description and method of extraction are two core issues for roadextraction. However, there are some shotcomes for the existed approaches in these twoaspects. In the first aspect, the shotcomes include: weak ability for road featuredescription, and insufficient using of comprehensive advantages of multi-source remotesensing data. While in the second aspect, the shotcomes include: strong dependencybetween method and data, poor flexibility for method extension. In order to solve theabove problems, we take road feature description and extraction methods as the primarystudy contents this paper. The main studies of this paper are:1) A geometric feature extraction is proposed based on Non-SubsampledContourlet Transform and Hidden Markov Chain (NSCT-HMC) model. Taking NSCTas the basic multi-scale geometric analysis tool, and HMC as statistical model for NSCTsub-bands, this method reveals the deep statistical feature of NSCT sub-band cofficients.Based on the statistical model, NSCT-HMC feature is proposed by using statisticalmodel discretization, which reveals the deep geometric feature for roads.2) A two-stepped super resolution reconstruction model based on NSCT domainsuper resolution and spatial domain gradient profile priori constraints is proposed. Thegeneral super resolution is performed in Maximum A-posteriori Probability (MAP)model. The MAP model takes the result of NSCT domain super resolution as the initialestimation of high resolution image. The final super resolution result is obtained byusing an iterative approach for solving the MAP model. Though the the super resolution,data source for sophisticated geometric feature of road is provided.3) Considering the strong nonlinear characteristic for high dimensional imagefeature, we use manifold as the basic theoretical framework, propose a noval fusionmethod for multi-source features based on Joint Preserved Mapping (JPM). Through a uniform model, this method can be used as fusion tool for multi-source and multi-scaleremote sensing features.4) A learning-based road extraction framework is proposed. In order to improvethe flexibility and extensibility of road extraction method, we construct road extractionframework based on learning mechanism. Based on road samples, all the cofficients forroad extraction are obtained from samples. So the dependency between method and datais great reduced.Finally, to validate the proposed method, experiments are performd on remotesensing images with different background and forms. The proposed method is alsocompared with some typical methods. The results show that road extraction methodproposed in this paper has great improvement on data applicability and accuracy fordifferent types of roads.