Automatic Change Discovery Approaches For Typical Geographic Elements Integrating GIS Vector Data And Deep Learning

The automatic updating of GIS geographic elements by using remote sensing images has been a long-standing prominent and difficult research and development issue in the field of remote sensing and GIS.It has great theoretical and application value.Image-image,and image-vector change detection is one of the core technologies for this goal.Due to the complex spatial relationship of high-resolution remote sensing images,the issue of "same spectrum from different objects and same objects with different spectrum" is serious.Moreover,in the process of change detection,it is easy to face the problem of pseudo-change caused by the factors such as sensor angle deviation,illumination radiation difference and the confusion of ground features.So far there is a significant gap in this direction.This dissertation leverages automatic change discovery approaches of typical elements of basic geographic information data as the research goal,and uses two-phase high-resolution remote sensing images,GIS vectors and other auxiliary data as the main data sources.We introduce the deep learning technology,a representative new intelligent algorithm,and combine with GIS domain knowledge to explore the automatic change discovery approaches of typical elements.The main contributions of the dissertation are as follows:(1)A framework for automatic change discovery of typical elements integrating GIS vector data and deep learningA pass-specialize-refine framework for automatic change discovery of typical elements integrating GIS vector data and deep learning is proposed.We first use the multi-temporal and multi-source remote sensing images to quickly obtain general changes.On this basis,GIS vector semantics and deep learning are coupled to realize the automatic change discovery of thematic features.The introduction of GIS realizes the self-evolution of deep learning network model and typical element sample database.The incorporation of GIS realizes the self-evolution of deep learning network model and typical element sample database.By combining the empirical domain knowledge,using the expert knowledge base and deep convolution neural network,the accuracy of typical element change detection results is further improved.(2)Object-based change detection approaches based on multi-feature ensemble classifiersIn order to solve the problems of poor universality and more pseudo-changes in complex background interference,this dissertation proposes a novel multi-feature fusion change detection approach based on visual saliency and random forest.The method analyzes the advantages of visual saliency in the judgment of real change areas,and uses the results of saliency analysis to guide the automatic selection of object-level samples.Then by combining the random forest classifier,the accurate judgment of the real change area is effectively solved.In order to further overcome the false detection problem in the results,on the basis of combining the historical land use map,this dissertation proposes a novel multi-feature fusion change detection approach based on rotation forest and coarse-to-fine uncertainty analyses.The method uses object scale set constraints and a rotating forest classifier to compare the changes of image objects from coarse-to-fine.It unravels the problem that the traditional method directly uses vector polygon as a unit and local changes in the polygon are difficult to detect.Multiple experimental results prove that the proposed approach is superior to the state-of-the-art methods,and can further improve the accuracy of the change detection results.(3)A novel water body automatic change discovery approach based on improved U-Net network and a superpixel-based conditional random field modelIn view of the problems that the traditional water body extraction methods have parameters that are not self-adaptive,the edges of the extraction results are rough,and the positioning is inaccurate,this dissertation proposes a GIS-driven method for automatic change discovery of water bodies based on improved U-Net network and a superpixel-based conditional random field model.The method adopts a fully-supervised learning semantic segmentation strategy,and utilizes the priori information of historical land use vector map to enrich the water body sample database.It adaptively adjusts the parameters of the deep learning network and improves the accuracy of water body extraction in the new phase image.Experimental results prove that the proposed approach is superior to the state-of-the-art methods.The water body extraction results are closer to the reference image,indicating the superiority of the proposed approach.Afterward,by combining the change detection results and the prior knowledge of historical land use vector map,the old temporal water bodies is spatially superimposed with the new temporal water bodies.By making full use of the domain knowledge base,on the basis of change detection and two-phase water body comparison analysis,the corresponding spatial analysis rules of water body change discovery are established,which greatly improves the accuracy of the final results.(4)A novel building automatic change discovery approach based on GIS vector semantics and multi-scale residual concatenation networkAiming at the problem that traditional building extraction methods cannot effectively express building features,resulting in inaccurate building recognition in complex backgrounds,this dissertation proposes a novel building automatic change discovery approach based on GIS vector semantics and multi-scale residual concatenation network(MRCNet).The proposed approach uses the prior information of historical land use vector map to collect the building samples,and constructs the building sample database.On this basis,it uses the trained MRCNet network to extract buildings of the new temporal image.The MRCNet network can adaptively learn and analyze from shallow local features to deep abstract features.The improved network adopts multi-level feature integration and multi-scale feature integration strategies to obtain rich multi-scale context feature information.The proposed approach uses GIS vector constraints to eliminate the obvious "stitching traces" generated at the image stitching when extracting large-scale buildings.Compared with the existing building extraction network model,the proposed method has a significant improvement in building extraction accuracy,which can not only maintain the integrity of the large building roof,but also further reduce the missing inspection of small buildings.Then,by combining the change detection results and the prior knowledge of historical land use vector map,the old temporal buildings is spatially superimposed with the new temporal buildings.By making full use of the domain knowledge base,on the basis of change detection and two-phase building comparison analysis,the corresponding spatial analysis rules of building change discovery are established.Experiments show that the proposed approach can make up for the difficulty of determining the changed building objects in the complex background.It effectively compensates for the problem of insufficient extraction of changed building objects,and can achieve robust determination of changed building objects.In order to further verify the effectiveness and feasibility of the proposed method in practical application,this dissertation develops an experimental prototype system and takes the 1:50,000 terrain feature data regulations as an example.By summarizing the mapping rules of typical elements and analyzing the differences between the selection index of typical elements and remote sensing image,this dissertation refines the change discovery results of typical elements.In addition,the final result can better serve the requirements of basic geographic information data update.Through several experimental cases,the proposed approach can be successfully applied to the automatic change discovery of these typical elements.And the results show that the proposed approach is valuable in some certain applications.