Spatial Structural Feature Description For High Resolution Remote Sensing Images Based On Variogram And Data Field

As an important data source, high-resolution remote sensing images have been widely applied. It is an urgent job to achieve high-resolution image interpretation and information extraction precisely and efficiently. Spatial structural feature is one of prominent features of high-resolution images. Thus, it is meaningful for high-resolution images interpretation and information extraction to describe structural feature efficiently. In essence, spatial structural feature can be reflected by the distribution of neighborhood pixels or dependencies between pixels. This thesis studies spatial structural feature description based on Variogram and data field systematically, analyzes and extends Variogram and data field theory respectively to improve spatial strutural feature description for high-resolution remote sensing images combined with characteristics of high-resolution remote sensing images. Also, the validity of the proposed methods has been verified by the classification and information extraction.The main work of the paper is as follows:(1) The validity of different Variograms in spatial structural feature description of high-resolution images has been studied. Different Variograms have been reviewed and the relationship of them has been analyzed in theory. The relationship between eight Variogram curves of different samples (etc., crops, woodland, road, build-up area) has been analyzed. Experiments show that shapes of eight Variogram curves of each type of samples are relatively similar, except for the function value. In addition, for four kinds of samples, Rodogram and Pseudo Cross Variogram, Madogram and Pseudo-Cross Madogram, Direct Variogram and Pseudo-Cross Variogram show similar characteristics in curve shapes and function values respectively. Furthermore, the effectiveness of different Variograms in characterizing spatial structural feature of high-resolution images has been studied. Spatial structural feature computed by using different Variograms and spectral characteristics have been used to classify high-resolution remote sensing images.(2) The transition region description and extraction based on the Multivariate Variogram has been studied. Transition region can be regarded as coarsening edges in the image on the whole and is located between object and background locally. It is rich in grayscale and has obvious structural feature. By comparing the merit and demerit of several typical transition region extraction methods, a transition region description and extraction method based on Multivariate Variogram has been proposed. The main idea is to use Multivariate Variogram to describe the local structural feature at first and then to extract the transition region from structural feature image based on a threshold operating. After that, a best threshold can be found with the extracted transition region, which can be used to segment the image. Comparing with the local entropy-based method, level difference-based method and data field-based method, experiments indicate that the proposed transition region description and extraction method can get results with better continuity and less noise.(3)The spatial structural feature modeling method based on the extended data field has been studied. Firstly, Variogram and data field method are compared both from theory and application. Despite two functions are essentially the same in theory, the choosing way of pixel pair in two functions is different, which determines the differences in weights. Unlike Variogram, the choosing way of pixel pair in potential function of data field is biased, which causes limitations when the spatial structural characteristic based on data field theory is used to improve the classification accuracy. Furthermore, experiments of two data sets also justify the conclusion. Based on the analysis of the limitation, the extended data field has been proposed in this thesis and the functional relationship between potential function of extended data field and Variogram has been represented. The end of the paper focused on extended data field-based structural feature description. Extended data field has been used to describe the spatial structural feature of high-resolution images and then was combined with spectral characteristics to classify images. Experiments show that the extended data field is superior to the original data field, Variogram, CLBP and Gabor in characterizing spatial structural feature of high-resolution remote sensing images because it can get more local statistical information and reflect the local structural better as well.(4) Structural feature description based on Variogram and data field and build-up areas detection have been studied. Compared with non built-up areas in high-resolution remote sensing images, there are complex internal structures in built-up areas. By analyzing build-up areas in remote sensing images, we summarize the significant structural characteristics as follows:(a) remarkable texture feature,(b) rich line characteristics, and (c) pixels in build-up areas do not exist in isolation and each pixel has interaction with others. And then, we analyze the rules and characteristics of Variogram and data field in describing different structural features. Experiments show that Variogram-based method is fit to distinguish the homogeneous region and heterogeneous region, which reflect the texture of build-up areas very well. While data field-based method is fit to distinguish homogeneous regions with different attributes. It restrains the noise and reflects the connection characteristic of building area well. Furthermore, structural feature description methods based on Variogram and data field have been studied. After analyzing problems and theory, we use Variogram to describe the significant texture structural and use the density of edge points to reflect the rich line feature, after two features are fused effectively to obtain feature-map. Then apply the data field to describe the interaction between pixels in feature-map, fully combining advantages of both two theories in describing the structural characteristic. Finally, build-up areas can be extracted by threshold operating. Experiments show that the proposed method was proved to be superior to Harris and data field-based method, voting-based method in extracting build-up areas.