High Resolution Remote Sensing Image Multi-scale Segmentation Support By Spectral Graph Theory

Object-oriented remote sensing image analysis techniques is of great advantage in dealing with spatial relations, and has become one of the main research directions in remote sensing image analysis and processing. Object-oriented image segmentation is one of the key steps in object-oriented remote sensing image analysis. High resolution remote sensing image has many features of land objects, such as various types, complex structure and rich layers. To better analyze and interprete geographical phenomenon, multi-layer and multi-scale segmentation of land obejects should be implemented. Till now, a lot of multi-scale segmeniation algorithms have been developed. However, most of them are not aiming at high-resolution remote sensing image segmentation and have difficulty to communicate and use layers of information. In view of these limitations, this dissertation introduces the spectral graph theory into the image segmentation process to explore an effective multi-scale segmentation mechanism for high-resolution remote sensing image based on the spectral graph theory. The nature of multi-scale segmentation of remote sensing images is variation of land object level and change of associated obeject number, in which data dimension number changes. The dissertation considers matrix spectrum defined by the graph based on spectral graph theory, and further studies the information contained in the graph, and establishes the relation between the discrete and continuous space through the methods of geometry, analysis and algebra. Spectral graph theory describes the similarity between objects in order to reflect the structural properties and features of different objects or layers with eigenvalue decomposition to express and analyze problems in dealing with complex them.The dissertation starts from the multi-scale characteristics of ground objects on remote sensing images, proposes a new approach for multi-scale image segmentation based on the spectral graph theory. The structural relationship between objects and layers are described and analyzed by applying spectral graph theory to analyze spectrum information of remote sensing images, that is, eignvalues and eignvectors. At the same time, the involved key theories, processing algorithms and applications are researched and the major works are listed as follows:1. Based on the study of image clustering task, this dissertation summarizes and analyzes the internal relations of the spectral graph theory and image segmentation, and summarizes the methods of image processing and analysis based on the spectral graph theory. From the characteristics of remote sensing data, it analyzes the close relationship between the spectral information and segmentation results and the features of spectral methods to segment the image. This dissertation choses the spectral clustering algorithm—Normalized Cut in spectral graph theory to accomplish object-oriented high-resolution image segmentation.2. Two methods of multi-scale image edge extraction are proposed based on gray and texture features and the spectral clustering method introducing spatial relations. The oriented edge of grayscale images is detected by the oriented gradient information based on texture features and intensity features. Histogram difference method is used to extract the texture region edge of the image. The multi-scale segmentation results is obtained by setting a reasonable weight of the two steps. Based on the analysis and comparison of the traditional edge detection operators, this dissertation proposes the oriented energy method to model the image features at first, then computes the oriented energy to measure the similarity of pairs of pixels and introduces spatial relations to modify them. Finally, the image edge detection is done by spectral clustering.3. This dissertation presents a hierarchical segmentation method of high-resolution remote sensing image based on oriented watershed segmentation. The method consists of two steps:firstly, over-segmentation image is generated by the oriented watershed segmentation method. The process is that the image overall edge strength is firstly obtained by weight combination of the oriented edge multi-scale grayscale images, the oriented edge of texture regions and the edge detected by the spectral method. And then the the initial region segmentation results is obtained from the above edge images using the traditional watershed transformation and the corrected edge strength by fitting edge direction. Secondly, the multi-scale image segmentation results is obtained by merging region of the previous step according to the strength of the regional boundary.4. A hierarchical segmentation based on spectral clustering (HSSC) algorithm for high-resolution remote sensing image is proposed, which is multi-level and multi-scale segmentation algorithm based on graph theory. It introduces the weighted aggregation rules in algebraic multigrid technology to spectral clustering and clusters in bottom-up way by the combined strategy, then the fine-grained to coarse-grained map tree diagram is generated. Finally, optimal image segmentation in the fine-grained graph has gradually transformed into the coarsening graph with less nodes and fast image segmentation is completed. The HSSC algorithm uses the characteristics of spectral clustering that it do not simply assume the data distribution in the image-level segmentation, then it can approximately guarantee the global optimal solution. At the same time, image multi-level segmentation reduces the computing time by the use of pyramid structure in top-down way. Experiments show that the proposed algorithm can be applied to high resolution remote sensing image segmentation, and the segmentation results are expressed as natural and irregular cluster structure which reflects the objective requirement of hierarchical segmentation.In summary, segmentation experiments were done through aerial imagery and high-resolution IKONOS remote sensing, and their results were analyzed and evaluated. The experimental results show that segmentation method used in this dissertation is reasonable and effective.