Research On Multiscale Methods In Object-Based Image Analysis

As a bridge of the Remote Sensing Technology and the Geographic Information System(GIS), the Object Based Image Analysis(OBIA) is a new normal of Remote Sensing Image Analysis, with incomparable advantages than traditional Pixel Based Image Analysis(PBIA). OBIA can provide a strong technical support for automated remote sensing image analysis and image understanding. It became an important branch and research focus of Geographic Information Science. On the other hand, objective world is hierarchical, showing multi-scale characteristics. Objects and phenomena present as meaningful entities only under certain scale, and show different landscape patterns at different scales. As a direct reflection of the real world, objects in remote sensing image are also hierarchical and showing multi-scale characteristics. Therefore, like the PBIA, the OBIA should also be multi-scale. Base on OBIA, this paper studies multiscale phenomena and problems in high-resolution(HR) remote sensing image, and proposes multi-scale object-based analysis methods.Chapter I is an introduction, which describes the background and significance of the work of this paper. This chapter introduces the basic concepts of the OBIA, reviews the emergence and development of OBIA. It analyzes the situation of current researches on OBIA, including the sub-direction, the applications, the advantages and disadvantage and the development trends. It also points out the problems in current researches. Finally, it describes the main work and chapters arrangements.Chapter II studies the construction of multi-scale objects. This chapter analyzes the characteristics of HR images, introduces the basic theory and common methods of image segmentation, summarizes the characteristics and requirements of the HR image segmentation. According to the imaging mechanism of remote sensing image, it presents a image generation model, and proposes multi-scale segmentation algorithm which is efficient and can utilize multi-spectral information comprehensively, called Dynamic Statistical Region Merging(DSRM). The statistical region is built up based on the image generation model. To reduce propagation error on the blurry edges and gradual change regions, the DSRM tries to test the most similar regions first, and dynamically updates the similarities between regions and the test sequences. The experiment shows that, compared with the Static Statistical Region Merging(SSRM), the segmentation accuracy of the DSRM is higher. Although the merging process of the DSRM is more complex than the SSRM, its time complexity is still approximately linear due to improved its sorting process. It can meet the needs of large remote sensing image segmentation. Compared with the business software e Cognition, in the same segment granularity, the segmentation accuracy of the DSRM is higher.Chapter III studies the problem of segmentation scale selection. This chapter introduces the concept and content of the optimal segmentation scale, describes its significance and value in image segmentation, and summarizes its basic characteristics. Then it analyzes the current optimal scale selection method, and categorizes them as the empirical selection method, the model calculation method and the index method. It also analyzes the relationship between the scene complexity and the segmentation scale, presents a method for calculating the scene complexity based on th Waston visual model. Then the scene complexity is used for scale selection for the multi-scale segmentation proposed in the chapter II. The experiment shows that, the performance of this adaptive–scale segmentation is better than any fixed-scale segmentation. So, it is effective to select segmentation scale according to the scene complexity.Chapter IV studies the multi-scale object description, called multi-scale object tree. This chapter describes the purpose and significance of construction the multi-scale object description. A robust image interpretation system should retain the segments in scale-unfixed state, outputing a multi-scale objcet description which is beneficial for comprehensively utilize multi-scale objects and the relationship between objects. It analyzes the differences between the multi-scale object description and the image multi-scale description. The former does not change the image, while directly managements the objects on different scales. It also introduces several common theories and methods for the image multiscale description, and presents a data structure for multi-scale object, called Multi-scale Object Tree(MOT). To quickly build a MOT, a method based on graph theory is preaent. It also provides a variety of operation method on the MOT.Taking the building extraction for example, chapter V describes the application of MOT. It analyzes the characteristics of the building in HR image, and proposes a building model which integrates of the spectral, the texture, the shape, the context and the topology information. Based on the MOT, it proposed a building extraction method which using shadows as contextual information guiding building extraction. Taking large image for experiments, the results show that the method has a good overall performance, and has some practical value.Taking the road extraction for example, chapter V studies the object shape analysis methods based on its skeletons. A reasonable and effective shape analysis can improve the accuracy of the target recognition and information extraction. So, the post-processing is also an important content of the OBIA. It analyzes the shape characteristics of the road, introduces several common road shape filtering methods, and presents an adaptive scale filtering method based on road skeletons. Take the width of each skeleton point as filtering scale, a bi-direction cumulative smoothing can provie road centerline with accurate location and smooth shape, can effectively delete non-road areas and repair roads with incomplete occlusion. Adding a small amount of human intervention, it can jump the complete occlusion on road.Chapter VII is the conclusion, pointing out the future work.The work of this paper has theoretical and engineering significance. On the one hand, it can promote the development of the OBIA by integrating multi-scale analysis into its all aspects. On the other hand, it can also broaden the application of the OBIA in HR image analysis, such as building extraction and road extraction.