Research On Region-based Multi-scale Image Representation And Its Application In Remote Sensing

With the increasing use of very high spatial resolution remote sensing images in environmental, urban, ecological, agriculture and many other applications, object-based image analysis (OBIA) has become the main trendency in remote sensing image analysis. Due to the various kinds of objects on the earth, these objects can not be represented precisely at only one scale. It makes scale the key issue in object-based image analysis. As these different objects can be represented with hierarchal structures, hierarchical multiscale analysis is an effective and efficient solution for this problem. However, there are many theoretical and technical obstacles for multiscale OBIA, such as the precise representation of different objects at different scales, the transition of different scales, hierarchical relations of image objects, and so on. These limitations are great chanllages for multiscale OBIA. Therefore, it is necessary to develop new new techniques for multiscale OBIA, and make it more effective and efficient.To meet the requirements for multiscale object-based image analysis applications, this dissertation studies the region-based multiscale representation of remote sensing images, form theory, method and applications respectively. The main works are listed as follows:1) Firstly, this dissertation studies and summarizes the theoretical basis of multiscale region-based image representation; and then the concept of scale for region-based image representation is defined; finally, a novel data structure for hierarchical region-based image representation is proposed. The proposed structure takes the traditional binary partition tree (BPT) as the foundation, and introduces a scale parameter as index for each node, witch makes the proposed structure is convenient to represent images at arbitrary scales. This work laid a theoretical foundation for later study.2) Secondly, this dissertation studies the implementation of building a hiearichical region-based image representation stuture. The implementation starts from an initial partition of an image, and followed by a sequence hierarchical stepwise region merging process. In this work, three key issues of the implementation are studied:initial partition, region model®ion merging criteria, and the scale index of each node. Three novel ideas&algorithms are proposed:a) a regional adaptive marker-watershed transformation is proposed for the initial partition, b) a multiple feature combinded region model and its accociated merging criteria is proposed for high quality region merging, c) and finally, the image approximation error is proposed as the scale index of each region. Experiments on different remote sensing images have proved that the proposed approach is effective and efficient.3) Thirdly, the redundancy of the hierarchical structure is studied, and an approach for simplify the structure is proposed. In this work, the redundancy of the hierarchial image representation structure is studied. The structure can be observed from two aspects:local and globle, to separate the redundancy nodes from the whole structure. Experiments have shown that the proposed approach can effectively separate the reduncancy nodes, and retain all the useful nodes of the tree structure. Experiments on different images have show the tree simplification process dramatically reduce the number of regions of the hierarchial structure.4) At last, we propose a novel object-based building detection approach based on the multiscale hierarchial region-based image representation structure. In the proposed framework, it is convenient to take advantages of multiscale features of each region, which makes the application not sensitive to the scales, and avoid the traditional iterative image segmentation and object detection processes. In addition, a novel boundary-constrained structure feature is proposed, which is able to separate the whole building regions from its over-segmentation subparts and under-segmentation superparts. This is a crucial guarantee for object detection in the scale sets framework.