Research On The Segmentation Of High Resolution Remote Sensing Image Based On Object-oriented Markovian Random Fields Model

In the past decade, high resolution remote sensing satellites had been launched successfully. And in the next decade; many such satellites with sub-meter resolution will be launched in China. With the improvement of the resolution in remote sensing field, many new problems have emerged, which greatly challenge to the traditional methods for remote sensed image processing.In the high-resolution remote sensed images, details describing different elements of an object have been characterized with rich geometrical information. The size, shape, and adjacent relationship between different objects have been better presented. In accordance with such characters, the object-oriented methods for the analysis of high-resolution remote sensed images is proposed, in which the object is employed as the primary processing unit for extracting features because of it has more structural information than a pixel. In the view of the statistics, high-resolution remote sensed image is a high random signal on account of the high randomness of the spectal responses of the same object and the adjacent relationships between neighboring objects. Hence, it is an effective way to study the processing of high-remote sensed images by random field based methods, among which Markov random field model (MRF) methods have been widely used because of their effective description of spatial information, perfect theory foundation, and easy integration into the Beyesian framework. In this paper, the object-oriented methods are used for the segmentation of high-resolution remote sensed image. Firstly, objects are generated by over segmentation. Then, MRFs are employed to modeling the primary processing units of objects and to obtain the final segmentation results of high-resolution remote sensed images. The major contents of this paper are consists of five parts:(1) Some relative theory are systematically summarized and described, such as the theory of object-oriented image analysis, the theory of MRF-based image segmentation, and the theory of the quantitative and qualitative evaluation methods of segmentation results.(2) Based on the analysis of problems existing in the commonly used processing methods of high-resolution remote sensed images and their statistical characters, an object-oriented MRF-based image segmentation algorithm is proposed, named as object MRF model (OMRF). Firstly, Mean-Shift algorithm is employed to obtain the overly segmented results and the primary processing units are generated, based on which the object adjacent graphs (OAG) can be constructed. Then, MRFs are easily defined on the OAGs, in which special features of pixels are modeled in the feature field model, and the neighbor system, potential cliques and energy functions of OAGs are exploited in the labeling model. Finally, two high-resolution remote sensed image data sets, GeoEye and IKONOS, are used to testify the performance of OMRF. And the experimental results have shown the better adaptability of this method to the segmentation of high-resolution remote sensed images.(3) Based on OMRF and the obvious characters embedded in the high-remote sensed images that there are plenty of textural and shape features as well as spectral features, a multi-feature OMRF (MFOMRF) model is proposed for image segmentation. After the generation of objects by overly segmented, the spectral, textural, and shape feature are extracted for each node in the OAGs, all of these features are constructed in a feature vector, based on which the feature model is defined on the OAG Finally, two high-resolution remote sensed image data sets, GeoEye and IKONOS, are used to testify the performance of MFOMRF. And the experimental results verified that MFOMRF has the capability to obtain better segmentation results, especially for textural and shape richer images.(4) In the high-resolution remote sensed image, there is mass of data to be processed, and land objects exhibits strongly hierarchical and multiscale characters. In order to overcome the disadvantages of pixel-based hierarchical MRF model directly used on high-resolution remote sensed images, a hierarchically multiscale object-oriented MRF model (HMSOMRF) is proposed for image segmentation. Mean-Shift algorithm is employed to obtain multiscale segmentation results, which can form the hierarchical structure according to the correspondence of different objects in different scale, and the hierarchically multiscale object adjacent tree (HMOAT) can be easily defined. After the calculation of the spectral, textural, and shape features of each node, the hierarchical MRF model can be easily defined on the HMOAT for the segmentation of high-resolution remote sensed images. Finally, two high-resolution remote sensed image data sets, GeoEye and IKONOS, are used to testify the performance of MFOMRF. And the experimental results have shown the superiority of this method to the pixel-based hierarchical MRF segmentation method both on the effectively and accuracy, which implies it is suitable for the segmentation of high-resolution remote sensed images.(5) Finally, the validation and practicality of these algorithms proposed in chapter 3,4 and 5, are also used to the land-use classification, and compared with the commercial remote sensed software. Experimental results have shown that our algorithms have higher accuracy than ENVI and eCognition. So we can conclude that our algorithms have better practical and applicable value.