Application Of Fusion Model Of Dempster-Shafer Evidence Reasoning In Remote Sensing

As the development of the sensor techniques, remote sensing image information fusion has been widely applied to feature recognition, image classification, information extraction and so on. Therefore, it has been an international research hotspot in remote sensing. But because of the multidimensional specialities of fusion data, there are some problems existing in fusion processing, including uncertainty, inaccuracy, imperfection, inconsistency and time-dependence. A fusion model based on Dempster-Shafer(DS) evidence reasoning is an effective approach to these problems. Aiming at the application of this model to remote sensing image classification assisted by computer, more detailed research works will be done in this thesis.Firstly, pretreatment of information sources is necessary. ETM images are the main information data utilized in this paper, firstly, image correction of these data are carried out; then, the statistical analysis of the information content and spectral feature analysis of target objects are conducted; otherwise, make use of pixel-based fusion to integrate features of various sensors to integrate features of images with different spatial resolution.Secondly, application of classification by the fusion model of Basic DS Evidence Theory(BDSET) in remote sensing. The latest theory and application advances are summarized. The basic concepts, including discernment frame, Basic Probability Assignment(BPA) function and evidence combination rules, etc. are introduced and the main principle and the basic characteristics are presented. According to the application principle of BDSET in remote sensing, the classification flow of BDSET is mainly discussed. Where, four aspects are emphasized—determination of discernment frame, that is, number of the classes; acquisition of reasonable BPA function; realization of Dempster combination rule and assessment methods of classification precision. The experimental results show that better accuracy can be acquired with this DS fusion classification model than the conventional supervised minimum- distance classification method.Thirdly, improvements of BDSET fusion model are presented. For the limitation and problems of BDSE, the improvements are conducted from five aspects: (1) Extending the original algorithm by getting rid of the one subset limit. BPA function of the original algorithm is altered and the attenuation factor is introduced in order to determine the uncertainty of BPA function quantificationally.(2) Conducting PCA Transformation with image sources to reduce the correlation between difference sources.(3) Improving the combination rule to deal with evidence conflict.? Establishing decision-making trees to construct layered classification DS model. According to the spectral comparability, rough classes are acquired and subset classes are gained based on rough class. Layered classification methods are utilized to the analyzed space for decreasing computation complexity of the model.(5) Texture features are brought into BDSET, the classification precision are improved and classification results with four common texture are compared.(6) At last, taking expansibility of BDSET into consideration, the fuzzy logic is integrated with BDSET and a fusion model based on Fuzzy DS Evidence Theory(FDSET) is applied to remote sensing classification. BPA established by the membership of learning samples, combination of evidences and incorporation of information sources are considered furtherly.The experimental results suggest that those methods mentioned above are effective in improving classification accuracy and capability of algorithm. Compared with the BDSET fusion model in lots of experiments, these methods reduce the misclassifying instances effectively.In general, some theoretical discussion and method researches are carried out about classification in remote sensing with fusion models of uncertainty theories. Some researches of this paper are attempts of developing and perfecting computer-aided remote sensing decision-level classification.