Research On Multisource Image Fusion Algorithm And Its Applications Based On Statistics And Reasoning

Recently, with the rapidly development of remote sensing technique, different types of sensors reveal the characteristics of each landcover from different aspects, and form multisource data of the same region. How to combine the multisource data of huge quantity to one unity, to extract more integrated and exact information, and offer proof for man to investigate resources, environment, and disaster, is a key thesis in the field of remote sensing image processing. Based on the imaging principle of different sensors, multisource image fusion can take use of different assembled modes, to realize image enhancement, improve image quality, enhance image classification and target identification, and fusion results comprise more integrated and exact information than each single source. This thesis realizes the improvement of preserving spectral information based on the analysis of pixel-level fusion principle, with the consideration of some idiographic sensor, and discusses feature-level and decision-level fusion algorithms, to enhance the precision of landcover identification and classification.Firstly, based on the analysis of some conventional pixel-level fusion algorithms like IHS, PCA, SCN, one problem is that when they enhance the spatial information, they will also destroy the spectrum. This thesis presents the statistical modification method to improve these algorithms, which is to adjust some statistical component of high spatial resolution image to match that of Multi/Hyperspectral image. Experiments results show that after statistical modification methods to improve the original algorithms, the spatial information of fusion images is enhanced, and the phenomenon of spectrum distortion is improved.Then, in order to fuse the SAR and multispectral images, which are of different imaging mechanism and to apply the fusion results to landcover classification, the thesis researches in the fusion algorithm based on feature extraction. The fusion algorithm firstly extracts texture features of SAR, and feature image is used to form a feature set together with spectral features of multispectral image, this feature set is finally used as input for Maximum Likelihood Classfier for landcover classification. Experiments are conducted and the results show that classification accuracy by the proposed algorithm is enhanced compared with pixel-level fusion and no-fusion, respectively.Finally, Two applications of fusion, which are multisource target-region identification and multisource classification, are conducted. This thesis discusses the fusion algorithm based on the DS Theory of different sensors, which form the mass functions by the single sensor decision, and then combine these mass functions, then gain the fusion results by the decision rule. Based on this, an improved method of the construction of mass function is put forward. and the results show that, uncertain of single source is reduced. Result of classification and identification is more exact and integrated than that of single source.