卫星多源遥感图像融合技术的研究

【作者】 刘晓翔；

【导师】 郝重阳； 樊养余；

【作者基本信息】 西北工业大学 ， 信号与信息处理， 2004， 硕士

【摘要】 近年来，多源遥感图像融合已经成为遥感应用领域和图像工程领域的研究热点。它是为解决快速、有效处理多源传感器提供的海量数据这一问题而提出的一门新兴技术。多源遥感图像融合技术可以将多源传感器的图像数据进行关联和复合，产生出比单一信息源更精确、更完整、更可靠的估计和判断。图像融合分为三个层次：像素层、特征层和决策层。本论文的工作是在像素层和特征层展开的，取得了一些有新意的成果。论文的主要工作和成果包括： 在像素层，论文研究了多传感器数据融合理论及遥感图像预处理的过程和步骤，归纳了多源遥感图像像素层融合的常用算法，并针对目前遥感数据呈海量化、复杂化这一发展趋势同遥感信息提取的能力和效率滞后这一矛盾，在SFIM算法的基础上，将IHS变换与SFIM相结合，将原算法中的均值滤波器改进为自适应加权均值滤波器，提出了一种改进的SFIM算法，通过对一组多光谱图像和全色图像的双传感器融合仿真对比试验，证明了该算法在保持原多光谱图像光谱信息的同时，能够有效提高融合图像的空间分辨能力。由于该算法简洁，更适合用于那些需要快速交互处理和实时可视化的融合系统。 在特征层，论文研究了基于Markov随机场的图像分类方法。在详细讨论Markov随机场基本理论的基础上，论述了基于MAP-MRF框架的图像分类算法及该算法实现过程中的组合优化问题。针对遥感图像非监督分类中的参数估计问题，重点讨论了EM-MRF迭代算法的原理和实现，并将EM-MRF迭代算法引入到多源遥感图像融合的过程中，提出了两种分别基于集中式融合模型和分布式融合模型的图像融合方法。通过对合成图像和真实遥感图像的仿真，证明了EM-MRF分类方法有较高的分类精度和鲁棒性，本文的融合方法能进一步提高分类的精度。更多还原

【Abstract】 In recent years, multisensor image fusion techniques have attracted extensive attention in remote sensing application and image project area. Multisensor remotely sensed image fusion technique can combine multisensor images and produce a more precise, integrated and reliable estimation and description of them than a single image. According to the level in which the fusion implements, image fusion is divided into pixel-level fusion, feature-level fusion and decision-level fusion. The research in this paper is executed mainly at pixie level and feature level image fusion.At pixel level, on the basis of studying the elementary theories of multisensor data fusion and procedures of pretreatment in remote sensing image fusion, this dissertation summarizes the common methods which are applied in multisensor remotely sensed image fusion. Afterwards, in order to decrease the contradiction between the more complex and mass remote sensing image data and relatively slow speed of information extraction, an improved SFIM image fusion method is proposed. This modified algorithm is on the base of SFIM fusion technique, combines IHS method and SFIM method and then replaces the former mean filter by an adaptive weighted mean filter. Compared with the results of several common fusion techniques through a set of simulation tests between multispectral images and panchromatic images, it is proved that the new method can get an excellent result for the aim of improving spatial resolution while preserving the spectral information of multispectral images. Due to its simpleness, the proposed approach is more applicable for fast interactive processing and real time visualization.At feature level, this paper focuses on Markov random field-based image classification algorithms. Firstly, the cardinal theories of Markov random field, MAP-MRF based image classification algorithm and the combinatorial optimization methods during its implementation are elaborated. And then the paper describes EM-MRF iterative algorithm and its realization for the parameter estimation in unsupervised image classiifcation process. The EM-MRF-based image classification strategy is introduced into multisensorfeature-level image fusion, distributed and centric based fusion methods are proposed. Finally, simulated results through sythetic and real remotely sensed image illustrate the effectiveness and advantage of the proposed methods.更多还原