基于遗传算法和MRF的亚像元定位方法研究

发布时间：2018-09-13 16:58

【摘要】：高光谱遥感是一种多维信息获取技术,它不仅可以获得描述地物分布的二维空间信息,而且可以获得对应地物的一维光谱信息。高光谱遥感图像的光谱分辨率很高,随着光谱分辨率的提高,其对地物的认知能力也不断的提升,但是高光谱遥感图像的空间分辨率仍然很低,混合像素普遍的存在于高光谱遥感图像中。针对混合像素的处理,硬分类方法会导致地物信息的大量丢失,正因为如此,提出软分类方法,具体来说包括端元提取,丰度反演和亚像元定位三个部分,端元提取算法提取高光谱图像中所含有的端元,丰度反演是计算各端元在混合像素中所含有的丰度,亚像元定位技术预测各端元在混合像素中的分布。本文针对高光谱遥感图像亚像元定位方法中一些关键问题进行了研究。具体工作如下:首先,对亚像素和像素之间的吸引力模型(SPSAM),像元交换算法(PSA)等基础的亚像元定位算法进行了研究。SPSAM直接对亚像素进行赋值,而且吸引力值计算方法极为粗糙,使其亚像元定位结果中出现很多独立的像素。PSA具有高效的迭代速度,但是其缺点是对噪声和亚像元的初始分布非常敏感。其次,对遗传算法在亚像元定位中的应用进行了分析,由于遗传算法中交叉算子选择所交换基因的随机性,使得其迭代效率很低,最终的亚像元定位结果精度也不高。本文提出了一种基于改进的遗传算法(MGA)的亚像元定位算法,该算法既结合了遗传算法(GA)中种群思想的优点,又结合了像元交换算法(PSA)中高效的迭代速率的优点,使其迭代效率进一步增强。最后,上述亚像元定位算法是以光谱解混所得的丰度图像作为输入,由于现有的光谱解混算法很难达到其精度要求,使得最终亚像元定位结果存在误差的叠加,精度无法进一步提高。针对这些算法,本文首先描述了马尔科夫随机场(MRF)在亚像元定位中的应用,由于MRF可以结合空间和光谱信息,进一步描述了基于多光谱约束MRF的亚像元定位算法,虽然基于多光谱约束的MRF亚像元定位算法可以进一步提高SPM精度,但是由于其没有考虑亚像素平移图像(SSRSI)的空间信息,使得其精度有限,针对此缺点,本文提出了基于多空间约束和多光谱约束的MRF亚像元定位算法,进一步提高了亚像元定位的精度。
[Abstract]:Hyperspectral remote sensing is a multidimensional information acquisition technique, which can not only obtain two-dimensional spatial information to describe the distribution of ground objects, but also obtain one-dimensional spectral information of corresponding objects. The spectral resolution of hyperspectral remote sensing image is very high. With the improvement of spectral resolution, the cognitive ability of hyperspectral remote sensing image is improved, but the spatial resolution of hyperspectral remote sensing image is still very low. Mixed pixels are common in hyperspectral remote sensing images. For the processing of mixed pixels, the hard classification method will lead to a lot of loss of ground object information. Because of this, a soft classification method is proposed, which includes End-element extraction, abundance inversion and sub-pixel location. End-element extraction algorithm extracts endelements from hyperspectral images. Abundance inversion is used to calculate the abundance of each endelement in the mixed pixel, and sub-pixel localization technique is used to predict the distribution of each endelement in the mixed pixel. In this paper, some key problems in sub-pixel localization of hyperspectral remote sensing images are studied. The main work is as follows: firstly, the subpixel location algorithm based on the attraction model between sub-pixel and pixel, such as (SPSAM), pixel exchange algorithm (PSA), is studied. SPSAM directly assign sub-pixel value, and the calculation method of attraction value is very rough. Many independent pixels. PSA in the sub-pixel localization results have high iterative speed, but their disadvantages are that they are very sensitive to the initial distribution of noise and sub-pixel. Secondly, the application of genetic algorithm in sub-pixel location is analyzed. Because of the randomness of crossover operator selection, the iterative efficiency of genetic algorithm is very low, and the precision of final sub-pixel localization is not high. In this paper, a sub-pixel localization algorithm based on improved genetic algorithm (MGA) is proposed. This algorithm combines the advantages of population idea in genetic algorithm (GA) and the efficient iterative rate in pixel exchange algorithm (PSA). The iteration efficiency is further enhanced. Finally, the above sub-pixel localization algorithm is based on the abundance image obtained by spectral unmixing. Because the existing spectral de-mixing algorithms are difficult to achieve its precision requirements, the final sub-pixel localization results have a superposition of errors. The precision cannot be further improved. In view of these algorithms, this paper first describes the application of Markov random field (MRF) in sub-pixel localization. Because MRF can combine spatial and spectral information, the sub-pixel localization algorithm based on multi-spectral constraint MRF is further described. Although the MRF sub-pixel localization algorithm based on multi-spectral constraints can further improve the accuracy of SPM, but because it does not consider the spatial information of subpixel translation image (SSRSI), its accuracy is limited. In this paper, a MRF sub-pixel localization algorithm based on multi-spatial constraints and multi-spectral constraints is proposed, which further improves the accuracy of sub-pixel localization.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】