基于多目标优化的高光谱图像无监督波段选择

发布时间：2018-03-11 19:36

本文选题：波段选择　切入点：无监督　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：近年来,高光谱遥感数据波段选择(Band Selection,简称为BS)的研究由于其在分类和目标识别具有良好的应用受到了学者的广泛关注。高光谱图像数据包含数百个连续波长的窄光谱。其丰富的光谱信息提供了准确的地物识别潜力。但其庞大的数据量在数据的传输计算和存储等方面都带来了很严重的问题。特别是当数据的维数很高的时候,一些波段之间存在很强的相关性,包含大量的冗余信息,这使得分类精度会随着有限的训练样本的维数的增加而下降,即出现Hughes现象。因此很多基于统计的分类方法对高光谱遥感图像是无效的,所以高光谱遥感图像的波段降维是高光谱数据应用的一个很关键的环节。无监督的波段选择以其不需要先验知识、保留原有图像的物理特性、较好地保留有用信息去除波段间冗余性等优点,在高光谱数据的后续地物分类、地物识别等领域得到了广泛的应用。本文对高光谱图像的无监督波段选择进行了系统的研究,所取得的主要研究成果为:通过多目标优化对高光谱遥感数据进行了无监督波段选择的研究。所谓的无监督是指不需要任何先验知识即不需要特定的地物人工标识。针对波段选择的要求,本文将高光谱图像波段选择问题建模为一个多目标优化问题,得到的波段组合权衡了有用信息保留和冗余信息的消除(去相关)两方面,输出一组满足目标函数的非支配折衷解,然后从中选取用户需要的波段组合。提出的方法克服了现有技术的一些不足,得到的波段组合能够较好地保留有用细节信息同时尽可能多的消除波段间冗余信息,更好地适用于降维后地物分类、识别等运用。波段选择问题可以看作为一个组合寻优的问题,所以将多目标优化算法用于无监督的波段选择研究上可以取得不错的效果。本文将以地物分类精度和Kappa系数作为波段选择的评价标准,交叉算子和变异算子作为优化的主要操作算子对于波段的寻优起着很关键的作用,论文里采用了非一致交叉算子和非均匀动态变异算子,提高了对种群局部空间的搜索能力。实验验证了提出的方法具有可行性和有效性。实验利用Indian_pinesCp数据集和Salinas数据集进行了仿真实验,通过与一些已有的非监督波段选择算法进行比较,提出的波段选择算法在选择5-30波段范围内获得了较高的分类精度和Kappa系数以及较强的鲁棒性。
[Abstract]:In recent years, The research on band selection of hyperspectral remote sensing data (BSs) has attracted wide attention due to its good applications in classification and target recognition. Hyperspectral image data contain hundreds of narrow spectra of continuous wavelengths. Its rich spectral information provides accurate recognition potential, but its huge amount of data brings serious problems in data transmission, calculation and storage, especially when the dimension of data is very high. There is a strong correlation between some bands, including a large amount of redundant information, which makes the classification accuracy decrease with the increase of the dimension of the limited training samples. Therefore, many statistical classification methods are invalid for hyperspectral remote sensing images. Therefore, the band dimensionality reduction of hyperspectral remote sensing images is a key link in the application of hyperspectral data. The unsupervised band selection, which does not require prior knowledge, retains the physical properties of the original image. It has been widely used in the field of ground object classification and recognition of hyperspectral data. In this paper, the unsupervised band selection of hyperspectral images is studied systematically. The main research results obtained are as follows: the unsupervised band selection of hyperspectral remote sensing data is studied by multi-objective optimization. The so-called unsupervised means that no prior knowledge is required, that is, there is no need for specific artificial markers of ground objects. Recognition. Requirements for band selection, In this paper, the band selection problem of hyperspectral images is modeled as a multi-objective optimization problem. The band combination obtained is a trade-off between useful information retention and redundant information elimination (de-correlation). Output a set of non-dominant compromise solutions satisfying the objective function, and then select the band combination required by the user. The proposed method overcomes some shortcomings of the prior art. The band combination can preserve useful detail information and eliminate redundant information as much as possible, so it is more suitable for the classification of ground objects after dimensionality reduction. The band selection problem can be regarded as a combinatorial optimization problem. Therefore, the multi-objective optimization algorithm can be applied to the research of unsupervised band selection. In this paper, the classification accuracy of ground objects and the Kappa coefficient will be taken as the evaluation criteria for band selection. Crossover operator and mutation operator, as the main operators of optimization, play a key role in the optimization of the band. In this paper, the non-uniform crossover operator and the non-uniform dynamic mutation operator are used. The proposed method is proved to be feasible and effective. The Indian_pinesCp dataset and Salinas data set are used in the simulation experiment. Compared with some unsupervised band selection algorithms, the proposed band selection algorithm achieves high classification accuracy, Kappa coefficient and robustness in the range of 5-30 bands.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【相似文献】