基于进化优化的高光谱特征选择算法研究

发布时间：2018-03-05 20:12

本文选题：高光谱遥感　切入点：特征选择　出处：《华东师范大学》2014年硕士论文　论文类型：学位论文

【摘要】：高光谱特征选择是在保留图像本身物理信息的基础上,找出最具有代表性的波段组合的搜索过程。它本质上是一个全局优化过程。进化算法是一种模拟生物自然进化、基于种群的全局优化方法。本文在分析比较传统搜索方法的基础上,尝试将进化算法应用于高光谱图像特征选择,主要的研究内容包括：(1)首先,根据高光谱特征选择问题与进化算法相结合所要面临的问题,通过设计合适的编码方式和目标函数,给出了相应的解决方法。以新目标函数为评价准则进行基于进化优化的高光谱特征选择。实验证明基于新目标函数的进化搜索算法得到的波段组合图像具有比较好的分类性能。(2)其次,在进化算法的基础上,添加了局部搜索算子,提出基于Memetic算法的高光谱特征选择。在对经典的最佳指数法做出改进的基础上,从高光谱图像波段间相关系数和单幅波段图像的信息量两个方面设计局部搜索算子。本文中通过Online择优和Offline择优两种方式对Memetic算法进行了研究,并且在Online择优方式中提出了两种局部搜索方法。实验结果表明添加了局部搜索算子的Memetic高光谱特征选择算法不但提高了特征选择能力,还提高了分类能力。(3)最后,我们尝试将多目标进化算法应用于高光谱特征选择,从而避免单目标进化算法中部分参数选取的随机性。
[Abstract]:Hyperspectral feature selection is a search process to find the most representative band combination on the basis of preserving the physical information of the image itself. It is essentially a global optimization process. On the basis of analyzing and comparing traditional search methods, this paper attempts to apply evolutionary algorithm to feature selection of hyperspectral images. The main research contents include: 1) first of all, According to the problems faced by the combination of the hyperspectral feature selection problem and the evolutionary algorithm, the appropriate coding method and objective function are designed. The corresponding solutions are given. The new objective function is used as the evaluation criterion to select the hyperspectral features based on evolutionary optimization. The experimental results show that the band combination images obtained by the evolutionary search algorithm based on the new objective function have a comparison. Good classification performance. Based on the evolutionary algorithm, the local search operator is added, and the hyperspectral feature selection based on Memetic algorithm is proposed. The local search operator is designed in terms of the correlation coefficient between bands of hyperspectral images and the amount of information in a single band image. In this paper, the Memetic algorithm is studied by means of Online preference and Offline preference. The experimental results show that the Memetic hyperspectral feature selection algorithm with local search operator not only improves the feature selection ability, but also improves the classification ability. We try to apply multi-objective evolutionary algorithm to hyperspectral feature selection to avoid the randomness of partial parameter selection in single-objective evolutionary algorithm.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】