高光谱遥感图像波段选择算法研究

发布时间：2018-03-21 20:38

本文选题：高光谱遥感　切入点：波段选择　出处：《浙江大学》2014年硕士论文　论文类型：学位论文

【摘要】：高光谱遥感是近年来发展起来的具有高光谱分辨率的遥感科学和技术,对地物的分辨识别能力非常高,但由于其高分辨率也对数据处理带来了一系列的挑战,主要是数据量大,冗余信息多,某些波段噪声含量大,这对数据处理的效率和精度造成一定的影响。本文针对高光谱遥感的一系列问题,对不改变波段物理意义又能达到缩小数据源的波段选择算法展开了研究。波段选择算法主要分为监督和非监督两类,对于一般无先验知识的非监督的波段选择算法来说,信息量大,独立性好是选择波段的主要原则。在众多波段选择算法中,线性预测波段选择算法是相对来说原理明确、效率高、结果有效的算法。但经分析,该算法还存在一系列的缺点,造成波段选择的结果非最优,效率也有待提高。针对原线性预测波段选择算法的三个主要问题,本文进行了比较彻底的改进。第一是噪声波段的去除算法,提出了通过计算图像小波域的熵,估计出波段图像的噪声并将噪声较大的波段去除的思路；第二是初始波段选择算法的改进,用偏度峰度、互信息、K-L散度衡量波段的信息量大小,同时用信息量和独立性两个准则来选出初始波段,既考虑了波段的信息量,又提高了初始波段选择的效率；第三是线性预测后续波段选择的改进,每次迭代都去除线性预测误差最小的波段,这样可以逐渐减少数据源,提高波段选择的效率。针对以上改进思路,本文分别用高光谱图像处理非常重要的分类和解混两个应用来对算法进行了实验验证,实验中采用支持向量机、最近邻算法进行分类,用非负矩阵分解进行解混,两个实验都从精度和效率两方面验证了改进的线性预测波段选择算法的优越性,证明了这是一种有效的高光谱图像数据降维方法。
[Abstract]:Hyperspectral remote sensing is a kind of remote sensing science and technology with high spectral resolution developed in recent years. Its ability to distinguish and recognize ground objects is very high. However, its high resolution also brings a series of challenges to data processing, mainly because of the large amount of data. There are many redundant information and high noise content in some bands, which has a certain effect on the efficiency and precision of data processing. This paper aims at a series of problems in hyperspectral remote sensing. The band selection algorithm which can reduce the data source without changing the physical meaning of the band is studied. Band selection algorithms are mainly divided into two categories: supervised and unsupervised. For general unsupervised band selection algorithms without prior knowledge, large amount of information and good independence are the main principles of band selection. The linear predictive band selection algorithm is relatively clear in principle, high in efficiency and effective in the result. However, the analysis shows that the algorithm still has a series of shortcomings, resulting in the result of band selection is not optimal, and the efficiency needs to be improved. Aiming at the three main problems of the original linear prediction band selection algorithm, this paper makes a relatively thorough improvement. First, the noise band removal algorithm is proposed, and the entropy of the image wavelet domain is calculated. The second is the improvement of the initial band selection algorithm, which uses the bias kurtosis and mutual information K-L divergence to measure the information content of the band. At the same time, the information quantity and independence criteria are used to select the initial band, which not only considers the information content of the band, but also improves the efficiency of the initial band selection. Each iteration removes the band with the least linear prediction error, which can gradually reduce the data source and improve the efficiency of band selection. In view of the above improved ideas, this paper uses the very important classification and mixing of hyperspectral image processing to verify the algorithm. In the experiment, support vector machine and nearest neighbor algorithm are used to classify the algorithm. The nonnegative matrix decomposition is used to solve the problem. Both experiments verify the superiority of the improved linear predictive band selection algorithm in terms of accuracy and efficiency. It is proved that this algorithm is an effective method for dimensionality reduction of hyperspectral image data.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】