|
|
高光谱成像是将数字成像与光谱技术相结合，探测目标的二维几何空间与一 维光谱信息，并将其获取到的高光谱分辨率的连续且窄的波段的图像数据称为高 光谱图像。这项技术已成功应用于环境监测、农业、航天等领域。
在高光谱图像中，精细的光谱分辨率使得波段间通常具有很高的相关性，造 成了数据的大量冗余。因此波段选择作为数据降维的一种方法，在高光谱图像处 理中是最为关键的步骤之一。波段选择的好坏直接关系到利用所选波段进行高光 谱图像假彩色合成结果的优劣。基于图表示的波段选择方法是众多算法中最有效 的方式之一，它把波段看作高维空间中的点，巧妙地利用聚类的思想找到最具有 代表性且相互之间相关性较小的波段，大大提高了波段选择的效率。然而该方法 在有噪声波段存在时不再适用。
本文针对这一问题，首先介绍了高光谱图像的基本概念和原理，然后在基于 图表示的波段选择方法的基础之上，证明了该方法在所选波段数目大于聚类个数 的情况下会优先选入噪声波段，然后利用噪声对二阶微分算子敏感的特性，引入 拉普拉斯算子对噪声波段进行处理，研究不受噪声影响的基于图表示的改进波段 选择算法，并编写了相关的 Matlab 语言程序。
最后利用真实高光谱遥感图像数据集所进行的实验结果验证了本文方法的 有效性。
关键词： 高光谱图像波段选择假彩色合成图表示拉普拉斯算子
ABSTRACT
Hyperspectral imaging combines the digital imaging with spectroscopy, detecting
the information of target’s geometric space and spectrum. The image obtained by a series of continuous and narrow bands with high spectral resolution is called hyperspectral image. It has been successfully applied to environmental monitoring, agriculture and aerospace, etc.
In terms of hyperspectral image, the bands are usually highly correlated due to the spectrum resolution, which leads to great redundancy in hyperspectral data sets. Therefore, band selection used as an approach for dimensionality reduction is one of the most important steps in hyperspectral image processing. The correctness of the band selection is directly related to the result of false color composite of hyperspectral image. The graph representation based band selection(GRBS) is one of the most effective ways to solve this problem, which greatly improves the efficiency of band selection by interpreting bands as the nodes of a graph in high-dimensional space. It can find the most representative and interrelated bands using the idea of clustering tactfully. However, it is no longer applicable when noisy bands exist.
As to above problem, this paper firstly introduces the basic principle of hyperspectral image, then on the basis of GRBS, we prove that GRBS tends to select noisy bands when the number of selected bands is larger than the number of clusters. In order to solve this issue, the Laplace operator is introduced for processing and propose the algorithm termed improved graph-represented band selection(IGRBS). Also, we have made up the relevant Matlab language program.
Finally, we verified the validity of this method by using real hyperspectral image
data.
Keywords:Hyperspectral imageBand selectionFalse color composition Graph representationLaplace operator
2.3.1 高光谱数据表示方式11
2.3.2 波段间的相关性12
2.3.3 休斯(Hughes)现象12
2.3.4 高光谱数据特性分析小结13
3.2.1 GRBS 准则函数15
3.2.2 参数的选取17
3.2.3 GRBS 实验结果18
3.3.1 情形一证明19
3.3.2 情形二证明21
4.2.1 噪声波段对拉普拉斯算子的响应25
4.2.2 IGRBS 准则函数27
4.4.1 IGRBS 去噪实验32
4.4.2 IGRBS 选择波段进行假彩色合成33
4.4.3 IGRBS 分类精度测试38
第五章 结论与展望40
