高光谱图像降维及分割研究

【作者】 杨诸胜；

【导师】 郭雷；

【作者基本信息】 西北工业大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 高光谱遥感数据具有波段众多、光谱分辨率高、数据量大等特点。这使得对地物的分辨更加准确,但这不仅给数据的存储和传输带来了困难,也给数据的处理带来了巨大的挑战。因此如何有效的降低高光谱图像的维数,减少数据量是高光谱图像分析中的一个重要问题。 特征提取和波段选择是目前国内外主要使用的降维算法。主成分分析(PCA)作为一个多元数据分析的工具,在基于特征提取的降维算法中被普遍使用,但这种降维改变了原始波段的物理意义,使图像的解译变得困难。为了克服这个缺点,本文在对PCA在进行深入研究的基础上,提出了四种波段选择算法,并将它们应用于高光谱图像分割中,本文主要工作如下: 1.提出了一种基于权值的波段选择算法。在主成分变换中,各个主成分可以看作是以变换矩阵元素为权值由原始波段进行加权而得到的,因此,变换矩阵元素的一定组合可以反映原始波段的信息含有量,根据它可以选择出有效的波段,实验证明这种方法简单可行。 2.提出了一种基于贡献率的波段选择算法。根据原始数据协方差阵的特征值和特征向量,可以计算各个波段列给定主成分的贡献率,对重要主成分贡献率的和直接反应了波段信息量的大小,因此可根据它选择波段,实验证明,该方法效果较好,且计算量小。 3.提出了一种基于分段主成分分析的波段选择算法。前面两种方法在选择波段的过程中,相关性较强的两个波段可能被同时选择,而且由于主成分变换是一种全局变换,某些局部比较重要的波段可能被漏选。解决这个问题的一个简便的方法是先对原始波段进行分段,在分段的基础上进行波段选择。 4.提出了一种小波变换和主成分分析相结合的波段选择算法。主成分变换可以很好的压缩信息,但由于它的全局性,使得光谱特征和局部特征没有被更好的保留下来;小波分解是在光谱维对每个像元进行的,它可以很好的保留光谱特征和局部特征。该算法充分利用两种变换的优势,先对原始图像进行小波变换,然后在变换域进行波段选择。实验表明,与前三种算法相比,该算法可以将图像降到更低的维数。 5.高光谱图像分割是图像理解、目标跟踪、目标识别等许多图像分析和应用的一个关键步骤,本文总结了常用的无监督图像分割方法,并将本文提出的波段选择算法作为这些分割的降维方法,取得了较好的效果。更多还原

【Abstract】 The characteristics of hyperspectral remote sensing data are numerous channels, high spectral resolution, and large amounts of data, which makes It easy to discriminate objects in the scene, however, the vast amounts of data not only makes it difficult to transport and store them, but also process them. Therefore, it is very important to reduce the dimension in the hyperspectral image analysis.Dimensionality reduction can generally fall into feature extraction and band selection. As a multianalysis tool, Principle Component Analysis (PCA) is commonly used in the feature extraction. However, such dimensionality reduction changes the physical meanings of the original bands, which makes it difficult to interpret the hyperspectral image. In this paper, in order to overcome the shortcoming we carry on a deep investigation into PCA, and propose four band selection algorithms which select a subsets from the original bands, in the end, we apply these algorithms to the hyperspectral segmentation.1. Proposed a Band selection algorithm based on feature weighting. In the PCA, the principle components are weight sum of the original bands, the elements of the transform metric are the weight, a certain combination of the elements of the metric can reflect the importance of certain band which can be used to select the bands, the results of the experiments indicate that the algorithm is sample and feasible.2. Proposed a Band Selection algorithm based on contribution to the principle components. With eigenvalues and eigenvectors of the covariance metrics of the original data, the contribution of a given band to a certain principle component can be calculated, the sum of the contribution of a given band to all important principle components can reflect the information of the band, which can be used to select the bands, the results of the experiments indicate that the algorithm is validate and needs little computation.3. Proposed a Band Selection algorithm based on Segmented PCA. In the previous two band selection algorithms, two bands with high correlation coefficient may be both selected, and because of the globality of the transform, some bands which are important in the local may be unselected, a Segmented PCA can overcome these shortcomings.4. Proposed a Band Selection algorithm based on a hybrid Wavelet-PCA. PCA is sufficient for reducing data volume, however it fails to preserve the spectral and local characteristic of the original data, Unlike PCA, wavelet decomposition which can preserve the spectral and local characteristic focuses on reducing each individual pixel in the spectral domain, this algorithm makes full use of the advantages of PCA and wavelet decomposition, which first performs an initial reduction using a wavelet decomposition, and then the band selection is applied. The results of the experiments indicate the algorithm can reduce the data into a lower dimensionality compared to the previous three algorithms.5. The segmentation of hyperspectral imagery is a key step in hyperspectral image understanding, target tracking, target recognition. In this paper, we summarized someunsupervised segmentation methods that were commonly used of hyperspectral image. In the experiments of segmentation, band selection algorithms proposed in this paper are used to reduce the dimensionality of the hyperspectral image.更多还原