高光谱遥感能以纳米量级宽度的窄波段及多达数百个的波段,对目标进行连续的光谱成像,但其海量数据及相邻波段高度相关造成的数据冗余却制约着它的应用。因此,对高光谱遥感影像分类须进行有效的处理、寻找最优特征,以增强地物的最大可分性。本文首先针对EO-1 Hyperion高光谱影像波段维数高、相关性强和数据量大等特点,利用独立成分分析(ICA)方法进行影像特征提取,并提出一种改进的ICA算法(M-ICA)。试验证明该算法的运算效率明显高于传统FastICA算法,其平均迭代次数和平均迭代时间分别仅为后者的14.49%和17.32%。然后为验证该方法在地物类型提取方面的有效性,分别对M-ICA的特征提取结果,主成分分析(PCA)的特征提取结果和未经特征提取的原始数据进行分类试验。结果表明,经M-ICA处理后的数据的地物类型提取总体精度达到90.57%,比后两者分别高出约10.8%和20.0%。由此表明,M-ICA是一种收敛速度很快的地物类型影像特征提取算法,能有效地实现对高光谱影像数据降维并提高地物特征的可分性。

Hyperspectral remote sensing is a frontier technology that is currently being investigated by researchers and scientists with regard to the detection and identification of minerals, vegetation, crops, ocean and other backgrounds. Hyperspectral remote sensing imagery has tens or even several hundreds of bands with bandwidth of nanometer level and contains abundant spectral information, but its redundancy made by mass data and high correlation in adjacent bands restricts the application of hyperspectral remote sensing. So, rational feature extraction will be required to find the optimal characteristics for maximum separability before the hyperspectral imagery classification. This thesis, taking Hyperion hyperspectral visible/infrared sensor aboard an earth observation satellite platform (EO-1) for example, applied independent component analysis (ICA) to carry on the feature extraction on Hyperion hyperspectral imagery with high dimensions, a strong correlation and huge data. Then, a modified independent component analysis algorithm (M-ICA) was put forward. Finally, the results by the algorithm (M-ICA) and principal component analysis (PCA) were compared with experimental data of selected test area, while using the maximum likelihood classification method to classify objects to verify its validity on object type extraction. Experimental results show that the operation efficiency of M-ICA algorithm is apparently higher than the traditional FastICA algorithm, which average iteration number and time are only 14.49% and 17.32% of the latter. In addition, the total classification accuracy of M-ICA result image is up to 90.57%, which is raised 10.8% and 20.0% more than that of PCA result image and that of Original Hyperion imagery. Therefore, the results indicate that M-ICA is a good feature extraction method with fast convergence speed, can effectively realize the dimension reduction of the hyperspectral imagery, and improve the objects separability in different categories.