【作者】 刘锋；

【导师】 蓝天；

【作者基本信息】 北京理工大学 ， 光学工程， 2008， 硕士

【摘要】 高光谱数据能提供地物数十至数百个窄波段的光谱信息,通常波段宽度小于10nm,可产生连续的光谱曲线,具有极高光谱分辨率。然而,由于高光谱图像维数众多,数据量巨大,给后续处理带来巨大的挑战。本文以Hyperion图像为对象,研究了高光谱的特征提取和端元提取技术。论文首先介绍了高光谱数据处理面临的问题,包括高光谱的巨大数据量如何处理,高光谱图像混合像元的产生以及端元的提取,分析了对高光谱图像进行特征提取和端元提取的必要性。针对PCA特征提取算法反映地物的整体信息,忽略细节信息的缺点,对小波PCA算法进行改进,通过使用不同的小波基,运用小波PCA算法进行特征提取,最终选取最优的小波基。研究投影寻踪算法中的模拟退火优化算法,将此算法应用于寻找最优的投影向量。分析PPI端元提取算法,将进行预处理的PCA算法换成小波PCA算法,增加了对地物细节的反映,由此进行端元提取取得了较好的效果。论文还分析了高光谱图像的处理流程,包括图像的几何校正,图像的分类技术(K-mean分类,最大似然分类),高光谱图像镶嵌,并完成了通过高光谱图像提取地物的光谱曲线。对所开发的各算法的应用与演示,都取得了良好的结果。更多还原

【Abstract】 Hyperspectral data can provide hundreds of narrow-band spectral information with the band width of less than 10 nm. It can produce a continuous spectrum, with very high spectral resolution. However, due to the hyperspectral image dimension of the large volume of data, we have to deal with these follow-up challenges. In this paper, by studying Hyperion hyperspectral image, feature and endmember extraction technology is developed.First,the problems of hyperspectral data processing are introduced, including how to deal with huge volume of hyperspectral data, how mixed-pixels developed in hyperspectral image and endmember extraction. the necessity of hyperspectral image feature and endmember extraction is analyzed.Feature extraction algorithms for PCA reflects the overall features of information, but overlooks details of the feature. Due to this shortcoming, the wavelet PCA algorithm is developed with different wavelet basis;ultimately the best wavelet basis is selected. Simulated annealing in Projection Pursuit algorithm is studied, and then this algorithm is applied to find the optimal projection vector.By analysis of PPI endmember extraction algorithms, the PCA pre-processing algorithm can be replaced by wavelet PCA algorithm. The wavelet PCA algorithm can increase reflections of the surface feature details, so the PPI endmember extraction algorithm has achieved fairly good results with the improved algorithm.The paper also analysis the process of hyperspectral data processing,including image geometric correction, classification(K-mean classification, the maximum likelihood classification),hyperspectral image mosaic, and the spectral feature curve extraction from hyperspectral data.All the algorithms accomplished in this paper carry out excellent results in applications and demonstrations.更多还原