Hyperion高光谱影像坏线修复的局部空间-光谱相似性测度方法

Hyperion高光谱影像中的坏线将直接影响后续应用的准确性。针对Hyperion高光谱辐射率数据的特点，考虑影像中坏线像元与邻近像元在空间和光谱上的相似性，提出了一种局部空间-光谱相似性测度（local spectral-spatial similarity measure，LS₃M），以实现对Hyperion高光谱数据的描述和坏线修复。LS₃M由空间和光谱两部分的相似性测度构成，前者为欧氏距离度量，后者组合了Canberra距离和光谱相关角（spectral correlation angle，SCA）。考虑到Hyperion高光谱不同波段的辐射率特性，引入信息熵对SCA进行约束。针对相似像元的邻近搜索问题，引入相似度均值与方差对光谱相似性阈值进行动态调整。为验证该方法的有效性，选取了沙漠、草原、森林、城郊、沿海城市和内陆城市6种典型场景的Hyperion高光谱数据进行模拟坏线的定量误差分析和真实坏线的定性评价；通过与邻域均值法及常规光谱相似性测度的对比，证实LS₃M法坏线修复精度更高，稳定性更好。

A Local Spectral-spatial Similarity Measure for Bad Line Correction in Hyperion Hyperspectral Data

Hyperion hyperspectral data are frequently contaminated with bad lines, which directly affect the accuracy of the following-up applications. In this paper, a local spectral-spatial similarity measure, or simply LS3M, was proposed, in aids of the Hyperion hyperspectral data characterizing and bad line repairing. Not only the spatial similarity between target pixel and neighborhood similar pixels in the Hyperion hyperspectral image, but the similarity in spectral dimension was considered. That is, LS3M was constituted of spatial similarity measure and spectral similarity measures. The former was measured by Euclidean distance and the latter one combined Canberra distance (CD) and spectral correlation angle (SCA). Considering the radiance characteristics of Hyperion hyperspectral at different bands, information entropy was introduced to constrain the combination of CD and SCA measures with respect to the discrepancies of Hyperion radiance characteristics in different wavelengths. In terms of neighbor search of similar pixels, this paper introduced similarity mean and variance to realize the dynamic setting of the threshold of the spectral combination process. To verify the proposed method, Hyperion hyperspectral data of six typical scenes (i.e., deserts, grasslands, forests, suburbs, coastal cities and inland cities) were utilized to make an experimental study on simulated bad lines and real bad lines in Hyperion image. Besides, a comparison with neighborhood average method, single spectral similarity measures and combined spectral similarity measures was conducted. These results demonstrated that LS3M owned a higher accuracy and better stability on bad line repair, especially in depicting object boundary and topological relations between ground objects.