基于分量替换高分辨率遥感图像融合方法的对比研究

基于分量替换的高分辨率遥感图像融合是一种十分重要的融合方法，但对该方法的融合原理的深入分析在国内还鲜有报道。为此对分量替换融合方法进行原理探究和实验对比分析。首先从线性代数的角度来阐述分量替换融合算法的实质，并对两种典型的构造可替换波段的分量替换方法（基于光谱响应函数模拟低分辨率全色波段的Gram-Schmidt融合法（GS1）和基于多元一次线性回归拟合低分辨率全色波段的Gram-Schmidt（GS2）融合法进行原理说明；其次，选取QuickBird全色与多光谱图像数据，进行三种有代表性的基于分量替换的融合方法（PCA，GS1，GS2）的对比分析。通过对比融合前后典型地物的光谱特性变化来评价融合影像的光谱保真性也是该研究的一大特色。结果表明:3种融合方法都具有很高的光谱保真性，GS2融合方法具有最优的光谱保真性，PCA和GS1融合算法次之，尤其是GS1融合处理后的图像存在部分光谱失真的现象，GS2算法的光谱保真性明显优于GS1。从基于分量替换融合方法的实质可以诠释出造成GS1融合图像光谱失真的根本原因，GS1融合算法使用的只是实验室理想环境下所获取的名义上的光谱响应特征。传感器的实际成像过程受到诸多因素的影响，如传感器在轨工作环境、大气、观测角度不同等的影响。GS1算法单纯通过不同波段光谱响应函数的线性组合来模拟低分辨率全色波段并不十分准确，GS2直接利用MS和Pan波段像元灰度值进行线性回归，克服了上述不确定性问题。通过以上对比研究发现，如何利用多光谱数据准确地模拟低分辨率全色波段，直接影响到融合后影像的光谱保真性，是目前高分辨率遥感图像融合的关键技术。

Component Substitution Pan Sharpening of High Resolution Remote Sensing Imagery

Image fusion has great significance for image classification, feature extraction and feature identification. An effective image fusion technique can not only improve spatial details but also preserve the spectral information. A variety of image fusion techniques are devoted to merge multispectral (MS) and panchromatic (Pan) images. Among these techniques, component substitution (CS) methods are attractive because they are fast and easy to implement. But up to present in China, no study has been reported with respect to the in-depth analysis of the CS fusion principle. In this paper, we use linear algebra knowledge to analyze the principle of CS image fusion technique. Three CS based image fusion techniques are been compared, such as principal component Analysis (PCA), Gram-Schmidt transform pansharpening which simulates low resolution MS by spectral response function (GS1) and Gram-Schmidt transform pansharpening which simulates low resolution MS through multivariate regression of MS+Pan data (GS2). QuickBird image has been processed on the above three fusion algorithms. Experimental results show that although all these three algorithms have good spectral fidelity property, the GS2 algorithm is generally more efficient than PCA and GS1 algorithm. The spectral distortion is especially obvious in GS1 fused image. The red and NIR reflectance of vegetable surface features of GS1 fused image are obviously higher than that of original MS bands. Meanwhile, the green and NIR band spectral reflectance of high reflectance area (build-up land) of GS1 fused image are obviously lower than that of original MS bands. The reason to explain this phenomenon is that the GS1 method only considers the nominal spectral responses. Actually, the influence of other phenomena, such as on-orbit working conditions, variability of the observed scene, postprocessing effects, in particular, atmospheric influence can significantly modify the nominal spectral response. The GS2 method avoids this drawback by performing a linear regression between Pan and MS bands. The comparative analysis of these three methods shows that how to accurately simulate low resolution panchromatic band directly affects the spectral fidelity and how to construct a low-resolution panchromatic is the key technology in the current high-resolution remote sensing image fusion.