基于相关系数匹配的混合像元分解算法

A New Spectral Mixture Analysis Method Based on Spectral Correlation Matching

遥感影像中的混合像元不仅影响地物识别和分类的精度,而且已成为遥感科学向定量化发展的主要障碍。为此,人们已经提出了多种混合像元分解算法,其中应用最为广泛的是最小二乘法。该方法虽然具有意义明确、简单易行等特点,但也易受局部噪声、大气效应、环境辐射等因素的影响。本文将混合像元分解问题归结为一个基于光谱匹配的非线性最优化问题,并针对最小二乘法的不足提出了一种新的基于相关系数匹配(spectral corre- lation matching,SCM)的混合像元分解技术。通过在北京市北三环及其以北区域内案例的研究表明:在城市区域内,利用图像选取终端端元的办法,基于相关系数匹配的混合像元分解算法的总体精度高于带全约束的最小二乘法(LS)的分解结果。对比分析表明:在目标光谱的绝对值整体放大或缩小而光谱形状得到了很好的保持、及局部噪声使得光谱值显著变化但光谱形状得到了一定程度保持时,基于光谱形状的相关系数法可以得到比基于光谱绝对值的最小二乘法精度更高的分解结果。

Mixed pixels widely exist in remotely sensed images.It not only influences the accuracy of target detection and classification,but also greatly hinders the development of quantitative remote sensing.A large number of spectral mix- ture analysis methods have been proposed,among which the Least Square(LS)method is a widely used technique in remote sensing to estimate fractions of materials(endmembers)existing in an image pixel.With its character of simple and effective in many application studies,it has some defects such as sensitivity to local noise,atmospheric effects and environmental radiation etc.Because the Root Mean Square Error(RMSE)is used as model fit in the LS method,when the magnitude of the spectrum changes significantly while the spectral shape is kept well which would be caused by atmos- phere,shadow and so on,the unmixing accuracy of LS method will be reduced remarkably.In this study,the spectral unmixing problem is considered as a nonlinear optimization question and a new spectral mixture analysis method based on the Spectral Correlation Matching(SCM)is proposed to overcome the defects of LS method.Different with LS method, the SCM method uses the correlation coefficient to describe the similarity between the objective and test spectrum.Based on the overall shape feature of the spectra instead of the absolute differences between the objective and test spectrum,the SCM method can reduce the influence of atmospheric effects,environmental radiation etc.In order to evaluate the per- formance of SCM method and compare it with LS method,a case study was carried out in the north-third-ring area in Bei- jing city using a Landsat ETM + and IKNOS image.The ETM + images was resampled to 28m,and the IKNOS image was first classified into four land cover types corresponding to the endmembers,then the real fraction of each endmember was calculated within a 7×7 window.The results indicated that the proposed SCM method was a better alternative to least square method,with higher accuracies for each endmember estimation than LS method.It suggests that the SCM method can be applicable to solve unmixing problem in remote sensing.