Comparison of statistical inversion techniques for atmospheric sounding

Statistical inversion techniques are a critical component of the analysis of spectroradiometric data for the inference of atmospheric vertical profiles (e.g. temperature and water vapor). Statistical inversions provide final products or initial estimates for physical inversion algorithms. Both situations require the implementation of accurate methods. Due to the ill-conditioning of the inversion problem, standard statistical methods may not be adequate. This paper presents the application and comparison of several multivariate regression models for atmospheric sounding. Methods reviewed include principal components regression (PCR), canonical correlation regression (CCR), maximum redundancy (MR), and partial least squares (PLS). The inversions are compared in terms of temperature and water vapor retrieval accuracy and are shown to be equivalent to maximum likelihood physics-based inversion models