Ten years of MIPAS measurements with ESA Level 2 processor V6 – Part I: retrieval algorithm and diagnostics of the products

The MIPAS instrument on the ENVISAT satellite has provided vertical profiles of the atmospheric composition on a global scale for almost ten years. The MIPAS mission is divided in two phases, the full resolution phase, from 2002 to 2004, and the optimized resolution phase, from 2005 to 2012, which is characterized by a finer vertical and horizontal sampling attained through a reduction of the spectral resolution. While the description and characterization of the products of the ESA processor for the full resolution phase has been already described in previous papers, in this paper we focus on the performances of the latest version of the ESA processor, named ML2PP V6, which has been used for reprocessing the entire mission. The ESA processor had to perform the operational near real time analysis of the observations and its products needed to be available for data assimilation. Therefore, it has been designed for fast, continuous and automated analysis of observations made in quite different atmospheric conditions and for a minimum use of external constraints in order to avoid biases in the products. The dense vertical sampling of the measurements adopted in the second phase of the MIPAS mission resulted in sampling intervals finer than the instantaneous field of view of the instrument. Together with the choice of a retrieval grid aligned with the vertical sampling of the measurements, this made ill-conditioned the retrieval formalism of the MIPAS operational processor. This problem has been handled with minimal changes to the original retrieval approach but with significant improvements nonetheless. The Levenberg-Marquardt method, already present in the retrieval scheme for its capability to provide fast convergence for non-linear problems, is now also exploited for the reduction of the ill-conditioning of the inversion. An expression specifically designed for the regularizing Levenberg-Marquardt method has been implemented for the computation of the covariance matrices and averaging kernels of the retrieved products. The regularization of the Levenberg-Marquardt method is controlled by the convergence criteria and is deliberately kept weak. The resulting oscillations of the retrieved profile are a-posteriori damped by an innovative self-adapting Tikhonov regularization. The convergence criteria and the weakness of the self-adapting regularization ensure that minimum constraints are used and the best vertical resolution obtainable from the measurements is achieved in all atmospheric conditions. Random and systematic errors, as well as vertical and horiz