Assessment of a Variable Projection Algorithm for Trace Gas Retrieval in the Short-Wave Infrared

Abstract

An important part of atmospheric remote sensing is the monitoring of its composition, which can be retrieved from radiance measurements, e.g. in the short-wave infrared (SWIR). For deriving trace gas concentrations in the SWIR spectral region a radiative transfer model is fitted to observations by least squares optimization. The aim of this thesis is to present the well-established variable projection method for solving separable nonlinear least squares problems and to examine and configure it for trace gas retrieval. For this, a Python implementation of the algorithm, called varpro.py, will be outlined and later utilized in retrievals with real satellite observations. These are meant to assess the efficiency, accuracy and robustness of three iterative algorithms for nonlinear least squares problems which have been built into varpro.py. Furthermore, a new feature - applying bounds to the non-linear fit parameters - will be included in the implementation and evaluated for its quality and usefulness. As a result of these tests, a new 'default' configuration will be suggested based on the algorithm with the best performance for trace gas retrieval. Also, ideas for analysing and testing strategies which could lead to even more insights will be proposed. Finally, possible future applications for trace gas retrieval will be motivated and suggestions for further research and modifications of varpro.py will be made.