Retrieval of Aerosol Properties from TROPOMI Measurements

Abstract

Aerosols affect Earth's radiation budget by scattering and absorbing solar radiation (direct effect) and by influencing the cloud formation processes (indirect effect). Accurate assessments of aerosol properties, such as optical depth and layer height, are important for the global monitoring of air pollution in the lower atmosphere. A number of passive satellite sensors enable to monitor aerosol properties on both regional and global scale using spectral information at various wavelengths. For example, measurements in the oxygen A-band from the Scanning Imaging Absorption Spectrometer for Atmospheric Chartography (SCIAMACHY), the Global Ozone Mapping Experiment (GOME) and GOME-2, the Greenhouse Gases Observing Satellite (GOSAT), and the TROPOspheric Monitoring Instrument (TROPOMI) are used to retrieve both aerosol optical depth and layer height. The ultimate generation of passive satellite sensors, as for example, the TROPOMI/S5P has an extraordinary spatial, temporal and spectral resolutions. The challenge of processing hyperspectral data is to increase the performance of the retrieval algorithms in order to achieve near-real-time requirements. The goal of this thesis is the design of algorithms for retrieving aerosol parameters from TROPOMI/S5P measurements in the oxygen A-band. The designed algorithms can be grouped into two categories. The first category includes Bayesian-based retrieval algorithms for a specified aerosol model and a set of candidate models. In the latter case, two solutions estimates, namely (i) the maximum solution estimate, corresponding to the model with the highest evidence, and (ii) the mean solution estimate, representing a linear combination of solutions weighted by their evidences, are proposed. The algorithms use a linearized radiative transfer model relying on the discrete ordinate method with matrix exponential, and as acceleration approaches, the telescoping technique, the method of false discrete ordinate, the correlated k-distribution method, and the principal component analysis. The inverse problem is formulated as a least-squares problem and solved by means of the iteratively regularized Gauss-Newton method. The second category includes neural network retrieval algorithms. These are trained (i) to emulate the radiative transfer model, which is then used in conjunction with a Bayesian approach to solve the inverse problem, and (ii) to learn the inverse model using as input either the synthetic radiances or their principal components. The retrieval performances of the retrieval algorithms are analyzed on synthetic and real data.