Preview

PDF, English - main document Download (91MB) | Terms of use

Abstract

Nitrogen dioxide (NO2) is an important air pollutant, monitored globally by satellite instruments from space, in situ measurements at the surface, aircraft or balloons, and ground-based spectroscopic instruments. Satellite instruments, such as TROPOMI, provide NO2 column densities (i.e. vertical concentration integrals) by spectral analysis of backscattered sunlight. However, the concentration profiles themselves cannot be obtained. This marks a considerable deficit, as they could be pivotal for studies on health effects of air pollution, air chemistry, and improved satellite retrievals. This thesis presents the model “NitroNet”, a new NO2 profile retrieval based on an artificial neural network that predicts NO2 profiles from TROPOMI NO2 column densities and other ancillary variables. The network’s training data is obtained from the regional chemistry and transport model WRF-Chem, operated on a central European domain for the month of May 2019. The contents of this thesis include the validation and optimization of the WRF-Chem model, followed by the training, hyperparameter optimization, and validation of the NitroNet neural network, where “validation” refers to the comparison of either model to satellite observations and measurements from in situ and ground-based spectroscopic instruments. Furthermore, the applicability of NitroNet on other spatio-temporal domains is successfully demonstrated.