Model informed assessment of observation impact for emission rate optimization using four-dimensional variational data assimilation

Air pollution poses a significant threat to human health and the environment. Although efforts are being made to develop emission inventories, and characterise and quantify emission rates, the absence of direct measurements introduces uncertainty into model-based estimations, leading to uncertain air pollution predictions. Chemistry transport models, such as the EURopean Air pollution Dispersion-Inverse Model (EURAD-IM), can combine physical, chemical, and meteorological information to simulate the state of the atmosphere. A common methodology for incorporating observational data into chemistry transport models for further analysis is data assimilation. The EURAD-IM uses four-dimensional variational data assimilation, which allows for the correction of emission inventories. However, the accuracy of these corrections depend on the information provided by the observations. It is therefore a fundamental requirement of the observational data to be representative of the state of theatmosphere. In order to improve the representativity of the model analysis, a methodology to generate a representative split of ground-based observation data is developed in this work. A clustering algorithm is used to determine observation stations with common statistical features e.g. mean and variance of the annually averaged diurnal cycle. By sampling each cluster individually, the resulting assimilation and validation data sets are representative of the available measurement data. The representativity is evaluated as the difference in the root mean square error of the assimilation data to the validation data after the model analysis. It is shown that the representativity of the new data sets is improved compared to anoperational configuration. To assess the influence of observation configurations on the model result, a series of observing system simulation experiments is performed. Hereby, artificial observations are extracted from a model simulation, the nature run, which serves as ground truth to gauge the accuracy of subsequent model simulations. The nature run is generated with a known emission perturbation for CO (carbon monoxide), NOx (nitric oxide and nitrogen dioxide), NMVOC (non-methane volatile organic compounds), and SOx (sulfur oxide) from EURADIM simulations. This allows to examine the influence of the observation network on the emission rate correction directly. A systematic benchmark study addressing the relation between spatial coverage of the observational network and the quality of the model analysis is performed. The ability of the model to reproduce the emission rate perturbation of the nature run is evaluated for different observation networks with varying coverage. It is shown that the best performance is achieved by an observation configuration covering approximately 11 % of the model grid. Increasing the observation density further reduces the accuracy of the emission rate corrections. In addition, a surrogate model for the evaluation of the impact of individual observations is developed, focusing on the definition of an optimal observation network for emission rate optimization. It combines an adjoint sensitivity field of the simulated day with the emissioninventory data and the root-mean square error (RMSE) of the model concentrations of a reference simulation (without data assimilation) with respect to the observation. An algorithm for the generation of an observation configuration based on the observation impact is developed and the resulting observation configuration is tested in the framework established by the benchmark study. The generated observation network outperforms the best benchmarking configurations in terms of RMSE reduction and convergency rate with a lower number of observations. The algorithm is extended to supplement satellite-based column density observations to ground observation networks. It is shown that additional satellite observations to an optimal ground-based observation configuration are detrim…