Master thesis by Tomas Fernandez Bouvier

Title: Automatic detection of outliers in meteorological observation data

Abstract: In this thesis we study and develop a suite of tests aimed for probing meteorological observations
before they are assimilated into a numerical weather prediction model. Moreover we design and implement a framework for benchmarking the latter tests and tune their ”hyper-parameters” in order to maximise the hit rate while keeping the false alarm rate low. We provide a bayesian method to account for the multiple evidences arising from the tests. Finally, we apply the methods developed in this work to the observation datasets used in the DANish regional ReAnalayses (DANRA) and the Copernicus Artic Regional Re-Analyses (CARRA), both project making use of historical data for which the quality of observations may be so sub-optimal that it requires an extensive quality control. The automatic correction output by our method tends to be consistent with an expert researcher manual-detection for DANRA while for CARRA it appears difficult to match the expert skill.

Supervisors: Professor Eigil Kaas, NBI/NCKF, senior researchers at DMI, Xiaohua Yang og Bjarne Amstrup.

Censor: Peter Aakjær.