Variational Quality Control of Non-Gaussian Innovations in the GRAPES m3DVAR System: Mass Field Evaluation of Assimilation Experiments

The existence of outliers can seriously influence the analysis of variational data assimilation. Quality control allows us to effectively eliminate or absorb these outliers to produce better analysis fields. In particular, variational quality control (VarQC) can process gray zone outliers and is thus broadly used in variational data assimilation systems. In this study, governing equations are derived for two VarQC algorithms that utilize different contaminated Gaussian distributions (CGDs): Gaussian plus flat distribution and Huber norm distribution. As such, these VarQC algorithms can handle outliers that have non-Gaussian innovations. Then, these VarQC algorithms are implemented in the Global/Regional Assimilation and PrEdiction System (GRAPES) model-level three-dimensional variational data assimilation (m3DVAR) system. Tests using artificial observations indicate that the VarQC method using the Huber distribution has stronger robustness for including outliers to improve posterior analysis than the VarQC method using the Gaussian plus flat distribution. Furthermore, real observation experiments show that the distribution of observation analysis weights conform well with theory, indicating that the application of VarQC is effective in the GRAPES m3DVAR system. Subsequent case study and long-period data assimilation experiments show that the spatial distribution and amplitude of the observation analysis weights are related to the analysis increments of the mass field (geopotential height and temperature). Compared to the control experiment, VarQC experiments have noticeably better posterior mass fields. Finally, the VarQC method using the Huber distribution is superior to the VarQC method using the Gaussian plus flat distribution, especially at the middle and lower levels.