Generating atmospheric forcing perturbations for an ocean data assimilation ensemble

Running ensemble of reanalyses or forecasts has proved successful at improving their performances, despite the cost. Generating ensemble simulations requires generating perturbations within the models, and for the assimilated observations and subsidiary conditions. This paper proposes a statistical model to generate atmospheric forcing random perturbations in a flexible and cheap way, for all the variables required to calculate bulk formulae. The training data are a ten-year set of differences between ERA-INTERIM (from ECMWF) and MERRA (from NASA) atmospheric forcing fields, unbiased with a high-pass filter. The model is designed to generate global spatially-varying multi-variate and unbiased perturbations with consistent spatial structures. Based on linear regressions, the model allows for regression coefficients and residual standard deviations to vary with the time of the year. Once defined, the model does not rely on any other external data to generate the perturbations, and can hence be used on- or off-line in the goal of seeding the ensemble more appropriately. Once designed, the model has been validated by comparing three years of generated perturbations to three years of differences between the reanalyses out of the training period. Statistical tests show that the distributions of the sets comply reasonably well, except for precipitation and snow. The major issues come from regions with high kurtosis or where no clear patterns can be established. In terms of structure, the perturbation standard deviation of both sets show similar patterns for all variables.