The goal of Module A during the second phase of MiKlip (MiKlip II) is (i) to improve estimates of initial conditions for the ocean, sea ice and the land soil moisture, (ii) based on those initial conditions to improve initialisation procedures and finally (iii) to optimise procedures to span the MiKlip forecast ensemble. The work package WP3 of Modul A is the sub-project  Atmospheric and Oceanic Data Assimilation & Ensembles Generation (AODA-PENG2). Its focus will be on the ensemble generation. Using two different approaches in MPI-ESM1.2, ensemble simulations will be performed and investigated to sample the sources of initial uncertainty through appropriate ensemble generation. Then the subsequent error growth during the predictions is used to quantify the development of prediction uncertainties.

As one approach, for dynamically oriented ensemble generation, the breeding method will be tested. The method differs from the classical breeding in the sense that specific time scales and spatial scales can be selected a priori to generate growing uncertainty modes. The breeding technique is a method used for constructing disturbances on the ocean prognostic variables. These disturbances are added to the initial model state of each forecast. The method extracts error modes which develop most strongly in space and time where the long term variability of the ocean physical processes is responsible for inter-annual and decadal changes.

A second approach will be based on the already implemented Ensemble Kalman Filter technique in the oceanic component of MPI-ESM 1.2. The current global implementation of the Ensemble Kalman Filter shows hindcast skill when the inherent data assimilation ensemble is used for forecast initialisation. However, recent studies have shown that the quality of the Ensemble Kalman data assimilation can be considerably improved in two ways: the use of the localized variant together with artificial covariance inflation and the increase of the ensemble size to more than 30 realisations. After upgrading our current implementation accordingly, we will further investigate the impact on hindcast skill and also analyse the ensemble spread and error growth.

Progress so far

WP 3.1

We implemented the Bred Vector (BV) method for ensemble generation (Romanova and Hense, 2015) on a longer time-scales, and such specifying the “errors of the month/year”. The method differs from the classical breeding in the specific time scales and spatial scales which can be selected a priori to generate growing uncertainty modes. It is based on modification of the velocity, temperature and salinity initial variables upon the error growth measured by the weighted total energy norm. The breeding routines were externally created in order to be easily implemented in other coupled climate model systems with respect further research or operational usage.

The produced BV hindcast begins in year 1961 and finishes in year 2016. Nine BV were calculated on twelve months looping period over five iterative steps and the forecast operated for ten years forward. To assess the BV hindcast skills and compare them to other initialisation methods, the CES evaluation system and VECAP standard evaluation tool were used. Additionally, another set of scores (Romanova et al. 2017) was applied, which includes Cumulative Frequency Analysis, Reliability Diagrams (RD), Quantile Score (QS) and Cumulative Ranked Probability Skill Score (CRPSS). The applied evaluation scores showed improvement of the BV predictive skills.

A supplementary study was done on the climatic states at each iteration step for La-Nina (year 2010) and El Nino (year 2015) years in order to outline the differences in the error growth in dependence of the background conditions (fig. 1). The study showed that the most sensitive regions in the ocean responsible for inter-annual to decadal variability are localised by fastest uncertainty growth rates and there is a fundamental difference in dependence of the background climate conditions. In this aspect and considering the results from the evaluation of the variables, we consider the BV ensemble generation method as more advanced way to produce climate predictions compared to the lagged or random initialisation.



Fig. 1 Ocean surface temperature growth at the end of the fourth iteration step in the BV procedure for the La Nina year 2010 (left) and El Nino year 2015 (right).


We increased the impact of observations on the simulated ocean state by advancing from the global variant of the ensemble Kalman filter (EnKF, Brune et al. 2015) from MiKlip I to the localised variant of EnKF for assimilating observed oceanic temperature and salinity profiles into MPI-ESM-LR. To stabilise the large scale oceanic circulation during assimilation we applied a two-phase spin-up prior to assimilation with (i) 300 years of atmospheric nudging to monthly 1958 values from the ERA40 reanalysis with no explicit restrictions to the oceanic component, and (ii) 50 years of atmospheric nudging as in (i) combined with oceanic localised EnKF assimilation of monthly temperature and salinity profiles from the EN4 dataset accumulated over the period 1950-1959. After the spin-up, assimilation with atmospheric nudging to ERA40/ERAInterim re-analyses and the oceanic localised EnKF runs over the period 1958-2016 with 16 ensemble members. We analysed the resulting 10 year hindcasts starting every November 1959-2016. Hindcast skill is generally improved when compared to hindcasts initialised by atmospheric nudging and oceanic global EnKF assimilation (see Brune et al. 2017). When compared to MiKlip pre-operational system hindcast skill for surface temperatures is improving in the Tropical Pacific and Northern Atlantic.

We currently implement the localised oceanic EnKF in MPI-ESM-HR. The assimilation set-up has to be adapted to the higher resolution. Assimilation quality of surface temperatures in MPI-ESM-HR/EnKF is good. Yet, special emphasis and has to be put on a stable oceanic meridional overturning circulation during assimilation. This leads to a significant increase in effort while testing assimilation set-ups.


Fig.2 Correlation with HadCRUT4 air surface temperature observation: a) local EnKF/MPI-ESM-LR hindcasts lead year 1, b) local EnKF/MPI-ESM-LR hindcasts lead years 2-5, c,d) differences in correlation between local EnKF/MPI-ESM-LR and Pre-Op.


  • Brune, S., L. Nerger, J. Baehr (2015): Assimilation of oceanic observations in a global coupled Earth system model with the SEIK filter. Ocean Modelling 96, Part 2:254–264, DOI 10.1016/j.ocemod.2015.09.011

  • Brune, S., A. Düsterhus, H. Pohlmann, W. Müller, and J. Baehr (2017), Time dependency of the prediction skill for the North Atlantic subpolar gyre in initialized decadal hindcasts, Climate Dynamics, in press, doi:10.1007/s00382-017-3991-4.

  • Romanova V., and Hense A. (2015) Anomaly Transform Methods Based on Total Energy and Ocean Heat Content Norms for Generating Ocean Dynamic Disturbances for Ensemble Climate Forecasts. ClimDyn. DOI: 10.1007/s00382-015-2567-4

  • Romanova, V., Hense, A., Wahl, S., Brune, S., Baehr, J. (2017): Skill assessment of different ensemble generation schemes for retrospective predictions of surface freshwater fluxes on inter and multi-annual timescales - Meteorologische Zeitschrift, doi:10.1127/metz/2017/0790


Institut für Meereskunde Universität Hamburg
Prof. Dr. Johanna Baehr
Dr. Sebastian Brune

Meteorologisches Institut Universität Bonn
Prof. Dr. Andreas Hense
Dr. Vanya Romanova

Time dependency of the prediction skill for the North Atlantic subpolar gyre in initialized decadal hindcasts

2018 - Climate Dynamics 51(5-6), 1947–1970

Brune, S. | Düsterhus, A., Pohlmann, H., Müller, W. A., Baehr, J.

Hindcast skill for the Atlantic meridional overturning circulation at 26.5°N within two MPI-ESM decadal climate prediction systems.

2016 - Clim. Dyn.

Müller, V. | H. Pohlmann, A. Düsterhus, D. Matei, J. Marotzke, W. A. Müller, M. Zeller and J. Baehr

Assimilation of oceanic observations in a global coupled Earth system model with the SEIK filter

2015 - Ocean Modelling, Vol. 96(2), pp. 254–264

Brune, S. | L. Nerger and J. Baehr