Module A - Summary of the first phase of MiKlip

In MiKlip Module A, the aim was to explore the benefits of using different initialization methods and identify the associated mechanisms leading to improved predictive skill. Several ensemble generation methods that aim to represent the uncertainty in ocean initial conditions were tested. A further aim was to investigate the benefits of initializing soil moisture as a potential source of predictability at interannual-to-decadal time scale. Attempts were made to improve data assimilation techniques and initial conditions for the ocean and ice components.

The main results that have been achieved in the first MiKlip phase and remaining challenges are highlighted below.

Objective A1: Improving initial conditions

A-Coordination: A new ocean synthesis GECCO2 covering the period 1948-2014 was produced (Köhl, 2014; and is now provided on a routine basis for new MiKlip initializations. GECCO2 extends the previous GECCO synthesis into the most recent years, while at the same time providing higher horizontal and vertical resolution, and improved physics. GECCO2 now also covers the global ocean domain and performed first sea ice assimilation attempts, which need to be intensified in the future. The GECCO2 syntheses is examined within the Ocean Reanalysis Intercomparison Project; e.g., Storto et al. (2015) analysed the performance of different syntheses with respect to the global and regional steric sea level changes, and Toyoda et al. (2015) investigated the representation of the ocean mixed layer depth in different reanalyses.

First coupled data assimilation experiments were conducted with CESAM ( as identical twin experiments that assimilate pseudo-data (which were generated with the model itself from the default values of the control vector). The identical twin experiments showed the accurate recovery of the default parameter values through assimilation of the data, with a strongly (by more than five orders of magnitude) reduced gradient of the cost function. These results prove the usefulness of the variational assimilation system around the ESM and demonstrate the feasibility of assimilating climate data to generate initial conditions simultaneously to improving model parameters.

AODA-PENG implemented a singular evolutive ensemble Kalman filter (SEIK) from the parallel data assimilation framework PDAF (Nerger and Hiller, 2013) into the ocean component of the coupled MPI-ESM. A global setup of SEIK with a small ensemble size and an assimilation interval of 1 month was tested (Brune et al. 2015, submitted). In a combined nudged atmosphere and SEIK ocean assimilation, the variability of surface temperature was found to be similar to the observations. Regionally weaker correlations with observations compared to the atmosphere and ocean nudged Baseline 1/Prototype assimilations were found. This can be expected because the nudging assimilation is explicitly built to reproduce the observations by the model, while the SEIK leaves more degrees of freedom to the model, reproducing the observation within a given uncertainty range. However, the hindcasts initialized from the combined nudged atmosphere and SEIK ocean assimilation show regionally higher skills for surface temperature compared to the Baseline 1/Prototype hindcasts.

Objective A2: Improving initialization and ensemble generation procedures

MODINI proposed an initialization method in which observed momentum flux anomalies are added to the coupled model’s seasonally varying climatology (to minimize coupling shock at the start of the forecast). First hindcasts have been produced that showed the improved skill in the tropics. They have shown that the initialization runs have skill at reproducing ENSO events, the Pacific Decadal Oscillation, and part of the variability of the East Asian Summer Monsoon (Ding et al., 2014a) and the Southern Annular Mode (Ding et al., 2014b). The results reported above were carried out using the Kiel Climate Model (KCM; Park et al., 2009), also MODINI has been implemented in the MPI-ESM for the set of initialized hindcasts (Thoma et al., 2015). The importance of the tropics for driving the climate system is well known (see, for example, Greatbatch et al., 2012, for a detailed assessment). Nevertheless, initialization in the tropics is problematic, as discussed by Bell et al. (2004). Away from the equatorial zone, the dominant balance in the ocean is the geostrophic balance. But in the equatorial zone the geostrophic balance breaks down. Indeed, the dominant balance in the zonal momentum equation along the equator is between the zonal pressure gradient and the zonal wind stress and, furthermore, this is not necessarily a steady state balance but rather a dynamically evolving one. This means that when the ocean hydrography is initialized in the equatorial zone, there is in general no guarantee that the resulting pressure gradient will be balanced correctly by the zonal wind stress in the initialization system. These imbalances can lead to forecast errors (for example, leading to a La Niña in the subsequent forecast rather than an El Niño). MODINI has the advantage that it corrects the balance between the zonal wind stress and the zonal pressure gradient in the ocean.



Fig 1. Maps of anomaly correlation and mean squared error skill score of ensemble mean historical forecast skills of SAT calculated against HadCRUT4 median, averaged over the 2–5 and 6–9 year periods. Crosses denote values significantly different from zero exceeding at a 5% level applying 500 bootstraps. Gray shaded areas mark missing values with less than 90% data consistency in the observations (Figure adapted from Thoma et al., 2015, Fig 2).


PastLand: The scope of PastLand was focused on model improvements, sensitivity studies, evaluation of observational datasets, as well as the state and parameter optimization of the land surface model. Important achievements were: (1) A 5-layer soil hydrology scheme was developed for JSBACH by Hagemann and Stacke (2014), which was a crucial prerequisite for the parameter optimization based on top layer soil moisture observations. The new scheme represents soil moisture memory much better compared to the former bucket scheme due to a deep soil layer below the root zone, which serves as a long term buffer during dry periods. This scheme is now included in the MiKlip prediction system. (2) Sensitivity studies for soil moisture initialization reveal memory between a few days up to several seasons for soil moisture (Fig. 2). This demonstrates that soil moisture assimilation has the potential to improve the prediction skill of the ESM. The analysis also showed complex interactions between memory and climate states that differ regionally and temporally. This leads to the assumption that the predictive skill might not only be influenced by the spatial distribution of soil moisture initial states, but also by the particular time of initialization (Stacke and Hagemann, 2015). (3) The analysis of satellite-based soil moisture observations (ESA-ECVSM) with respect to their potential and limitations to serve as validation or assimilation data for land surface models (Löw et al., 2013). The evaluation methods applied in this study can be adapted for the evaluation of other land surface observations in PastLand2. (4) The parameter and state optimization using the optimization tool TAF (Kaminski et al., 2012) in JSBACH. This method was shown to be suitable for optimization using data products based on the whole root zone but is limited in its application to top layer soil moisture.



Fig 2. Seasonal variation and distribution of root zone soil moisture memory (Figure from Stacke and Hagemann, 2015).


A-Coordination: Using UCLA/MITgcm three initialization methods for decadal predictions were tested, namely full state initialization, anomaly initialization, and full state initialization including heat and fresh water flux correction. It was shown that full-state initialization employing flux correction leads to better forecast skill for sea surface temperature (SST) as a result of improved mean state and smaller biases in predictions (Polkova et al., 2014). Forecast skill in the first lead years (up to yr2-5) in the regions with a deep mixed layer was attributed to the SST re-emergence mechanism (initialized winter SST anomaly persist below seasonal thermocline without being damped back to the atmosphere in summer-time and become re-entrained into the deep mixed layer and reappear at the ocean surface the following winter, e.g. Deser et al, 2003). Therefore, more realistic representation of the mixed layer depth in full-field initialized and flux-corrected hindcasts resulted in better forecast skill. Also for the Atlantic meridional overturning circulation (AMOC) full state initialization and flux correction showed better results (forecast skill up to yr6-9) than anomaly initialization (up to yr2-5). The flux correction in the current setup leads to the best results, followed by the full state initialization and anomaly initialization.

A new ensemble generation procedure based on a method used in seasonal prediction (Molteni et al. 1996) was developed based on empirical oceanic singular vectors (SVs). In terms of forecast skill for surface air temperature, hindcasts based on the singular vector approach show improvement over atmospheric lagged initialization over the North Atlantic Ocean up to lead year five. For other variables, the analysis of the results showed either no significant improvement over the lagged initialization, or improvements were compensated by additional problems.

AODA-PENG tested two ensemble generation methods, namely the anomaly transform (AT; Romanova and Hense (2015) and multi-assimilation (MA; Kröger and von Storch, 2015, in prep.) These methods including the SV scheme from A-Coordination were evaluated in comparison to the lagged initialization commonly used in the decadal prediction community by performing experimental ensembles of ten-year-long hindcasts. The hindcast skills for all newly implemented methods currently show all similar performances compared to the lagged initialized hindcasts from Baseline 0/1 (Fig. 1). This can be seen from Fig. 3 which presents the predictive skills measured by the correlation coefficient of the ensemble mean versus the observed sea surface temperatures SST from the HadISST data set at lead time yr1 (left) and yr2-5 (right). and A-Coordination: The systematic comparison of the ensemble generation methods appeared to be non-trivial because of different scopes of the all three methods. While SVs are designed to capture the fastest linear growing errors in the ocean, the AT is designed to capture typical balanced anomaly structures in space and time. On the other hand, the MA studies the mixture of internal model variability with the effects of assimilating the data only in areas where observations are available. Nevertheless, we used the standard metrics of ensemble spread evaluation (beta score and ensemble spread score) to evaluate the differences between the methods. The beta score summarizes information from the analysis rank histogram (ARH), which compares the spread of the ensemble with the variability of the observations. A beta score value of zero indicates a perfect or flat ARH, meaning that the spread of the ensemble covers the variability of the observations well. Such an ensemble is called “calibrated”. Values less (larger) than zero indicate an underdispersive (overdispersive) ensemble, meaning that the ensemble spread is too small (large) compared to the variability of the observations.

The typical behavior of the lagged initialized predictions is the underestimation of a spread at initial time followed by a growth in spread during the first 2 – 4 lead years. In this respect, neither AT nor MA showed much improvement in the spread for the beta score and ensemble spread score for ocean heat content and SST, while SV seemed to overestimate spread at initial time (Fig.4). The overestimation of spread by SVs could depend on a too large scaling of the initial perturbations. However, in Module A, the SV was scaled to represent the error in the EN3 observations. Hence, evaluating the spread with respect to the EN3 data showed that the SV hindcasts in fact also underestimate the spread at initial time but not as much as in the atmospheric lagged initialization.

On the other hand, the performance of different ensemble generation methods seems to be sensitive to verification metrics and verification dataset. For instance, comparing the spread with respect to HadISST and Reynolds SST pointed out some differences in the spread score patterns, namely the spread of the SV-hindcasts evaluated against HadISST is much larger than that evaluated against Reynolds SST. Further, the ratio of the interannual variability explained by HadISST and Reynolds SST with respect to the AMSRE data showed that the HadISST seem to underestimate the variability in the extratropical Southern Hemisphere and along the western boundary currents (Marini et al., 2015, in prep).



Fig. 3 SST predictive skill measured in terms of the correlation coefficient of the ensemble mean vs the HadISST observations at lead time yr1 (left) and yr2-5 (right).
Fig. 4 Beta scores for the OHC700m (left) and the SST (right) in different regions. Every line corresponds to a spatial area and a type of ensembles. The blue color always represents atmospheric lagged initialization, and the red corresponds to AT (top), SVs (middle), and the multi assimilation runs approach (bottom). Verification values come from the NODC dataset for OHC700m and HadISST for the SST.


AODA-PENG also evaluated the importance of direct and indirect effects of tropospheric aerosol upon the potential predictability of interannual to decadal atmospheric variability. To this end four rather unique ensemble simulations using ECHAM6 were carried out:
(1) a no-aerosol AMIP-style 10 member ensemble for the period 1995 – 2004 using monthly mean values of sea surface temperatures (SST) and sea ice concentrations (SIC) from the HadISST data set as forcing boundary conditions,
(2) a 10 member AMIP-style ensemble (over 1995-2004) based on the HadISST with spatial and temporally varying aerosol radiative properties according to Kinne et al. (2006), implementing the direct radiative effects of tropospheric aerosols without feedbacks to the model atmosphere,
(3) a 10 member AMIP-style ensemble (over 1995-2004) with the HAM module added to ECHAM6 simulating the direct aerosol effects based on emission rates and with interaction to the simulated atmospheric general circulation, and finally
(4) a 10 member AMIP-style ensemble (over 1995-2004) with the HAM module added to ECHAM6 simulating both the direct radiative aerosol and indirect cloud-aerosol effects based on emission rates and interaction with the simulated atmospheric general circulation.

This type of experimental setup allows by analysis of variance the separation of the unpredictable internal atmospheric variability from the potentially predictable ones: SST induced variability with the potential of decadal predictability and the importance of aerosol induced variability with and without interaction and the direct vs. indirect effects. In Fig.5 the relative contribution of the aerosol effects to the total variability as the sum of internal, SST and aerosol-induced are shown. Fig 5 (left) present the results for the direct/indirect ECHAM6/HAM vs no-aerosol ensemble and Fig.5 (right) for the direct only vs no-aerosol ensemble for the annual mean, vertically averaged atmospheric radiation budget. The comparison shows that the direct-only ensemble exhibits a higher sensitivity upon the atmospheric radiative energy budget because the overall relative contributions are clearly larger than for the indirect/direct ensemble.



Fig. 5 (Left) Results of the variance analysis for annual means of the atmospheric radiative energy budget comparing the direct/indirect ECHAM6/HAM ensemble versus the no-aerosol pure HadISST ensemble and (right) for the direct only ensemble versus the no-aerosol pure HadISST ensemble. Both maps show the estimated relative contribution of the aerosol variance effects to the total variance.



Bell, M., M.J. Martin, and N.K. Nichols, 2004: Assimilation of data into an ocean model with systematic errors near the equator, Q. J. R. Meteorol. Soc., 130, 873–893 doi:10.1256/qj.02.109.

Brune, S., Nerger, L. and Baehr, J., 2015: Assimilation of oceanic observations in a global coupled Earth system model with the SEIK filter, submitted to Ocean Modelling.
Deser C., Alexander M.A., and Timlin M.S., 2003: Understanding the Persistence of Sea Surface Temperature Anomalies in Midlatitudes. J. Climate, 16, 57–72.

Ding, H., R.J. Greatbatch, W. Park, M. Latif, V. Semenov, V. and X. Sun, 2014a: The variability of the East Asian Summer Monsoon and its relationship to ENSO in a partially coupled climate model., Clim. Dyn., 42, 367-379, doi:10.1007/s00382-012-1642-3.

Ding, H., R.J. Greatbatch and G. Gollan, 2014b: Tropical impact on the interannual variability and long-term trend of the Southern Annular Mode during austral summer from 1960/1961 to 2001/2002., Clim. Dyn. doi:10.1007/s00382-014-2299-x.

Greatbatch, R. J., G. Gollan, T. Jung, and T. Kunz, 2012: Factors influencing Northern Hemisphere winter mean atmospheric circulation anomalies during the period 1960/61 to 2001/02., Q.J.R. Meteorol. Soc., 138, doi:10.1002/qj.1947.

Hagemann, S., and T. Stacke, 2014: Impact of the soil hydrology scheme on simulated soil moisture memory., Clim. Dyn., 1-20, doi:10.1007/s00382-014-2221-6.

Kaminski, T., Knorr, W., Schürmann, G., Scholze, M., Rayner, P. J., Zaehle, S., Blessing, S., Dorigo, W., Gayler, V., Giering, R., Gobron, N., Grant, J. P., Heimann, M., Hooker-Strout, A., Houweling, S., Kato, T., Kattge, J., Kelley, D., Kemp, S., Koffi, E. N., Köstler, C., Mathieu, P., Pinty, B., Reick, C. H., Rödenbeck, C., Schnur, R., Scipal, K., Sebald, C., Stacke, T., van Scheltinga, A. T., Vossbeck, M., Widmann, H. and T. Ziehn, 2013: The BETHY/JSBACH Carbon Cycle Data Assimilation System: experiences and challenges., J. Geophys. Res., 118, doi:10.1002/jgrg.20118

Köhl A., 2014: Evaluation of the GECCO2 Ocean Synthesis: Transports of Volume, Heat and Freshwater in the Atlantic. Quart. J. Roy. Meteor. Soc., doi:10.1002/qj.2347.
Kröger J. and von Storch Jin-Song, 2015: Spurious initialization in near-term AMOC predictions, in prep.

Löw, A., Stacke, T., Dorigo, W., de Jeu, R. and S. Hagemann, 2013: Potential and limitations of multidecadal satellite soil moisture observations for selected climate model evaluation studies., Hydrol. Earth Syst. Sci., 17, 3523-3542, doi:10.5194/hess-17-3523-2013

Marini C., Polkova I., A. Köhl, and D. Stammer, 2015: Comparison of ensemble generation methods using oceanic singular vectors and lagged initialization for decadal climate prediction, in prep.

Molteni F., R. Buizza, T. N. Palmer, and T. Petroliagis, 1996: The ECMWF ensemble prediction system: Methodology and validation. Quart. J. Roy. Met. Soc., 122, 73-119.

Nerger L., and Hiller W., 2013: Software for ensemble-based data assimilation systems - Implementation strategies and scalability, Comp.&Geos. 55, 110–118, doi:10.1016/j.cageo.2012.03.026.

Polkova I., A. Köhl, and D. Stammer, 2014: Impact of initialization procedures on the predictive skill of a coupled ocean–atmosphere model. Clim. Dyn., 42, 3151-3169 doi:10.1007/s00382-013-1969.

Romanova, V. and A. Hense, 2015: Anomaly transform methods based on total energy and ocean heat content norms for generating ocean dynamic disturbances for ensemble climate forecasts, Clim. Dyn., 1-21.

Stacke, T. and S. Hagemann, 2015: Life time of soil moisture perturbations in a coupled land-atmosphere simulation, Earth Syst. Dynam. Discuss. 6(2): 1743-1788.

Storto, Andrea, et al., 2015: Steric sea level variability (1993–2010) in an ensemble of ocean reanalyses and objective analyses, Clim. Dyn., 1-21.

Thoma, Malte, et al., 2015: Decadal hindcasts initialized using observed surface wind stress: Evaluation and prediction out to 2024, Geophysical Research Letters 42.15, 6454-6461.

Toyoda, Takahiro, et al., 2015: Interannual-decadal variability of wintertime mixed layer depths in the North Pacific detected by an ensemble of ocean syntheses, Clim. Dyn., 1-17.

More on Module ANews-Icon