1. ECMWF CO 2 Data Assimilation at ECMWF Richard Engelen European Centre for Medium-Range Weather Forecasts Reading, United Kingdom Many thanks to Phil Watts, Frédéric Chevallier, Tony McNally, and Erik Andersson.

2. CO 2 Data Assimilation at ECMWF Minimize the following cost function: where : state vector increments : observation operator : observation departures Depending on the specification of their respective error covariance matrices B and R, the background term J b and the observation term J o will determine the new trajectory of the model. 4D-Var data assimilation

3. CO 2 Data Assimilation at ECMWF 4D-Var data assimilation

4. CO 2 Data Assimilation at ECMWF 4D-Var data assimilation High resolution trajectory run : First Guess Resolution: T L 511  40 km, 60 levels up to 0.1 hPa  Compare OBS - FG : Data quality control Minimisation using tangent linear + adjoint model Resolution: up to T L 159  125 km, 60 levels Successive minimisations at T L 95 and T L 159 with simplified physics. Update of the trajectory at high resolution in between. Observation time window : 12 hours  Adjust initial state iteratively to fit Observations Set up of current research mode assimilation (CY25R4):

5. CO 2 Data Assimilation at ECMWF CO 2 in the 4D-Var system CO 2 is currently treated as a so-called ‘column’ variable within the 4D-Var data assimilation system. This means that CO 2 is not a model variable and is therefore not moved around by the model transport. It is retrieved at the observation location using the 4D-Var fields of temperature, specific humidity and ozone. The CO 2 variable itself is limited to a column-averaged mixing ratio with a fixed profile shape. There are plans to include CO 2 as a tracer variable in the model, but this is not available yet.

6. CO 2 Data Assimilation at ECMWF AIRS Data Stream NASANESDIS ORA ECMWF ECFS 4D-Var system Black List Cloud Detection Assimilated Fields MARS ECFS CO 2 T, q, etc. MARS 2378 ch., all pixels324 ch., 1 out of 18 pixels when operational ~ 100 channels Bias correction

7. CO 2 Data Assimilation at ECMWF Channel flagging 324 input channels +/- 100 channels remaining Cloud Detection Blacklisting Quality Control

8. CO 2 Data Assimilation at ECMWF Cloud detection A new cloud detection algorithm has been developed by Tony McNally and Phil Watts (Revised version submitted to Q.J. R. Meteor. Soc.). Instead of detecting clear Field of Views, it detects clear channels. It makes use of the fact that channels with weighting functions peaking above the cloud are not affected by the cloud. To determine if a channel is affected by clouds, clear radiances calculated from the model are compared to the observations. After sorting the channels based on the peak of their weighting function, differences between simulated and observed radiances approach zero within the error margin of both model and observations. This is used to determine if a channel is clear or not. The method does not rely on the cloud estimation of the forecast model and maximises the amount of extracted information, which are significant advantages.

9. CO 2 Data Assimilation at ECMWF Cloud detection: example Surface600 mb100 mb

10. CO 2 Data Assimilation at ECMWF AIRS bias relative to ECMWF background

11. CO 2 Data Assimilation at ECMWF Bias correction Bias correction is necessary to properly assimilate AIRS observations in the 4D-Var system. The bias correction is done for each individual channel. When an air-mass dependent correction is used for AIRS, it is possible to remove part of the CO 2 signal and create air- mass dependent biases in the CO 2 analysis values. This effect is the strongest when atmospheric CO 2 is correlated with any of the predictors. At the moment a global mean bias correction is used for each individual AIRS channel. This bias correction is calculated with fixed CO 2 values in the radiative transfer.

12. CO 2 Data Assimilation at ECMWF Blacklist A so-called blacklist is used in the data assimilation system to remove all channels that are not fit to use. Examples are the blacklisting of the short-wave band of AIRS to avoid problems with solar radiation in the radiative transfer modelling, and the blacklisting of defunct channels. The blacklist can easily be adapted to include or exclude extra channels. Also, different data filters can be put in to use e.g. only data over ocean or only data with a scan angle less than a certain value.

13. CO 2 Data Assimilation at ECMWF Global assimilation with real AIRS data Assimilation experiments were done for the period 24 – 27 January 2003 and for 24 – 27 April 2003. 324 channels were available before the blacklisting process. Cloud detection algorithm is used to screen for cloud-free channels. Different CO 2 assimilation set-ups: Single column value with the same mixing ratio in the troposphere and the stratosphere using all available channels Single column value with the same mixing ratio in the troposphere and the stratosphere using 10 stratospheric channels Single column value with the same mixing ratio in the troposphere and the stratosphere using 10 tropospheric channels Double column experiment with all available channels

14. CO 2 Data Assimilation at ECMWF Assimilation using all channels First results with the sink variable analysis using all available channels show an unexpected spatial distribution. Siberia in wintertime should have increased CO 2 levels in the troposphere, not decreased. Note that the absolute values of the model results are not verified against flask observations. Results are presented here for spatial distribution only! Colour scales are also slightly different.

15. CO 2 Data Assimilation at ECMWF Assimilation using 10 stratospheric channels When only 10 stratospheric channels are used, the spatial distribution looks more like what we would expect. Highest concentrations in the tropics with decreasing values to higher latitudes. Note that the absolute values of the model results are not verified against flask observations. Results are presented here for spatial distribution only! Colour scales are also slightly different.

16. CO 2 Data Assimilation at ECMWF Assimilation using 10 tropospheric channels When only 10 tropospheric channels are used, we lose quite a bit of our coverage due to clouds. However, Siberia now indeed shows increased CO 2 levels in the troposphere, not decreased. Note that the absolute values of the model results are not verified against flask observations. Results are presented here for spatial distribution only! Colour scales are also slightly different.

17. CO 2 Data Assimilation at ECMWF Assimilation using 10 tropospheric channels Tropospheric CO 2 (left) and number of observations per grid box as used in the average (right). Northern hemisphere in winter has a lot of cloud cover in the lower and middle troposphere. Therefore, to get a decent time average, we need a relative long time series of data.

18. CO 2 Data Assimilation at ECMWF Double column analysis The introduction of a double sink provides a bit more flexibility in the estimation of CO 2. The tropopause is estimated from the background temperature profile and independent values for stratospheric and tropospheric CO 2 are then estimated. Errors (  a ) are estimated by calculating the CO 2 Jacobians within the analysis and using the 20 ppmv background error (  b ) and 1K observation errors (S y ) : The analysed values are then gridded, weighted by their individual analysis errors. This assumes that the errors are uncorrelated in time and space, which is a bit optimistic.

19. CO 2 Data Assimilation at ECMWF Double column analysis

20. CO 2 Data Assimilation at ECMWF Double column analysis The fully correlated mean errors on the left show the pessimistic error estimates. The number of channels used in the analysis (shown below) provides an indication of how deep in the troposphere we look.

21. CO 2 Data Assimilation at ECMWF Effect of background bias The effect of the background bias is shown by filtering out any CO 2 analysis values that have analysis errors too close to the background error.

22. CO 2 Data Assimilation at ECMWF Conclusion and Questions The ECMWF column CO 2 product is getting into shape. Still some work to do: use of best available background, bias correction, quality control, validation (?). What do we know about the correlation between the stratosphere and troposphere? This is important, because we have so many AIRS channels that are (partly) sensitive to the stratosphere. The specified background covariance matrix has therefore an important effect on how information from these channels is distributed through the vertical. Are we able to specify horizontal correlations for gridded surface fluxes? This will be important if we want to go one big step further and start CO 2 model inversions within the ECMWF 4D-var framework.