the reanalysis. However, since the heat fluxes and the ocean model are both imperfect, it is necessary to impose a negative feedback to the surface temperature during the spin-up phase. A high value of 400 wm2/ K is used. This can be thought of as assimilating SST data in some sense. However, it can still be the case that the heat flux and or wind stress or model error can be sufficiently large that the model SST deviates from the observed by -0.3 K. Surface salinity is relaxed to climatological values, to prevent the ocean from drifting to far from reasonable salinity values since the ocean model has no other effective method of adjusting to errors in the fresh water flux or errors in the model formulation.
An assimilation system has been incorporated into the system to assimilate data from the GTSPP project. The coverage is in general poor with large areas of the world being
unsampled in a one month period. In contrast there are areas where the ocean is very well sampled, notably within +1- 8 degrees of the equator in the Pacific. In this area there is almost complete cover every day. The version of the assimilation system originally implemented (denoted 01-1) did not recognise this large difference in data coverage. Instead data were separated into ten-day windows. An 01 analysis was then performed using the model as first guess and merged with the data. The code for the 01 implementation, together with the quality control software was kindly provided by Smith (See Smith et al 1991 for a fuller description). The data quality control is quite involved. A datum is first checked against climatology, then against an analysis made without that datum, called buddy checking. When there are several similar observations in close proximity (less than half a degree), the data are combined or super-obbed to give one observation with generally smaller though occasionally equal error. The background error covariance is not well known. It is approximated as an ellipse of half width 200km in the north south and 1500km in the eastwest, in the equatorial region. At middle latitudes the error is taken to be more isotropic, with a half-width of 400km. The scheme is univariate, so that velocity fields are not altered as a result of assimilation of thermal data. Rather the model is allowed to spin-up the appropriate velocity field.
By blocking data into 10 day windows ( 5 days either side of the analysis time) and then performing an 01, there is a tendency to generate high frequency gravity waves after the analysis stage when the model is integrated forward to the next analysis time. Modifications have been made to the scheme to blend the data in more continuously by calculating the analysis increment from the 01 every 10 days, but applying the correction, I, in small increments TIN every timestep, where N is the number of timesteps in 10 days, (denoted 01-2). A third scheme where an 01 analysis is made every day using data in a sliding window of 10 days has also been tested (01-3). No forecasts have yet been made from these latter versions although the analyses appear smoother, as the velocity field has longer to adjust to a thermal increment, so generating fewer gravity waves.
The scheme used by Leetmaa et al 1996 in which data were blocked into two week blocks and an analysis done almost every time step is similar to 01-3, except that the data window is 30 days with data weighted in time, highest weight being given to data corresponding to analysis time and zero weight to data +- 15 days from analysis time. The NCEP scheme, like 01-3, has the effect of giving greater weight to the data as each datum is used many times, but this violates one of the assumptions of 01 that the data are uncorrelated in space and time.
