The data assimilation algorithms derive the optimum analysis based on the background (previous forecast) and the observations, taking into account the error estimates of both background and observations. In order to infer corrections to the model fields from observation information, the observations have to be compared to the model state. This comparison is expressed in terms of observation increments, i.e. differences between a value representing the observation and a value representing the model. Either of these two values may be the result of a kind of interpolation, extrapolation, and possibly non-linear function of other values.
A number of checks (comparison of observed values with background or climatological values, test for consistency, blacklist, data selection for geographical area or height range) are applied in order to identify and reject outliers, so that they are not used in the subsequent analysis.
The background check rejects observations, those deviations from the background is larger than σfg times the expected standard deviation (efg2
Where the observed value is denoted as o, the forecasted value as fg, and the specified observational and background error as eo and efg
, respectively. Prescribed values for bounds to be used are denoted as σfg
Furthermore, the observed values have to be confined by prescribed climatological or physically meaningful bounds.
On the basis of these checks observations used in the assimilation are denoted as ACTIVE or ACCEPTED. The ACCEPTED flag indicates that the observation obtained passed all checks but not the background one. Instead, the ACTIVE flag indicates that the observation passed all checks.
Observations not used in the assimilation are denoted as REJECTED if they are dismissed due to insufficient quality (did not pass all of the quality control checks). They are denoted as PASSIVE if they are not assimilated but processed by the assimilation system just for monitoring purposes. The PASSIVE-REJECTED flag indicates that passively monitored observations did not pass the quality control checks. Observations denoted as DISMISSED are rejected without being written to any of the monitoring file.
At every analysis step, observations are held within reports (see table below as an example).