Observations — monitoring and quality control

Monitoring

We monitor meteorological observations received from a variety of sources worldwide. This is primarily to maintain and improve the use of these observations in our computer assimilation system (4D-VAR), which is part of the numerical weather prediction (NWP) system.

The quantity and quality of observational data received and assimilated are checked daily, and any problems followed up. Observations are compared with short-period forecast (background) fields and observation-minus-background (o-b) statistics are used for monitoring over various time periods. On a monthly basis any poor quality data that are identified are either added to reject lists and excluded from the assimilation or corrected prior to use.

Automatic quality control

The current automatic quality control system is based on Bayesian probability theory, and a careful statistical analysis of observation and background errors.

Bad data

Each observed element is given an initial 'probability of gross error' (PGE). For example, we expect about 1.5% of SYNOP pressure observations to be 'bad' and assign them an initial PGE of 0.015. This PGE is increased if the element has failed one of the earlier consistency checks, e.g. the pressure is checked against the pressure three hours earlier and the reported pressure tendency.

Good data

Even 'good' observations have small errors (e.g. barometer accuracy is about 0.2 hPa — inaccuracies in knowledge of the station height can introduce larger errors). We take account of the fact that observations include small scale detail, not resolved by the computer model. Including this factor, the observation error for good SYNOP pressure observations is estimated as 1.0 hPa. We also estimate the root-mean-square error of the background (forecast) fields.

These estimates depend on various elements.

• They are larger in the vicinity of fast-moving vigorous depressions than in large anticyclones
• Climatological elements — latitudes having large numbers of observations will have generally lower values (average about 1.3 hPa) than data sparse latitude zones (typical errors of 2-3 hPa)
• The most important single check is that against the forecast background (T+3 hours or T+6 hours), updating the PGE. It includes an estimate of the probability that the whole observation is wrong, e.g. position reported incorrectly

Several observation-type specific checks are applied.

• Calm aircraft winds are rejected
• An asymmetric check for cloud track winds
• Radiosonde standard and significant levels are checked against the background, and then averaged over the model layers omitting levels with PGE > 0.999
• 'Buddy check' — compares each observation against neighbouring observations