96 Andrew C. Lorenc
5.8 Monitoring
To the manager of a manufacturing company, Quality Control has a different
meaning to that we have implied so far; he wants to know about and prevent errors.
In NWP we call the equivalent process "monitoring". The purpose is to collect statistics on the performance of observing and processing systems, to detect systems
that are not performing as expected, and to feed this information back so the deficiency is corrected at source. To do this we need:
- a comprehensive database of basic and proces sed observed values, independent
estimates of the same quantities, and parameters affecting the processing
- software for categorising, sorting, and analysing the database
- effort to try categorisations and look for "unexpected" behaviour
- communications, willpower and persistence, to get errors from stages out ofyour
direct control rectified.
Design of the monitoring system is as important as design of the data-assimilation scheme; it should not be added on as an afterthought. Monitoring by NWP
centres of the operational World Weather Watch observations is probably the
major cause of the significant increase in observational quality which has been
seen in radiosondes (S. Uppala, personal communication), ships (C. Heasman, personal communication), and cloud track winds.
Another important product of monitoring is a good description of the observational error characteristics. If we are using the gross-error model, we need to know
the prior probability of error, and the error distributions of gross-error and of
"good" observations. Without these, the quality control is not objective 6 . Lorenc
and Hammon (1988) showed how the statistics of observations processed by their
quality-control scheme could be used (with some added "judgement") to "bootstrap" the assumed prior distributions.
For some observation types, more complicated error models are called for. For
instance there are many different ways that a radiosonde temperature and geopotential report can be corrupted. Gandin et al. (1993) have devised a "Comprehensive Quality Control" scheme which looks for sixteen. Because of the redundancy
of information in a radiosonde message, it is often possible to correct errors.
Many observations have bias errors, which monitoring statistics are useful in
detecting and correcting. For instance many ships and buoys have mean surf ace
pressure errors which persist until the instrument is recalibrated; the Met Office
routinely updates a list of corrections for them. The "observational" errors in satellite radiance soundings are biased by errors in the radiative transfer calculations
used in H; alI successful methods for using the radiances use empirical bias corrections obtained from a monitoring process.
6. By "objective" 1 mean more than the automatic application of ad hoc rules, rather that the
rules themselves have some statistical foundation.
5.8 Monitoring
To the manager of a manufacturing company, Quality Control has a different
meaning to that we have implied so far; he wants to know about and prevent errors.
In NWP we call the equivalent process "monitoring". The purpose is to collect statistics on the performance of observing and processing systems, to detect systems
that are not performing as expected, and to feed this information back so the deficiency is corrected at source. To do this we need:
- a comprehensive database of basic and proces sed observed values, independent
estimates of the same quantities, and parameters affecting the processing
- software for categorising, sorting, and analysing the database
- effort to try categorisations and look for "unexpected" behaviour
- communications, willpower and persistence, to get errors from stages out ofyour
direct control rectified.
Design of the monitoring system is as important as design of the data-assimilation scheme; it should not be added on as an afterthought. Monitoring by NWP
centres of the operational World Weather Watch observations is probably the
major cause of the significant increase in observational quality which has been
seen in radiosondes (S. Uppala, personal communication), ships (C. Heasman, personal communication), and cloud track winds.
Another important product of monitoring is a good description of the observational error characteristics. If we are using the gross-error model, we need to know
the prior probability of error, and the error distributions of gross-error and of
"good" observations. Without these, the quality control is not objective 6 . Lorenc
and Hammon (1988) showed how the statistics of observations processed by their
quality-control scheme could be used (with some added "judgement") to "bootstrap" the assumed prior distributions.
For some observation types, more complicated error models are called for. For
instance there are many different ways that a radiosonde temperature and geopotential report can be corrupted. Gandin et al. (1993) have devised a "Comprehensive Quality Control" scheme which looks for sixteen. Because of the redundancy
of information in a radiosonde message, it is often possible to correct errors.
Many observations have bias errors, which monitoring statistics are useful in
detecting and correcting. For instance many ships and buoys have mean surf ace
pressure errors which persist until the instrument is recalibrated; the Met Office
routinely updates a list of corrections for them. The "observational" errors in satellite radiance soundings are biased by errors in the radiative transfer calculations
used in H; alI successful methods for using the radiances use empirical bias corrections obtained from a monitoring process.
6. By "objective" 1 mean more than the automatic application of ad hoc rules, rather that the
rules themselves have some statistical foundation.