A Bayesian Approach to Observation Quality Control in Variational and Statistical Assimilation,
Abstract
Bayesian methods are ideally suited to the ongoing operational data assimilation needed for numerical weather prediction (NWP). Observational errors can be treated as random variables, and we have a long experience of previous observations over which to build up an estimate of their distribution. This experience tells us that observation error distributions are typically non-Gaussian; there are more large errors than expected. It is the handling of these gross errors that we call quality control. As well as the observations, we also need, and have, much other information about the atmosphere. Indeed this prior information is more valuable than that from the observations at any one time. We have a forecast background field, based on the accumulated knowledge from previous observations, which is usually rather accurate. A forecast based on the background, with no new observations, would probably be more accurate than one based on a batch of observations, with no background. So it is essential to give proper weight to this prior knowledge; the Bayesian approach allows us to do this.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1993
- Accession Number
- ADP008738
Entities
People
- Andrew C. Lorenc