Title: On the estimation and treatment of observation uncertainty in data assimilation for numerical weather prediction
Authors: Sarah Dance - University of Reading (United Kingdom) [presenting]
Abstract: In numerical weather prediction, forecasts are produced each hour via the numerical solution of a partial differential equation, starting from an initial estimate of the current state of the atmosphere. Variational data assimilation is used to produce these initial data, combining the latest observational data (of order $10^7$ observations) with numerical model forecasts (with state vector of dimension around $10^9$) to estimate the current state of the system. The variational assimilation problem has a Bayesian formulation, in which the observation and model forecast are assumed to be Gaussian variables with prescribed covariances. Until recently, most operational forecasting centres have assumed that the errors in the observations are uncorrelated. However, this is not always true, especially when considering multichannel satellite observations, and it has been shown that fully specifying observation error covariance matrices leads to more accurate forecasts. The latest approaches used in numerical weather prediction to estimate and treat observation uncertainty are reviewed. We will discuss covariance estimation, regularization and implementation in high performance computers.