Implementation of a new nonlinear quality control scheme for the RTMA/URMA at NCEP

In order to improve the utility of data from a wide range of sources in NCEP's Real-Time Mesoscale Analysis (RTMA) and Un-Restricted Mesoscale Analysis (URMA) 2DVar schemes a new nonlinear quality control scheme is being implemented. Unlike conventional data quality control, that involves the discontinuous binary acceptance/rejection decision which must be performed prior to the variational analysis procedure, nonlinear quality control is an objective Bayesian procedure that down-weights each observation as a function of its departure from the analysis itself in a way dictated by the assumed form of the observation's non-Gaussian error distribution. Through this procedure, the scheme is able, in principle, to use the iterated analysis to mediate between close observations and thereby perform what amounts to an implicit “buddy check” in order to arrive at an allocation of effective weights incorporating the objective probabilistic assessment of their relative effective precision. The probability model we employ is a family generalizing the “logistic” probability distribution, with exponentially decaying tails which, while fatter than those of a Gaussian, do not introduce multi-modality into the implied cost function we seek to minimize. This means that the scheme can be “switched on” from the outset, unlike an older version of NCEP's nonlinear quality control which could only be safely applied after many iterations of the analysis solver had already been used up. The new implementation has necessitated a re-tuning of the error characteristics of the data to which it has been applied. These aspects of the implementation of the scheme will be discussed, together with a preliminary assessment of the results we have achieved with it.