Olaf Stiller, Deutscher Wetterdienst (DWD)

A cross-validation method for observational data and its application to IASI cloud screening


The exploitation of remote sensing data for NWP strongly relies on quality control (QC) methods including (often quite sophisticated) screening procedures aimed at identifying observations affected by influences (as, e.g., from clouds or land surfaces) for which the employed observation operator is not adequate. Apart from the most simple QC checks, most screening methods consider collective properties of groups of observations for which they impose thresholds. In a more general sense, such QC schemes can be considered as cross-validation methods as the validity of observations is assessed from their consistency with other observations (plus the model background).

While these consistency assessments are generally quite heuristic, the work presented here gives a mathematically rigorous framework for computing the conditional probability of observations (or subsets of observations) given the background and other observations.

Ways for constructing a cloud screening scheme for IASI data within this probabilistic framework are presented and discussed. Such a scheme requires diagnostics for the identification of groups of observations which are affected by systematic influences (like, e.g., clouds). First a method is presented which flags quite generally observation groups which are inconsistent with the assumptions made by the DA system (error covariances and observation operators). This method, however, appears to be over-restrictive as the number of rejected data is comparably large (many of them seem to be rejected for reasons not linked to clouds). To increase the data yield, a more targeted method which restricts the search to specific types of influences has been developed. This more targeted scheme also allows to identify some limits for cross validation methods in general where, e.g., a reliable identification of very low clouds from infrared data alone is often not possible.