@article{gledocs_11858_10236,
author = {Stiller, Olaf},
title = {New impact diagnostics for cross‐validation of different observation types},
year = {2022-07-31},
volume = {148},
number = {747},
pages = {2853-2876},
publisher = {John Wiley & Sons, Ltd},
publisher = {Chichester, UK},

abstract = {New cross‐validation diagnostics have been derived by further partitioning well‐established impact diagnostics. They are related to consistency relations, the most prominent of which indicates whether the first‐guess departures of a given observation type pull the model state into the direction of the verifying data (when processed with the ensemble estimated model error covariances). Alternatively, this can be regarded as cross‐validation between model error covariance estimates from the ensemble (which are used in the data assimilation system) and estimates diagnosed directly from the observations. A statistical cross‐validation tool has been developed that includes an indicator of statistical significance as well as a normalization that makes the statistical comparison largely independent from the total number of data and the closeness of their collocation. We also present a version of these diagnostics related to single‐observation experiments that exploits the same consistency relations but is easier to compute. Diagnostics computed within the Deutscher Wetterdienst's localized ensemble transform Kalman filter (LETKF) are presented for various kinds of bins. Results from well‐established in‐situ measurements are taken as a benchmark for more complex observations. Good agreement is found for radio‐occultation bending angle measurements, whereas atmospheric motion vectors are generally also beneficial but substantially less optimal than the corresponding in‐situ measurements. This is consistent with reported atmospheric motion vector height assignment problems. To illustrate its potential, a recent example is given where the method allowed identifying bias problems of a subgroup of aircraft measurements. Another diagnostic relationship compares the information content of the analysis increments with a theoretical optimum. From this, the information content of the LETKF increments is found to be considerably lower than those of the deterministic hybrid ensemble–variational system, which is consistent with the LETKF's limitation to the comparably low‐dimensional ensemble space for finding the optimal analysis.},
note = { \url {http://resolver.sub.uni-goettingen.de/purl?gldocs-11858/10236}},
}