An assessment of ground‐based GNSS Zenith Total Delay observation errors and their correlations using the Met Office UKV model

Ground‐based GNSS Zenith Total Delay (ZTD) observations have been assimilated into the Met Office numerical weather prediction (NWP) models since 2007, and into the Met Office UKV model since its introduction in 2009. The UKV model is a 1.5 km resolution convective‐scale model and uses a 3D‐Var assimilation system. There is a plan to upgrade the UKV assimilation system from 3D‐Var to 4D‐Var in the near future, giving the opportunity to increase the temporal resolution of ZTDs assimilated. The ZTD observation‐error covariances used operationally are assumed to be uncorrelated in both space and time despite the expectation that ZTDs have temporally and spatially correlated observation errors due to the production method (e.g. batch processing using a sliding window and (time) relative constraints). To assess whether these error correlations should be accounted for in order to use ZTDs at higher temporal resolution, a posteriori diagnostics to estimate the extent of temporal and spatial error correlations in ZTD observations over the UK, BENELUX and Northern France are used. Over two separate month‐long periods, we find that ZTD observations within the same processing batch are correlated, and that correlations persist between different batches to at least 1 h. Spatially, ZTD observations are found to be correlated to a minimum of 62.5 km. We find that the extent of the diagnosed correlation between observations separated in space and time is affected by the value of the relative constraints parameter chosen by the processing centre in the GNSS processing software. The impact of the relative constraints parameter on the diagnosed error variances is greater than that revealed by innovation statistics alone. Observations of Zenith Total Delay (ZTD) from the delay in GNSS signals travelling from the satellite to ground‐based receivers are produced using different software and processing options by different data suppliers, and subsequently used for assimilation into NWP models. Here we find that the processing strategies can arise in temporal and spatial error correlations which vary according to the different processing options, and which have length‐scales which need to be accounted for when ZTDs are assimilated into NWP models.