Spatiotemporal estimation of analysis errors in the operational global data assimilation system at the China Meteorological Administration using a modified SAFE method

Abstract

AbstractQuantification of the uncertainties in initial analyses against the real atmosphere (“reality”) provides a fundamental reference for the evaluation and development of operational data assimilation (DA) systems. Due to the unknown reality, most existing methods for analysis error estimation use reanalysis datasets or observations as a proxy for reality, which are empirical, non‐objective, and biased. Unlike these methods, our study adopted a modified Statistical Analysis and Forecast Error (SAFE) estimation method to objectively and directly quantify spatiotemporal errors in analyses compared to reality based on unbiased assumptions. In the present study, the SAFE method was first applied to estimate the annual variation and spatial distribution of analysis errors in the Global Forecast System of Global/Regional Assimilation and PrEdiction System (GRAPES_GFS) at the China Meteorological Administration (CMA) since the beginning of its operational implementation (i.e., 2016–2021). Qualitative comparison to analysis error estimations in previous studies showed that SAFE can provide more reasonable spatial‐mean analysis error profiles than can the estimation with the ERA‐5 reanalysis as a reference (the approach hereafter called “ERAv”). Moreover, ERAv overestimates (underestimates) the spatial‐mean analysis error below (above) approximately 500 hPa compared to SAFE because it neglects the uncertainties inherent in reanalysis. Overall, the SAFE estimation reveals that relative reductions of about 12.5%, 29%, and 24.5% were achieved for the spatial‐mean analysis errors of wind, temperature, and geopotential height, respectively, in the GRAPES_GFS throughout the six‐year study period. These results can largely be attributed to the DA scheme being upgraded from 3D‐Var to 4D‐Var. SAFE can also provide more reasonable and accurate pointwise analysis errors than ERAv can.