Objectively estimating the temporal evolution of accuracy and skill in a global ocean reanalysis

The accuracy of ocean reanalyses is acknowledged to have increased dramatically during the last few decades, as a consequence of the evolution of the oceanic and atmospheric observing networks. Because of evolution of the atmospheric observing system, a similar accuracy increase also occurs in assimilation-free global ocean simulations. However, the temporal evolution of traditional accuracy metrics based on comparing reanalyses with observations is affected by the change in the number and spatial sampling of the verifying observations themselves. To disentangle the accuracy evolution and the changes in the verifying dataset, procedures insensitive to the latter are recommended. A simple and easy-to-implement methodology inspired by bootstrapping is proposed, specifically conceived for objectively estimating the accuracy and skill evolution of reanalyses over long (e.g. inter-decadal) periods. This allows the artificial impact of the observational network evolution on the verification metrics, which generally led to under-estimated accuracy at the beginning of the reanalysis period and much too sharp an increase later, to be overcome. Applied to the assessment of ocean temperature accuracy and skill in a state-of-the-art eddy-permitting global ocean reanalysis covering the period 1982-2012, the methodology illuminates the crucial impact on the upper ocean (top 100 m) reanalysis skill of both the atmospheric and oceanic observing network evolution, the former leading to a quicker skill increase. Conversely, the fundamental effect of the oceanic observing network below 100 m emerges from the study, as the assimilation-free simulation shows no skill at any time in the comparison against observations.