Estimation of the Impact of Sampling Errors in the VOS Observations on Air–Sea Fluxes. Part II: Impact on Trends and Interannual Variability

Abstract

Abstract Using the same approach as in Part I, here it is shown how sampling problems in voluntary observing ship (VOS) data affect conclusions about interannual variations and secular changes of surface heat fluxes. The largest uncertainties in linear trend estimates are found in relatively poorly sampled regions like the high-latitude North Atlantic and North Pacific as well as the Southern Ocean, where trends can locally show opposite signs when computed from the regularly sampled and undersampled data. Spatial patterns of shorter-period interannual variability, quantified through the EOF analysis, also show remarkable differences between the regularly sampled and undersampled flux datasets in the Labrador Sea and northwest Pacific. In particular, it is shown that in the Labrador Sea region, in contrast to regularly sampled NCEP–NCAR reanalysis fluxes, VOS-like sampled NCEP–NCAR reanalysis fluxes neither show significant interannual variability nor significant trends. These regions, although quite localized covering small parts of the globe, play a crucial role for the coupled atmosphere–ocean system. In the Labrador Sea, for instance, interannual and decadal-scale changes of the surface net heat fluxes are known to affect oceanic convection and, thus, the meridional overturning circulation of the Atlantic Ocean. From a discussion of current atmospheric data assimilation systems it is argued that in poorly sampled regions reanalysis products are superior to VOS-based products for studying interannual and interdecadal variations of atmosphere–ocean interaction. In well-sampled regions, on the other hand, conclusions about surface heat flux variations are relatively insensitive to the choice of the flux products used (VOS versus reanalysis data). The results are confirmed for two different datasets, that is, ECMWF 40-yr Re-Analysis (ERA-40) data and seasonal integrations with a recent version of the ECMWF model in which no actual data were assimilated.