Assimilating 200 Years of Weather: The 20th Century Reanalysis Version 3 System

A new historical reanalysis, the NOAA-CIRES-DOE 20^th Century Reanalysis Version 3 (20CRv3), has been completed. It provides a 4-dimensional reconstruction of global weather spanning 1836-2015, with an experimental extension back to 1806, by assimilating only surface pressure observations into a modern global land-atmosphere forecast system with prescribed sea surface temperatures, sea ice concentrations, and radiative forcings. The historical reanalysis setup provides a useful testbed for new data assimilation methods because it uses a modern forecast model but only assimilates one type of conventional observation. In addition, the surface pressure observation network varies significantly in time: fewer than 5 observations are available globally in a 6-hour assimilation window in the early 19^th century, compared to tens of thousands in the 21^st century. The data assimilation method must therefore be robust to these network changes and use each observation in early time periods as effectively as possible, while determining which observations are most useful in the later time periods. The method should also provide reliable estimates of uncertainty throughout the entire timespan, though uncertainties in the prescribed boundary conditions will likely have an impact in the early data-sparse period.

The 20CRv3 system assimilates observations using an 80-member ensemble Kalman filter with several upgrades to the algorithms implemented in the previous 20^th Century Reanalysis version 2c (20CRv2c). In particular, 20CRv3 uses a nonlinear quality control algorithm on the observations, adaptive covariance localization, and a relaxation-to-prior-spread inflation method. These updates, along with a newer forecast model, larger ensemble size, and more available observations, have yielded smaller errors that are more consistently predicted from the ensemble spread than in 20CRv2c. Comparisons with conventional, upper-air, and satellite reanalyses, as well as with independent observations, demonstrate that 20CRv3 performs well in its estimates of atmospheric variables. Investigations of the ensemble spread show that 20CRv3 also has demonstrably more reliable estimates of uncertainty quantification over the nearly 200-year period.