Two years ago we reported on studies at JPL using radio occultation (RO) temperature profiles, produced in 2003 by CHAMP and SAC-C, in three-way comparisons with AIRS temperature retrievals (Ver. 4 software) and ECMWF daily analyses. The expectation in such comparisons, based on both first principles and previous experience with the data, is that RO will offer the most precise and least biased temperatures, and the highest vertical resolution, between about 5 and 20 km altitude, and thus can serve as a standard for evaluating other techniques over that altitude range. This is the so-called "cal/val" role for which RO has long been recommended.

The early results offered some surprises: The RO data tended often to be the "odd man out," showing significantly poorer agreement with AIRS and ECMWF than they showed with one another. Several lines of argument and analysis led to the suspicion that there was a significant correlation between AIRS and ECMWF temperature errors, causing them to agree with one another more closely than their individual accuracies would lead one to expect. Such a correlation could arise in the AIRS initialization process, derived from a set of ECMWF training samples from 2002, provided those "first guess" values were not substantially adjusted in the AIRS retrieval process. After accounting for the correlation we estimated that individual RO profiles were accurate to 0.4-0.7 K over the 5-20 km range; ECMWF profiles were accurate to 0.7-1.2 K; and AIRS retrievals were accurate to 1.0-1.6 K. All values are one sigma.

Those results were obtained with profiles produced in 2003. Since then there has been a good deal of progress on all fronts. The AIRS team has released its Version 5 retrieval software featuring a number of changes from Version 4. ECMWF has likewise made a series of upgrades in their analyses, including introducing a greater number of vertical levels and a higher maximum altitude. Meanwhile, COSMIC has come on the scene as the definitive RO mission to date, providing greater quantities and a higher quality of RO data, and RO retrieval methods have continued to be refined.

We have now performed comprehensive new comparisons of AIRS, RO, and ECMWF temperature profiles using data from 2006 and 2007. There have been distinct changes in the relative performance of the three techniques, and the degree of correlation between AIRS and ECMWF appears to have been significantly reduced. The actual measurement (or model) accuracies appear to have improved in some cases and in other cases not. This presentation will summarize the latest results and conclusions from these comparisons.