A NOAA-STAR dataset is born…

Spread the word

What does a new entrant in the lower troposphere satellite record stakes really imply?

At the beginning of the year, we noted that the NOAA-STAR group had produced a new version (v5.0) of their MSU TMT satellite retrievals which was quite a radical departure from the previous version (4.1). It turns out that v5 has a notable lower trend than v4.1, which had the highest trend among the UAH and RSS retrievals. The paper describing the new version (Zou et al., 2023) came out in March, and with it the availability of not only updated TMT and TLS records (which had existed in the version 4.1), but also a new TLT (Temperature of the Lower Troposphere) record (from 1981 to present). The updated TMT series was featured in the model data comparison already, but we haven’t yet shown the new TLT data in context.

Readers will recall that the TLT product is nominally a weighted average of atmospheric temperature anomalies from the surface up to 5km or so. The weighting varies a little between land and ocean, and as a function of topography or surface type (some model-observation comparisons take this into account, but a global uniform weighting is often good enough). The nature of the measurement, using off-nadir scans of the instrument made the retrieval more noisy than other MSU prodcuts, and it has taken time to deal with those issues. Some long-timers might even recall the rather tumultuous history, involving over-confident claims of precision, the discovery of systematic biases because of orbital decay, corrections, independent replication and more errors, more corrections, etc. This history should temper any claims now that the structural uncertainty has finally been beaten down, but it’s worth digging in a little deeper to see where it comes in.

Intra-TLT differences

First, how do the three TLT versions compare? I’ve made two versions of this graph to highlight where and how the three lines differ. It certainly isn’t as simple as just a shift in the linear trend.

A NOAA-STAR dataset is born…Time series of TLT retrieval anomalies from UAH, RSS and NOAA STAR, baselined to 2001-2010. All lines show clear upward trends that are more coherent after 2000.

As expected, the year-to-year variations are very similar, but there are notable divergences between 1996 and 1999 that are (mostly) related to the treatment of data from NOAA-14 which had a large orbital and instrumental drift. The trends prior to 1995 (0.07/0.14/0.16 ºC/dec, for UAH, RSS and NOAA-STAR respectively) and after 2001 (0.14/0.20/0.17 ºC/dec) vary across the products too. Thus the similarity of the full period trend (1981-2022) between UAH and NOAA-STAR is somewhat coincidental (0.14/0.20/0.13 ºC/dec) made up of a larger trend in NOAA-STAR to ~1988, a smaller trend to 2000, and a slightly larger trend in the last two decades. This heterogeneity in differences is very likely structural uncertainty in how the records are constructed, and the span of trends in the three products is a likely an underestimate of true uncertainty. It’s not a democracy where the ‘right’ answer is decided by majority vote!

Kitchen sink comparisons

How does the TLT new record stack up against the surface records? Here we can compare to the in situ surface data records (GISTEMP, HadCRUT5, NOAA NCEI), the radiosondes, the reanalyses (ERA5 and JRA55), and (over a shorter period), the AIRS satellite retrievals. Each of these has it’s own issues but they bring a wealth of independent data to bear on the issue. Similarly to above, I include two versions of the graphs with different baselines.

Time series of surface and near surface products including TLT retrieval anomalies from UAH, RSS and NOAA STAR, surface station based records, radiosonde data and reanalyses, baselined to 2001-2010. All lines show clear upward trends that diverge mostly in the 1990s. Time series of surface and near surface products including TLT retrieval anomalies from UAH, RSS and NOAA STAR, surface station based records, radiosonde data and reanalyses, baselined to 1981-1990. All lines show clear upward trends that diverge slightly after 2000.

The overwhelming impression from these graphs is the similarity of all these records, and not just in the year to year variations. The upward trends differ slightly for sure, but they are all recognizably describing the same climate change. Curiously, the TLT records bracket the spread of the other independent datasets, suggesting that the structural uncertainty is simply larger in the satellite retrievals (including the different versions of the AIRS data).

But why should this be so? Historically, there has been a lot of discussion about non-climatic effects in the surface stations and ocean data – station moves, urban heating, instrument changes etc. While these are important effects, they are often local. Stations globally did not move at the same time, instrument changes happened at different times in different places, areas urbanized at different rates and at different times. Thus the implications for what one does about these issues mostly have local impacts. There are systematic changes that have bigger implications – for instance the change of data sources in shipping in the 1930s/40s/50s and the aliasing of errors in instruments and coverage in the ocean – and those corrections dominate the impact of adjustments on the global mean surface temperature trends.

Now, let’s think about how the TLT satellite data are processed. There are corrections for each satellite in the time series (now up to 16 instruments) for orbital decay, orbital drift, instrument calibration drift, etc. There is some overlap between successive satellites, but there are still uncertainties in what corrections are needed and what source of information should be used to do that correction. The key thing to remember that each of those uncertainties applies to the length and totality of that satellites record, and different choices will lead to different trends. Thus uncertainties in the satellite corrections almost invariably have an impact on the longer term global trends.

The AIRS satellite record is also interesting. This is from a single instrument on the NASA Aqua satellite, that, until last year, was in a controlled, non-drifting orbit. This means some of the issues that affect the MSU/AMSU instruments don’t apply. However, the trends in different versions of the retrievals (i.e. v6 to v7) can be quite variable. In this case, the uncertainty comes in with the retrieval algorithm and treatment of confounding effects like clouds or surface emissivity changes. As I understand it (and someone correct me if I’m wrong!), the AIRS retrievals work by assuming a (realistic) prior atmospheric profile (surface temperature, vertical profiles of temperature, humidity, cloud cover, aerosols, ozone etc.) for which the spectral signal can be calculated, and then the (small) deviations seen in the actual retrieved data can be easily associated with small deltas in the inputs. But the further away the prior profile is from the actual profile, the more complicated and error-prone the retrieval is. In version 6, the prior profiles were all from the early part of the time series, which mean that the early years had pretty accurate retrievals, but the later years (with climate trends in all the inputs), they were less accurate. For version 7, the prior profiles were better spaced through time which evened out the uncertainty, and thus impacted the trends too. There were other processing steps that changed as well. The point though is that a change in the algorithm affects the whole record and so can have a systematic impact on the trends.

Thus the additional of a new version of the TLT record from NOAA-STAR helps underline the continuing structural uncertainty in these records, and it’s clear, that unlike the surface temperature record, we aren’t (yet) seeing a convergence on the ‘right’ answers.

What’s to come?

Can we expect further improvements in the uncertainty estimates? Absolutely. For a number of the surface station-based products (notably HadCRUT5, GHCN and the ERSST products), better uncertainty models have been developed using a Monte Carlo approach of creating an ensemble of products each of which made slightly different choices in the corrections (within reasonable bounds). Something similar for the satellite data sets would be very interesting. This effort is computationally expensive and requires a lot of attention to detail (including encapsulation of different models for corrections, not just different parameters within a specific model) for it to be complete, but this is coming to be seen as the gold standard for capturing the ‘real’ structural uncertainty when there are complex and non-linear data processing workflows. Watch this space.


  1. C. Zou, H. Xu, X. Hao, and Q. Liu, “Mid‐Tropospheric Layer Temperature Record Derived From Satellite Microwave Sounder Observations With Backward Merging Approach”, Journal of Geophysical Research: Atmospheres, vol. 128, 2023. http://dx.doi.org/10.1029/2022JD037472

The post A NOAA-STAR dataset is born… first appeared on RealClimate.