Man, I Hate Being Wrong

Guest Post by Willis Eschenbach

We’ll get to what I did wrong in a moment, but first, in my last post, a number of folks questioned my calculation of the surface temperatures from the CERES surface dataset. This is a dataset which is calculated from the CERES measured top-of-atmosphere (TOA) radiation measured by the satellites.

Let me start with a description of the CERES surface dataset from the developers:

EBAF-Surface Product Features

Global gridded, monthly mean surface fluxes calculated using a radiative transfer model.

Radiative transfer calculations are performed hourly on the CERES 1° equal-area grid.

Cloud properties are derived from narrowband imagers onboard both EOS Terra and Aqua satellites as well as Geostationary satellites to more fully model the diurnal cycle of clouds.

Gridded monthly mean cloud and atmospheric properties are adjusted so that model results:

• approach CERES net balanced Top-of-Atmosphere fluxes (EBAF-TOA product), where the global net is constrained to the ocean heat storage.

• more closely match modeled downward longwave surface fluxes that include active cloud base measurements from Calipso and Cloudsat.

Clear sky is separately adjusted to the monthly mean from CERES EBAF-TOA clear-sky ‘filled’ observations.

And here’s a flowchart showing how they get from the CERES measured TOA radiances to the surface datasets.

Figure 1. Flowchart showing how the CERES surface radiation datasets are calculated from the top-of-atmosphere (TOA) radiation measurements.

There is a lot more information about the calculation of the surface dataset where those came from. And for folks who want to see how the sausage is made, there’s a deeper dive into the process at Estimation of Longwave Surface Radiation Budget From CERES.

The CERES surface datasets include a dataset of the upwelling longwave from the surface. But that’s not much use to me. I wanted surface temperatures rather than surface upwelling longwave emissions. However, the Stefan-Boltzmann equation lets us convert from longwave emission to temperature if we know the emissivity.

The good news is that for natural substances, in almost all cases the emissivity is quite close to 1.0. Here’s a list of a few emissivities:

Water, 0.96
Fresh snow, 0.99
Dry sand, 0.95
Wet sand, 0.96
Forest, deciduous, 0.95
Forest, conifer, 0.97
Leaves Corn, Beans, 0.94

and so on down to things like:

Mouse fur, 0.94
Glass, 0.94

You can see why the error from considering the earth as a blackbody in the IR is quite small.

I must admit, though, that I do greatly enjoy the idea of some mad scientist at midnight in his laboratory measuring the emissivity of common substances when he hears the snap of the mousetrap he set earlier, and he thinks, hmmm … but I digress.

In any case, since I did not know the actual emissivity of the various surfaces, I decided to use 1.0 for the emissivity of all of them. I reasoned that it would only make a slight difference in the absolute temperatures, and it would not affect relative temperatures or trends at all.

So … how well does my converted dataset agree with the other global datasets? Since the ocean is 70% of the surface, let me start by comparing the converted CERES data to the Reynolds Optimally Interpolated sea surface dataset. Here is the average ocean temperature per Reynolds, minus the average ocean temperature per converted CERES.

Figure 2. Average Reynolds Optimally Interpolated Sea Surface Temperature minus average converted CERES temperatures. Averages are over the period of the CERES datasets, March 2000 to February 2018.

They are nearly identical almost everywhere … so why the difference at the poles? It’s because the Reynolds OI SST is showing the water temperatures, and the CERES data is showing the ice temperatures at the actual surface … so the CERES data is much colder at the poles. If we omit everything above and below the Arctic and Antarctic circles, we get the following:

Figure 3. Average Reynolds Optimally Interpolated Sea Surface Temperature minus average converted CERES temperatures. Averages are over the period of the CERES datasets, March 2000 to February 2018. Areas polewards of the Arctic and Antarctic circles have been excluded.

Note that there is much less than one degree of error in the averages. The agreement over the ocean is impressive.

Next, the land. I compared the converted CERES data to the HadCRU land-only data and the Berkeley Earth land-only data, along with the MSU UAH land-only satellite lower troposphere temperatures. Unfortunately, those three datasets are temperature anomalies, not absolute temperatures. So I couldn’t do a comparison of absolute values as I could with the Reynolds OI SST dataset. In lieu of that, here are the annual changes in the anomalies of the four land-only datasets …

Figure 4. Anomalies, land-only global temperature averages, from Berkeley Earth, converted CERES, HadCRUT, and the University of Alabama Huntsville Microwave Sounding Unit (UAH MSU) lower troposphere temperature.

The pairwise correlations of the datasets are quite similar, with the expected exception of the UAH MSU lower troposphere temperatures. These UAH MSU temperatures are measuring the lower troposphere and not the surface, so they are smoother and they don’t correlate quite as well with the other surface temperature datasets.

         Berkeley CERES HadCRUT UAH MSU
Berkeley       NA  0.88    0.89    0.85
CERES        0.88    NA    0.86    0.80
HadCRUT      0.89  0.86      NA    0.77
UAH MSU      0.85  0.80    0.77      NA

With that as prologue, I titled this post “Man, I Hate Being Wrong”. It has that title because my last post contained wrong calculations. So … where was I wrong in my last post? The error was not in the conversion of the CERES surface radiation data to temperatures as some people thought. That calculated temperature dataset, as shown above, is quite close to the other global ocean and land temperature sets.

Where I went wrong was in the calculation of the individual trends. Last week I’d written a new algorithm to calculate trends. And I thought that I’d tested it … but I’d tested it without remembering that trends in sinusoidal datasets are heavily affected by the choice of endpoints. Grrr … wrong again. So the main graphic in my last post is incorrect, with all of the trends being too low by about a tenth of a °C/decade. That definitely angrifies a man’s blood.

Here is the correct analysis of the decadal trends in the temperatures around the globe:

Figures 5a and 5b. Temperature trends for the period March 2000 to February 2018, correctly calculated this time.

Finally, and perhaps not surprisingly, now that I have the correct results, the trends of the converted CERES data and the two surface-station-based datasets are different. The trend of the CERES dataset is much closer to the trend of the UAH MSU lower troposphere dataset. Here are the decadal trends, land-only:

Berkeley Earth: 0.27 °C/dec

HadCRUT: 0.29 °C/dec

UAH MSU: 0.17 °C/dec

Converted Ceres: 0.16 °C/dec

And here are the trends of the ocean data:

Reynolds OI: 0.12 °C/dec

Converted Ceres: 0.14 °C/dec

UAH MSU: 0.11 °C dec

This agreement between the trends of the converted CERES temperature data with those two oceanic datasets strongly suggests that the land-station based datasets are trending high … the CERES dataset agrees with the UAH MSU on both land and sea, and with the Reynolds UI SST at sea.

Where is the difference in trends between CERES and Berkeley Earth? As you might expect, it’s generally where there are fewer surface temperature stations—the upper Amazon, central Africa, the Arctic …

Figure 6. Difference in trends, Berkeley Earth trends minus converted CERES trends.

So my main conclusions are:

• My previous post sucks. As much as I’d like to just ignore it and move on, I’m not built that way. I’m sworn to tell the truth as I best know it in all my posts, and that means admitting when I’m wrong even when no one else notices where the mistake is located. In this case, I was 100% wrong. Regarding that post, pay no attention to that man behind the curtain.

• This analysis agrees with my earlier comparisons of the converted CERES temperature dataset with other datasets. The absolute values and the month-to-month variations in my conversion of the CERES surface longwave radiation data to temperature are very close to the other temperature datasets (Reynolds OI SST, Berkeley Earth Land Only, HadCRUT Land Only, and UAH MSU lower troposphere land and sea).

So I am at ease using the converted CERES data as a valid temperature dataset. However, the trends are different from the surface-station datasets … and that has almost nothing to do with the exact conversion from radiation to temperature—trends in one will be trends in the other.

• The trend of the CERES temperatures is quite similar to the MSU UAH trend over both land and sea, and to the trend of the Reynolds OI sea surface temperature. I suspect that this indicates that the station-based land trends are gradually affected over time by encroaching civilization—more blacktop, more roads, more auto traffic, more jet exhaust and more powerful jet engines at the airport stations, more air conditioners, more sewage plants, more buildings, taller buildings, the list of things that can bias temperatures upwards is long.

• Man … I hate being wrong.

My very best wishes to everyone, with hopes that y’all can avoid being publicly wrong, it’s no fun at all.

w.