This is a guest post by Steven Mosher, whose previous guest post was about skeptics demanding adjustments. This one is about discrepancies between surface temperature and satellite temperature datasets. I don’t really need to say more.
Guest post: Surface and Satellite Discrepancy
With the publication of a new version of RSS’s data product the controversy over the accuracy of satellite data is likely to intensify. Prior to the publication of this new data, I took some time to do exploratory data analysis of RSS and the Berkeley Earth surface product. My tentative conclusion was that there were two areas that merited deeper investigation: The performance of satellite products at high latitudes and the transition between MSU data and AMSU data. With the publication of the new data, my analysis will have to be revisited ; however, there are still some things to be learned from looking at the prior version. In a complete analysis the various uncertainties of the data would have to be considered. At this stage I am only looking for fruitful areas to explore based on the known differences between the products.
The most superficial way to compare satellite data with surface data is to compare the global anomalies. Figure 1 illustrates the difference between RSS and BE. RSS data is for TLT, from the surface to several kilometers, while BE is a combination of air temperature over land and water temperature.
Over the entire period of record for RSS the difference in trend is roughly .05C per decade. However, as we will see when we push deeper into the data the reasons for this difference are not simplistically resolved.
There are several reasons why we should not be surprised about a difference between the two series:
- They use different estimation techniques: Satellites estimate the temperature of the bulk atmosphere by inference, translating brightness at the sensor face into temperature of the air near the surface. Surface datasets use more direct measurements to do their estimates
- They estimate different things. Satellites estimate the temperature of the entire column of air, while the surface data represents a layer of air and a layer of water.
- They take measurements at different times: Surface air measurements are made twice a day at Tmax ( whenever that occurs ) and Tmin (whenever that occurs ). SST measurements are not made twice per day, but rather randomly with respect to time. Satellite measurements are not made at Tmax or Tmin, but rather are made whenever the satellite overpasses the section of earth being sampled. This will change over time and requires adjustment.
- They have different coverage. While it is generally assumed that satellites have global coverage, this is not exactly true. Both the RSS and UAH product interpolate over swath gores, and both exclude high elevation areas where emissions from the earth’s surface corrupt the brightness data.
- The satellite record covers a period in which there was a major change in the sensor type used at a specific date. This is the shift from MSU to AMSU which occurs in May of 1998.
There are other differences between the datasets , but for now I will not focus on them.
Figure 2 illustrates the difference between BE global and RSS global anomalies. Since they use different baseline periods, the series were shifted to the same period. This shift does not change the difference in trends we see between the series. Two linear regressions were performed on the data. A slope of 0 indicates that the difference between the series is not changing over time.
The green line below represents the slope of the difference between BE and RSS over the MSU period: 1979-May 1998. The red line represents the difference between the entire series—roughly .05C per decade.
As the figure suggests during the MSU period the two records do not differ. The both measured the same amount of warming. The principle difference between the records arises because of differences that occur after the inclusion of AMSU data. Unfortunately, the switch over to AMSU happens at an inopportune time. This somewhat complicates the diagnosis of the difference to merely being a change of instrument issue. However, on its face the data suggests that further investigation of this transitional period is warranted. In other words, the data says dig here.
The difference is coverage and time of observation is more difficult to assess unless we turn to actual temperature fields. As noted above UAH does not publish temperatures. They compute them, but don’t publish that data. RSS absolute temperature fields are available, a sample month is shown below to illustrate the spatial coverage of the data. White areas have no data.
The RSS field is not global. There is no data at the poles and no data at locations of high land elevation ( see the white gores near the Andes and Tibet ). RSS temperature fields are produced in a 2.5 Degree equal angle grid, while BE is a 1 degree equal angle. Both were resampled to a ½ degree grid to maintain the resolution of the data during raster algebra operations, and then the BE field was masked to match the RSS field. Where RSS has no data we remove the data from BE so that an apples to apples comparison can be made. Also note that the BE fields are generally warmer. The reason for this is that BE samples a layer at the surface, whereas RSS estimates the temperature of a volume of air from the surface to several kilometers. The bulk of this is around 700hPA.
To compare the difference between the masked fields and determine if coverage differences between the products has an effect the masked the temperature fields were integrated over space and an monthly anomaly was calculated for both using the base period of 1979-2015. The RSS curve, of course, does not change and the trend in the BE curve is diminished slightly as a result of matching the RSS coverage.
The differences between the series does not appear to result from a difference in coverage. And we still see a difference during the MSU and AMSU periods
Next we turn to comparisons over land and ocean. The primary reason for doing this comparison is that land temperatures in surface products are made twice a day whereas satellite products are adjusted to a single time of day. Also, the Satellite approaches make different assumptions about sensor returns over land versus sensor returns over the ocean. We start with the land fields.
Integrating the fields and turning them into anomaly series and differencing yields the following
The masked land series shows a difference that is twice as large. 0.113C versus .05C. However, the AMSU effect is still apparently present. Turning to the ocean data we see the following
We should recall that the ocean represents over 70% of the surface of the planet. Over that portion of the planet the satellite series shows a much better agreement with the surface data.
Further, we also note that the AMSU effect is still present. Over the 1979-1998 period there is effectively no difference between the rate of ocean warming measured by MSU and the rate of warming of SST measures. However, there is a difference once we include AMSU data into the satellite series.
The analysis above suggest two lines of inquiry: A more comprehensive look at MSU and AMSU differences and a more comprehensive look at how temperatures over land are estimated.
There are two aspects of land surface measurements that we can look at in this analysis. The first has to do with time of observation. The land data is measured twice a day. Once at Tmax and once at Tmin. Satellite data is not measured at any consistent time. Consequently RSS adjust their data to represent the temperature at local noon. For example, if the satellite overpass was at 8:49AM, the data is adjusted to represent a “local noon” measurement. RSS does this by applying a diurnal adjustment taken from GCM results. For the next series of charts, we will compare the trend in Tmin from BE with RSS “local noon” trend, and the Trend in BE Tmax with the RSS local noon. The concern here is as follows. The trends in Tmax and Tmin are not the same: Over this period BE Tmax trend is .3C per decade, while the trend in Tmin is .22C per decade. Averaged, they come out to a trend of .27C decade. However, RSS only measures the trend at local noon.
One thing that has been suggested is that the difference between RSS and land records has to do with UHI. Since UHI typically impacts nighttime temperature rather than daytime temperatures we might expect to see larger differences in the Tmin comparisons. We don’t. In summary, the primary difference between RSS and BE is in the land and not in the Ocean. And on the land the difference is greater if we look at Tmax rather than Tmin, suggesting that UHI is probably not a sufficient explanation for the difference between the series. The fact that the MSU are trends are small re enforces this point.
In reading through the RSS documentation there was one other assumption that got my attention: an assumption of constant emissivity. In order to estimate the temperature the RSS approach depends upon an assumption that the emissivity of the earth is constant over time. However, given the changes in landcover over the period in question (1979-Present) we know this is not strictly true. Cities change emissivity. Greening of the planet changes emissivity. And changes in snow cover could change the emissivity.
The next question was is there any spatial pattern of difference between the two records? To anwser this the monthly difference between the fields was calculated on a spatial basis. The pattern that emerged suggests another avenue for investigation. When we difference the BE fields and the RSS fields we can immediately note areas where there are temperature inversions. RSS is effectively estimating the temperature of the air kilometers above the surface. And in general it is colder than the surface. However when we difference RSS with BE over land we find areas where the temperature at altitude is warmer than the surface. As an example I have taken Radiosonde data for a single location as an illustration of what an inversion pattern looks like. Below I have plotted the sonde temperature at 00Z (UTC), 12Z and the records for BE Tmin, Tmax and Tave. The RSS temperature has been forced to fit the line at around hPa750. Since RSS measures the bulk air column and not a specific pressure level this is an illustration only, but it highlights as well the difficulties inherent in comparing a satellite product that integrates over an entire volume of space at changing times with a Sonde record that captures air temps at discrete levels at different times and a surface product that captures temperature twice a day.
Globally we see the following areas where RSS suggests a temperature inversion. Red areas are locations in which there never was a temperature version over the entire record. Blue depicts the regions where there was at least one month with a temperature inversion. The crosses depict the sonde stations where records are complete enough to do meaningful comparisons.
Below is a gray scale version of the percentage of time a particular area has temperature inversions according to RSS comparisons with BE
Using the inversion mask we can then compare the difference in those areas where there are no temperature inversions to those areas where there are inversions. First the non inversion areas.
The data here suggests the following. As with the Ocean there is little difference between land temperatures and satellite temperatures. And the differences that exist are strictly confined to the AMSU period of coverage. Over the ocean, over the land where there are no temperature inversions, during the MSU period, there is no difference between what satellites estimate and what the surface shows. This comparison hardly argues for a systematic bias in surface products. Turning to areas where RSS indicates there is a temperature inversion we see the following.
These results are summarized in the following table.
|All Land||No Inversion||Inversion|
At this point we can point to a couple areas that merit further investigations. First is the transition of MSU to AMSU. The transition occurred at an inopportune time ( the middle of 1998 ) basically in the middle of a temperature spike and the differences between MSU and AMSU may turn out to be unimportant. Given the new RSS dataset, this will be an interesting metric to re compute. Also, the data suggests that differences in land temperatures dominate and in particular those areas where temperature inversions dominate. Those are the areas I will dig in the new RSS data.
HatTip: To Eli and Tamino for inspiring the line of inquiry