I’ve moved to WordPress. This post can now be found at GISS Deletes Arctic And Southern Ocean Sea Surface Temperature Data######################
REFER TO UPDATE AT THE END OF THE POST
There are numerous blog posts and discussions about how the GISS global temperature anomaly product GISTEMP differs from the Hadley Centre and NCDC datasets. The repeated reasons presented for this are, GISS uses 1200km radius smoothing to fill in the areas of the globe with sparse surface temperature readings, and the area this has the greatest impact is the Arctic. Typically, a map or comparison of global temperature anomaly maps is included, similar to Figure 1. The top two maps were cropped from Figure 3 in the Real Climate post “2009 temperatures by Jim Hansen”. I added the third. The bottom map was created at the GISS Global Maps webpage. It’s a map of the GISTEMP Global Temperature Anomaly product with 250km radius smoothing for the calendar year 2005, the same year as the top two maps. I did not include a temperature scale because the bottom map was provided to allow a visual comparison of the spatial coverage of the HadCRUT product and the GISTEMP product with 250km radius smoothing. Examine the Arctic and the ocean surrounding Antarctica, the Southern Ocean. Notice a difference? In 2005, the HadCRUT data had better coverage of the Arctic and Southern Oceans than the GISTEMP dataset with 250km radius smoothing. What’s missing in the GISTEMP product? There’s no sea surface temperature data.
GISS DELETES POLAR SEA SURFACE TEMPERATURE DATA
The general regions where GISS deletes Sea Surface Temperature data are shown in Figure 2. Three areas are highlighted: two cover the Arctic Ocean, and a third surrounds Antarctica. The specific locations are clarified in the following. GISS then uses their 1200km radius smoothing to replace the sea surface data with land data.
Tilo Reber in his recent “Diverging views” post at Watts Up With That? noted that the GISS Current Analysis webpage includes the following statement:
“Areas covered occasionally by sea ice are masked using a time-independent mask.”
This means that vast regions of Sea Surface Temperature (SST) anomaly data in the Arctic Ocean and Southern Ocean are deleted from the GISTEMP record. GISS does not delete all of the Arctic and Southern Ocean SST anomaly data, just the data from the areas where the annual sea ice melt occurs, and those are good portions of them.
I have looked for but have not found an explanation for this exclusion of Sea Surface Temperature data in the papers provided on the GISTEMP references page.
THE AREA OF THE ARCTIC OCEAN WHERE GISS DELETES SST DATA
Figure 3 shows four Arctic (North Pole Stereographic, 65N-90N) maps prepared using the map-making feature of the KNMI Climate Explorer. The maps illustrate temperature anomalies and sea ice cover for the month of September, 2005. The calendar year 2005 was chosen because it was used in the RealClimate post by Jim Hansen, and September is shown because the minimum Arctic sea ice coverage occurs then. The contour levels on the temperature maps were established to reveal the Sea Surface Temperature anomalies. Cell (a) shows the Sea Ice Cover using the Reynolds (OI.v2) Sea Ice Concentration data. The data for the Sea Ice Cover map has been scaled so that zero sea ice is represented by grey. In the other cells, areas with no data are represented by white. Cell (b) illustrates the SST anomalies presented by the Reynolds (OI.v2) Sea Surface Temperature anomaly data. GISS has used the Reynolds (OI.v2) SST data since December 1981. It’s easy to see that SST anomaly data covers the vast majority of Arctic Ocean basin, wherever the drop in sea ice permits. Most of the data in these areas, however, are excluded by GISS in its GISTEMP product. This can be seen in Cell (c), which shows the GISTEMP surface temperature anomalies with 250km radius smoothing. The only SST anomaly data used by GISS exists north of the North Atlantic and north of Scandinavia. The rest of the SST data has been deleted. The colored cells that appear over oceans (for example, north of Siberia and west of northwestern Greenland) in Cell (c) are land surface data extending over the Arctic Ocean by the GISS 250km radius smoothing. And provided as a reference, Cell (d) presents the GISTEMP “combined” land plus sea surface temperature anomalies with 1200km radius smoothing, which is the standard global temperature anomaly product from GISS. Much of the Arctic Ocean in Cell (d) is colored red, indicating temperature anomalies greater than 1 deg C, while Cell (b) show considerably less area with elevated Sea Surface Temperature anomalies.
Basically, GISS excludes Arctic Ocean SST data from 65N to 90N and, for round numbers, from 40E to 40W. This is a good portion of the Arctic Ocean. Of course, the impact would be seasonal and would depend on the seasonal drop in sea ice extent or cover. The sea ice extent or cover has to decrease annually in order for sea surface temperature to be measured. I’ll use the above-listed coordinates for the examples that follow, but keep in mind that they do not include areas of sea ice in the Northern Hemisphere south of 65N where sea surface temperature data are also deleted by GISS. These additional areas are highlighted in Figure 4. They include the Bering Sea, Hudson Bay, Baffin Bay and the Davis Strait between Greenland and Canada, and the Sea of Okhotsk to the southwest of the Kamchatka Peninsula.
Note: GISS uses Hadley Centre HADISST data as its source of Sea Surface Temperature (SST) data from January 1880 to November 1981 and NCDC Reynolds (OI.v2) data from December 1981 to present. To eliminate the need to switch between or merge SST datasets, this post only examines the period from 1982 to present. And to assure the graphics presented in Figures 3 and 6 are not biased by differences in base years of the GISTEMP data and the Reynolds (OI.v2) SST data, the latter of which has only been available since November 1981, I’ve used the period of 1982 to 2009 as base years for all anomaly data.
WHY WOULD DELETING SEA SURFACE TEMPEATURE DATA AND REPLACING IT WITH LAND SURFACE DATA BE IMPORTANT?
Land Surface Temperature variations are much greater than Sea Surface Temperature variations. Refer to Figure 5. Since January 1982, the trend in GISTEMP Arctic Land Surface Temperature Anomalies (65N-90N, 40E-40W) with 250km radius smoothing is approximately 8 times higher than the Sea Surface Temperature anomaly trend for the same area. The Arctic Ocean SST anomaly linear trend is 0.082 deg C/ decade, while the linear trend for the land surface temperature anomalies is 0.68 deg C/decade. And as a reference, the “combined” GISTEMP Arctic temperature anomaly trend for that area is 9 times the SST anomaly trend.
By deleting the Sea Surface Temperature anomaly data, GISS relies on the dataset with the greater month-to-month variation and the much higher temperature anomaly trend for its depictions of Arctic temperature anomalies. This obviously biases the Arctic “combined” temperature anomalies in this area.
GISS DELETES SEA SURFACE TEMPERATURE DATA IN THE SOUTHERN HEMISPHERE, TOO
Figure 6 shows four maps of Antarctica and the Southern Ocean (South Pole Stereographic, 90S-60S). It is similar to Figure 8. Cell (b) illustrates the SST anomalies presented by the Reynolds (OI.v2) Sea Surface Temperature anomaly data. SST anomaly data covers most of the Southern Ocean, but GISS deletes a substantial portion of it, as shown in Cell (c). The only SST anomaly data exists toward some northern portions of the Southern Ocean. These are areas not “covered occasionally by sea ice”.
Figure 7 illustrates the following temperature anomalies for the latitude band from 75S-60S:
-Sea Surface Temperature, and
-Land Surface temperature of the GISTEMP product with 250km radius smoothing, and
-Combined Land and Sea Surface of the GISTEMP product with 1200km radius smoothing, the GISTEMP standard product.
The variability of the Antarctic land surface temperature anomaly data is much greater than the Southern Ocean sea surface temperature data. The linear trend of the sea surface temperature anomalies are negative while the land surface temperature data has a significant positive trend, so deleting the major portions of the Southern Ocean sea surface temperature data as shown in Cell (c) of Figure 6 and replacing it with land surface temperature data raises temperature anomalies for the region during periods of sea ice melt. Note that the combined GISTEMP product has a lower trend than the land only data. Part of this decrease in trend results because the latitude band used in this comparison still includes portions of sea surface temperature data that is not excluded by GISS (because it doesn’t change to sea ice in those areas).
ZONAL MEAN GRAPHS REINFORCE THE REASON FOR THE GISS DIVERGENCE
When you create a map at the GISS Global Maps webpage, two graphics appear. The top one is the map, examples of which are illustrated in Figure 1, and the bottom is a Zonal Mean graph. The Zonal Mean graph presents the average temperature anomalies for latitudes, starting near the South Pole at 89S and ending near the North Pole at 89N. Figure 8 is a sample. It illustrates the changes (rises and falls) in Zonal Mean temperature anomalies from 1982 to 2009 of the GISTEMP combined land and sea surface temperature product with 1200km radius smoothing. The greatest change in the zonal mean temperature anomalies occurs at the North Pole, the Arctic. This is caused by a phenomenon called Polar Amplification.
To produce a graph similar to the GISS plot of the changes in Zonal Mean Temperature Anomalies, I determined the linear trends of the GISTEMP combined product (1200km radius smoothing) in 5 degree latitude increments from 90S-90N, for the years 1982 to 2009, then multiplied the decadal trends by 2.8 decades. I repeated the process for HADCRUT data. Refer to Figure 9. The two datasets are similar between the latitudes of 50S-50N, but then diverge toward the poles. As noted numerous times in this post, GISS deletes sea surface temperature data at higher latitudes (poleward of approximately 50S and 50N), and replaces it with land surface data.
Figure 10 shows the differences between the changes in GISTEMP and HADCRUT Zonal Mean Temperature Anomalies. This better illustrates the divergence at latitudes where GISS deletes Sea Surface Temperature data and replaces it with land surface temperature anomaly data, that latter of which naturally has higher linear trends during this period.
There appears to be some confusion in the comments in the WattsUpWithThat thread of my post GISS Deletes Arctic And Southern Ocean Sea Surface Temperature Data about what this post illustrates. I prepared a graph for this post but chose not to use it since it appeared redundant to me. It should clarify what is being presented. It is a comparison graph of GISTEMP Arctic Surface Temperature anomalies for the grid 65N-90N, 40E-40W, which is a major portion of Arctic Ocean as shown above in Figure 4. One dataset is the combined land plus sea surface data; the other is the land-only data. The two datasets are identical. If you subtract one from the other, the difference is 0.0 (zero) for all months. This indicates that there is no Sea Surface Temperature data in the combined product in this grid.
Update Figure 1
Does the Sea Surface Temperature data exist for this area? Yes. It is illustrated above as the green curve in Figure 5.
To me, this indicates that GISS deleted the sea surface temperature data for this portion of the Arctic.
Maps and data of sea ice cover and temperature anomalies are available through the KNMI Climate Explorer:
Excellent stuff. The divergence is quite big already, so even the polar areas are not that big, they do count.
I have started comparison of HadCRUT/MSU trends since 1979 in 60x30° grids accross the whole globe, like you did RSS vs GISTEMP. If you are interested in results, just e-mail me.
Juraj V: Thanks for the offer on the RSS/GISTEMP comparison, but I've got to get back to my long-term project and a video I've been working on.
Thanks for the analysis Bob. I am waiting for the folks over at Rankexploits to finish their reviews of the current GISSTemp processing, and then the follow-on review of how those processes effect temp calculations.
As we skeptics suspect, a very small annual bias (.001 - .0001) introduced by the processing will affect (perhaps dramatically) long range estimates.
CoRev: But as illustrated, the bias in the Arctic created by deleting the SST data is not small.
That's an interesting analysis and it shows the weakness of using SST in place of air temperature over the oceans.
Ideally, I think, the global temperature data sets would consist of air temperatures measured at a uniform height. However, air temperatures are not available in great numbers over the oceans so SST is used instead. SST anomalies in areas of open ocean tend to track air temperature anomalies closely, but this relationship breaks down near the coasts and over areas of sea ice.
SST is therefore something of a compromise. Model estimates of global temperature typically use the modelled 2m air temperatures, so SST is not the ideal thing to compare them with.
An additional problem is that in areas of ocean which were permanently ice covered during the climatology period it is not clear that the SST anomaly is well defined.
Nebuchadnezzar: Thanks for the comment. You wrote, “SST anomalies in areas of open ocean tend to track air temperature anomalies closely…”
Isn’t it the other way around, with NMAT mimicking SST?
You wrote, “Model estimates of global temperature typically use the modelled 2m air temperatures, so SST is not the ideal thing to compare them with.”
I didn’t compare models in this post. And many of the model runs used in AR4 (eleven, I believe) also model SST.
You wrote, “An additional problem is that in areas of ocean which were permanently ice covered during the climatology period it is not clear that the SST anomaly is well defined.”
The climatology period for the SST data is 1982 to 1992, and the Reynolds OI.v2 dataset also includes Sea Ice Concentration data, which are well defined. GISS also includes the Reynolds climatology to define the areas where they mask SST data.
"Isn’t it the other way around, with NMAT mimicking SST?"
The sea surface will affect the air above it and vice versa. I didn't want to specify any particular direction for the cause. I was simply pointing out that in an ideal world we'd use marine air temperature and SST is a reasonable substitute for that in some areas.
"I didn’t compare models in this post. And many of the model runs used in AR4 (eleven, I believe) also model SST."
I'm aware you don't compare to models, but I think it's worth keeping in mind as these datasets are frequently compared to models.
The models do generally model SST - whether or not it's stored - but when the global average is computed, I believe they generally use the air temperature and this is what the observed 'global' temperature is considered to be representative of.
From your comparisons which dataset would you consider to be most representative of the global average near surface air temperature?
Anonymous: You asked, "From your comparisons which dataset would you consider to be most representative of the global average near surface air temperature?"
I avoid land surface temperature datasets whenever possible. I wrote this post because what GISS was doing finally struck me.
"I avoid land surface temperature datasets whenever possible. I wrote this post because what GISS was doing finally struck me."
That point wasn't immediately clear from the post given that it was about GISS replacing SST with land surface air temperature. It just struck me as pertinent to ask why. Sorry to drag you off topic.
I downloaded the gridded anomaly data to have a more detailed look. They don't delete the data per se, what they are doing is way worse. When they combine the the land and ocean data into a single set, they make no adjustments for the increased number of land grid cells versus ocean cells.
In 1880 64N to 90N was calculated from 212 grid cells, 150 of them ocean. In 2009 they used 722 grid cells.... 150 of them ocean. But they just total them and divide by the number of cells as if the land area actually physicaly grew!
South zone is worse! ocean cells are 41 while land goes from 0 to 1200! They've made a .25 degree difference in the north and .75 in the south by doing the math as if the ratio in the number of data points also represented the physical ratio in size of land and ocean.
long, gory, detailed version
knowledgedrift said, "They don't delete the data per se, what they are doing is way worse..."
GISS states, regarding SST data, "Areas covered occasionally by sea ice are masked using a time-independent mask.”
This means that the data that is available is not used. In other words, they delete it.
The 150 cells in the Arctic with SST data you mention in your post are those cells where the ocean doesn't freeze in winter, which would be north of the North Atlantic and Scandinavia. GISS deletes all SST data for the remainder of the Arctic.
Please identify where the 150 cells with SST data are before you write that they "don't delete the data per se" and imply my post is in error.
I didn't mean to imply that you were in error. I was very focused on the change in ratio between land and ocean and how they handled that math. I'd have to go back and have a detailed look, but what I THINK happens is this:
1. There are 150 ocean cells in the time series from beginning to end for that zone.
2. When a land cell and an ocean cell overlap, a value appears in that gridded cell that matches neither the ocean anomaly nor the land anomaly.
3.I did not try and reverse engineer how they arrive at the merged value. I know it is not a strict average but I only tested that on a couple of cells.
4.I'm not exactly certain how they depict this on the map.
5. When I said they don't delete it per se, that is what I meant. Ocean cells start off as just ocean cells. As more land cells appear over time, some of them coincide with the same grid points in the ocean database and they merge the data by some formula. I noticed a couple of instances where the combined data showed a zero instead, but assumed it was a fluke, most of them had data.
I can actually pull the specific latitude and longitude for the "land only" "ocean only" and "overlap" if that helps any.
knowledgedrift: For the SST data in the Arctic, what are the latitudes and longitudes of those 150 cells?
Couldn't figure out how to upload a spreadsheet to Wordpress, so here it is as a pdf. The grid is every 2nd latitude, hence no even numbers, that's why it starts at 65 not 64.
knowledgedrift (June 18, 2010 5:40 PM): Thanks for the pdf documenting of the SST data coordinates. They correspond with the Arctic Ocean area north of the North Atlantic, and north of Scandinavia, as illustrated in my cell c of Figure 3 from the post:
GISS deletes SST data, basically, east of 40E and west of 40W.
Great post. Something leaped out at me.
Fig9 shows a remarkable difference between GISS and HadCrut in the poles. I recently read Chylek's 2010 paper on the polar seesaw effect. When I raised some of the points highlighted in that paper with others they countered it with the Polar Amplification idea.
One weakness in the seesaw idea for explaining the recent changes in polar temperatures has been that while Arctic temperatures have risen Anarctic temperatures haven't seen a similar fall, rather they've been fairly level in recent times.
It looks to me like data set choice would be crucial in trying to make this sort of call about the polar temperatures and maybe more importantly whats controlling them. Do you know whether NASA and CRU aknowledge this difference? How they justify both methods existing? And whether the wider polar science community has more trust in one or the other?
HR: You asked, “It looks to me like data set choice would be crucial in trying to make this sort of call about the polar temperatures and maybe more importantly whats controlling them. Do you know whether NASA and CRU aknowledge this difference?”
They understand the differences. Refer to the RealClimate post by Hansen:
And refer to the draft of Hansen et al (2010):
You asked, “How they justify both methods existing?”
Hadley Centre presents the data without infilling and with minimal adjustments, while GISS elects to use the 1200km radius smoothing to infill the data. There’s no real justification needed.
You asked, “And whether the wider polar science community has more trust in one or the other?”
Polar researchers wouldn’t use either one. Hadley has little Arctic data, while the GISS data in most of the Arctic is make believe.
Bob, are you aware of Hansen's latest?
"We use ocean temperature change only in regions that are ice free all year (a map of this area is included in Appendix A) because our data set is intended to be temperature change of surface air. Surface air temperature (SAT), measured at heights of 1.25–2 m at meteorological stations, is of most practical significance to humans, and it is usually SAT change that is reported in climate model studies. Change of sea surface temperature (SST) should be a good approximation to change of SAT in ice-free ocean areas; climate model simulations [Hansen et al., 2007] suggest that long-term SAT change over ice-free ocean is only slightly larger than SST change. However, ocean water temperature does not go below the freezing point of water, while surface air temperature over sea ice can be much colder. As a result, SST change underestimates SAT change when sea ice cover changes. Indeed, most climate models find that the largest SAT changes with global warming occur in regions of sea ice [IPCC, 2007]. Thus, we estimate SAT changes in sea ice regions by extrapolating actual SAT measurements on nearby land or islands; if there are no stations within 1200 km, we leave the temperature change undefined."
Anonymous December 20, 2010 12:52 PM: Yes, I'm aware of that discussion in the recent Hansen paper. It, however, does not justify the deletion of SST data when the seasonal sea ice has melted and there is open ocean in areas of the Arctic. That's when SST data should be used.
Anonymous December 20, 2010 12:52 PM:
PS: I just took at quick look through the comments and found that I had linked the preprint of Hansen et al (2010) in the comment before yours.
Could you use the ccc-gistemp code to explore what differences there are between using SST when the relevant area is ice-free vs. Hansen's method?
Anonymous @ December 21, 2010 1:13 PM: I don't believe there's any need for the code if one wanted to determine the actual impact on local temperatures. We know what grids are masked, and the Reynolds OI.v2 data could be used to determine the effective open ocean area and the SST anomalies during a given month.
Post a Comment