As part of my exploration of different surface temperature datasets, I’m examining the relationship between average U.S. temperatures and other weather variables in NOAA’s Integrated Surface Hourly (ISH) dataset. (I think I might have mistakenly called it “International” before, instead of “Integrated” Surface Hourly).

Anyway , one of the things that popped out of my analysis is related to our record warm March this year (2012). Connecting such an event to “global warming” would require either lazy thinking, jumping to conclusions, or evidence that the warmth was not caused by persistent southerly flow over an unusually large area for that time of year.

The U.S. is a pretty small place (about 2% of the Earth), and so a single high or low pressure area can cover most of the country. For example, if unusually persistent southerly flow sets up all month over most of the country, there will be unusual warmth. In that case we are talking about “weather”, not “climate change”.

Why do I say that? Because one of the basic concepts you learn in meteorology is “mass continuity”. If there is persistent and widespread southerly flow over the U.S., there must be (by mass continuity) the same amount of northerly flow elsewhere at the same latitude.

That means that our unusual warmth is matched by unusual coolness someplace else.

Well, guess what? It turns out that our record warm March was ALSO a record for southerly flow, averaged over the U.S. This is shown in the next plot, which comes from about 250 weather stations distributed across the Lower 48 (click for large version; heavy line is trailing 12 month average):

Weather records are broken on occasion, even without global warming. And here we see evidence that our March warmth was simply a chance fluctuation in weather patterns.

If you claim, “Well, maybe global warming caused the extra southerly flow!”, you then are also claiming (through mass continuity) that global warming ALSO caused extra northerly flow (with below normal temperatures) somewhere else.

And no matter what anyone has told you, global warming cannot cause colder than normal weather. It’s not in the physics. The fact that warming has been greatest in the Arctic means that the equator-to-pole temperature contrast has been reduced, which would mean less storminess and less North-South exchange of air masses — not more.

Since NOAA encourages the use the USHCN station network as the official U.S. climate record, I have analyzed the average [(Tmax+Tmin)/2] USHCN version 2 dataset in the same way I analyzed the CRUTem3 and International Surface Hourly (ISH) data.

The main conclusions are:

1) The linear warming trend during 1973-2012 is greatest in USHCN (+0.245 C/decade), followed by CRUTem3 (+0.198 C/decade), then my ISH population density adjusted temperatures (PDAT) as a distant third (+0.013 C/decade)

2) Virtually all of the USHCN warming since 1973 appears to be the result of adjustments NOAA has made to the data, mainly in the 1995-97 timeframe.

3) While there seems to be some residual Urban Heat Island (UHI) effect in the U.S. Midwest, and even some spurious cooling with population density in the Southwest, for all of the 1,200 USHCN stations together there is little correlation between station temperature trends and population density.

4) Despite homogeneity adjustments in the USHCN record to increase agreement between neighboring stations, USHCN trends are actually noisier than what I get using 4x per day ISH temperatures and a simple UHI correction.

The following plot shows 12-month trailing average anomalies for the three different datasets (USHCN, CRUTem3, and ISH PDAT)…note the large differences in computed linear warming trends (click on plots for high res versions):

The next plot shows the differences between my ISH PDAT dataset and the other 2 datasets. I would be interested to hear opinions from others who have analyzed these data which of the adjustments NOAA performs could have caused the large relative warming in the USHCN data during 1995-97:

From reading the USHCN Version 2 description here, it appears there are really only 2 adjustments made in the USHCN Version 2 data which can substantially impact temperature trends: 1) time of observation (TOB) adjustments, and 2) station change point adjustments based upon rather elaborate statistical intercomparisons between neighboring stations. The 2nd of these is supposed to identify and adjust for changes in instrumentation type, instrument relocation, and UHI effects in the data.

We also see in the above plot that the adjustments made in the CRUTem3 and USHCN datasets are quite different after about 1996, although they converge to about the same answer toward the end of the record.

UHI Effects in the USHCN Station Trends
Just as I did for the ISH PDAT data, I correlated USHCN station temperature trends with station location population density. For all ~1,200 stations together, we see little evidence of residual UHI effects:

The results change somewhat, though, when the U.S. is divided into 6 subregions:

Of the 6 subregions, the 2 with the strongest residual effects are 1) the North-Central U.S., with a tendency for higher population stations to warm the most, and 2) the Southwest U.S., with a rather strong cooling effect with increasing population density. As I have previously noted, this could be the effect of people planting vegetation in a region which is naturally arid. One would think this effect would have been picked up by the USHCN homogenization procedure, but apparently not.

Trend Agreement Between Station Pairs

This is where I got quite a surprise. Since the USHCN data have gone through homogeneity adjustments with comparisons to neighboring stations, I fully expected the USHCN trends from neighboring stations to agree better than station trends from my population-adjusted ISH data.

I compared all station pairs within 200 km of each other to get an estimate of their level of agreement in temperature trends. The following 2 plots show the geographic distribution of the ~280 stations in my ISH dataset, and the ~1200 stations in the USHCN dataset:

I took all station pairs within 200 km of each other in each of these datasets, and computed the average absolute difference in temperature trends for the 1973-2012 period across all pairs. The average station separation in the USHCN and ISH PDAT datasets were nearly identical: 133.2 km for the ISH dataset (643 pairs), and 132.4 km for the USHCN dataset (12,453 pairs).

But the ISH trend pairs had about 15% better agreement (avg. absolute trend difference of 0.143 C/decade) than did the USHCN trend pairs (avg. absolute trend difference of 0.167 C/decade).

Given the amount of work NOAA has put into the USHCN dataset to increase the agreement between neighboring stations, I don’t have an explanation for this result. I have to wonder whether their adjustment procedures added more spurious effects than they removed, at least as far as their impact on temperature trends goes.

And I must admit that those adjustments constituting virtually all of the warming signal in the last 40 years is disconcerting. When “global warming” only shows up after the data are adjusted, one can understand why so many people are suspicious of the adjustments.

UPDATE: I’ve added 6 regional U.S temperature plots, for the Northwest, North Central, Northeast, Southwest, South Central, and Southeast.

Thanks for the comments and suggestions on yesterday’s post introducing a new U.S. population-density adjusted temperature (PDAT) dataset. As a result of the comments, I have stratified the U.S. into 6 subregions, and performed station temperature trend vs. population regressions, rather than just lumping all 280 stations together. This should help reduce the effect of any fortuitous (but real) regional warming which just happens to be located where more people live. It should better isolate the true urban heat island (UHI) effect on temperature trends.

The results are shown in the following plot, with the regional regression coefficients listed being the scaling factor between station temperature trend (deg. C/decade) and population density (persons per sq. km) to the 0.2 power (click for high res. version):

As can be seen, 4 of the 6 regions have quite strong dependence of the trends on population density. Only the Southwest U.S. has cooling with increasing population density, which is probably the result of people planting vegetation in what is mostly an arid region to begin with.

The impact of making regional — rather than whole U.S. — population adjustments on the U.S. average temperature variations results in only a slight increase in the resulting temperature trend I posted yesterday, which is still well below that computed from the CRUTem3 dataset (click for high res. version):

Of course, the regional trends would change substantially, since now I am actually warming the Soutwest U.S. temperatures with time, based upon station population density. But the Southeast U.S. trends will be cooled even more than before, because of the strong relationship between temperature trend and population density found there (see 1st plot, above).

UPDATE: Here are the population-adjusted temperature variations for the 3 northern U.S. sectors, with just the trailing 12-month averages plotted to reduce the messiness:

…and here are the 3 southern U.S. sectors:

The bottom line is that there is still clear evidence of an urban heat island effect on temperature trends in the U.S. surface station network. Now, I should point out that most of these are not co-op stations, but National Weather Service and FAA stations. How these results might compare to the GHCN network of stations used by NOAA for climate monitoring over the U.SA., I have no idea at this point.

Also, I need to clear up a misconception…the adjustments I perform do not remove the trends in the data. They remove only the component of the trend which is due to population density, using the regression coefficient alone (not the regression constant). There are no adjustments in January 1973 (the beginning of my data record), and then the adjustments increase linearly with time.

This is the first of what I hope will be monthly surface temperature updates for the contiguous U.S., based upon 280 International Surface Hourly (ISH) stations which have reasonably complete temperature records since 1973.

Following up on my previous post showing that ISH station warming trends during 1973-2011 were a function of population density, I have quantified the average temperature trend increase with population density (2000 population data) over the U.S, then applied a linear trend correction to each of the stations based upon that relationship.

A few of the findings:

1) Essentially all of the +0.20 deg. C/decade average warming trend over the U.S. in the last 40 years computed from the CRUTem3 dataset (which the IPCC relies upon for its official global warming pronouncements) evaporates after population adjustment (no claim is made for countries other than the U.S.)

2) Even without any adjustments, the ISH data have a 20% lower warming trend than the CRUTem3 data, a curious result since the CRUTem3 dataset is supposedly adjusted for urban heat island effects.

3) The only calendar month with obvious long-term warming is January, due to unusually cold U.S. winters during the 1970s.

4) Last month (March, 2012) is the second warmest monthly temperature anomaly in the 40 year record, and easily the warmest March, even after population adjustment.

For the time being, I’ve decided to post the results for comment rather than attempt to get the work published, which would be a much bigger effort. My hope is that the new dataset will stimulate debate in the climate research community over the existence of residual urban heat island (UHI) effects causing a spurious warming component in commonly used temperature datasets.

Unadjusted ISH Temperature Data vs. CRUTem3 Over the U.S.

As discussed in my previous post, the raw data come from the International Surface Hourly (ISH) database which is continuously updated at NCDC. I average the 4 synoptic reporting times (00, 06, 12, and 18 UTC) together to get a daily average temperature for each station. These are the most often reported times of day in the record, and using them alone maximizes the number of stations available for analysis while at the same time providing (what I believe to be) a more physically meaningful “daily” average than maximum and minimum temperatures do.

At least 80% of the daily data must be present to compute an average for a month from each station, and at least 90% of the months during 1973-present must also be available, as well as ALL calendar months from 1973 and 2011. Nominally 280 stations in the U.S. meet this requirement, a number which does not change substantially throughout the 40 year record.

When monthly anomalies (relative to 1973-2011) are computed in 5 deg. lat/lon grids covering the contiguous U.S. from both the CRUTem3 dataset and from the unadjusted data from 280 stations, here are the resulting monthly variations during 1973 through March, 2012 for ISH and through January 2012 for CRUTem3 (all images can be clicked to see the large, detailed versions):

The monthly correlation between the two time series in the above plot is 0.994. Curiously, even without any adjustments to the ISH data, the resulting ISH linear warming trend (+0.157 deg. C/decade) is about 20% lower than the CRUTem3 trend (+0.198 deg. C/decade).

If we difference the two time series, we get this (click for full res version):

There are a couple things to note. First, we see that the excess warming in CRUTem3 versus unadjusted ISH data is growing with time.

Secondly, there is some evidence of artifacts which are likely from the CRUTem3 dataset, such as a sudden downward adjustment starting about November, 1996. My understanding is that the CRUTem3 dataset has a station distribution which changes over time. Also, I believe there are adjustments made to the data from individual stations. In the ISH dataset, however, we have 280 stations with essentially complete data from beginning to end of the record, with no adjustments; it is difficult to see how such a jump could have arisen from the ISH data.

The Population Density Adjustment

Linear temperature trends computed from each of the 280 stations reveal a dependence on population density. Just as has been found in previous studies based upon spatial temperature patterns (cities being warmer than the surrounding countryside), we find that the warming trend with time increases rapidly with population at low population densities, then levels off at high population densities.

This nonlinear relationship is found here to go as population density (PD) raised to the 0.2. power (warming ~ PD^0.2). The following plot shows the results for all stations individually, as well as for averages in 4 population subgroups (click for full res version):

As seen in the above pair of plots, essentially the same regression coefficient is computed whether I use all stations individually, or average them into 4 population subgroups. The standard error of the regression coefficient is +/- 20%, which should give some idea of the statistical uncertainty in the population-based adjustments to the temperature data shown below.

Significantly, I will assume that this average relationship between temperature trend and population density is entirely due to the urban heat island effect, and remove it from each station. Of course, not all stations would have a UHI effect, but others will have a strong effect. The above plots’ regression lines show the average relationship across all stations, which I simply remove from all stations. This avoids qualitative decisions about individual stations’ histories, which would be difficult to reproduce by other investigators, and keeps the methodology simple.

Since (as we will see) this adjustment removes most of the warming trend in the U.S. since 1973, it will be the most criticized. It will be claimed that the warming trends are indeed real, and that it must be by coincidence that the most populated regions of the country have also warmed the most.

But that claim has no independent evidence, other than the thermometer data. It has no more support than my claim that the warming dependence on population is spurious, due to the UHI effect.

In fact, I think it has less support. We know based upon many published studies that the UHI effect is real, at least in spatial terms (cities average warmer than the surrounding countryside). The above plots show a similar effect on temperature trends, with a nonlinear functional dependence approximately like that seen in the spatial dependence found by other investigators. That this effect would be fortuitous seems to stretch credulity.

Results with Population Density Adjustment

The regression coefficient from the above plot was used to make a linear temperature trend adjustment to the ISH temperatures, starting with zero adjustment in January 1973. The resulting plot, analogous to the very first one above, for U.S. temperature variations since 1973 is shown next (click for full res. version):

Significant, the population adjustment erases essentially all of the U.S. warming over the last 40 years. Nevertheless, last month (March, 2012) is seen to be the 2nd warmest month in the 40 year record, and (as we will see) easily the warmest March.

The corresponding difference plot between the two datasets shows what I am interpreting to be considerable spurious warming in the CRUTem3 dataset:

U.S Temperature Variations, 1973-2012, by Calendar Month

When we examine the seasonal dependence of U.S. temperature changes over the last 40 years, we find that the only month with significant warming is January, and even that is only because there were so many cold Januaries in the late 1970s and early 1980s. The other months are essentially flat. Plots for individual months are shown next, and note that the January, February, and March plots end in 2012, while the others end in 2011 (click for full res. versions):

Conclusions

I am quite surprised that, even without any adjustments, the ISH data show 20% less U.S. warming than the CRUTem3 data over the 1973-2011 period. Since the CRUTem3 data are supposedly adjusted for urban heat island (UHI) effects, this seems quite curious, to say the least.

When the ISH temperature data are corrected for the average warming bias — shown here to be a function of population density — it essentially erases 40 years of U.S. warming: from +0.20 deg C/decade in the IPCC-blessed CRUTem3 dataset, to +0.01 deg. C/decade. For those interested in statistical uncertainties, the standard error in the regression coefficient I used would amount to about +/-20% uncertainty in the reduction in the warming trend.

The warmth of March, 2012 is indeed anomalous, at least in the context of the last 40 years. But as the plot for all March’s (above) shows, one month does not a warming trend make. 🙂

UPDATE #1: The Recent Warm Winter
Regarding the recent winter, if we plot trailing 3-month averages of the population-adjusted temperatures, we see that January through March of this year (2012) was the warmest 3 month period of the 40-year record. Of course, the warmest 3 months must occur at some point in the record, and since there is no long-term trend in the data, I would wager that it is a temporary blip, rather than a sudden shift into a new climate regime:

UPDATE #2: Why the Discrepancy with UAH LT Temperatures?
It has been pointed out that our UAH LT (tropospheric temperature) product has a warming trend for 1979-2011 of about +0.20 deg. C, so why the difference with my near-zero surface temperature trend (which is near-zero whether you start in 1973, or 1979)?

The monthly correlation between the two datasets is 0.87, so there is reasonably good agreement on that time scale, but a time series plot of their difference suggests some sort of step jump in 1995:
.
The direction of the change would be either spurious warming in UAH LT or spurious cooling in the ISH PDAT surface temperature data. The plot really doesn’t look like the CRUTem3 -minus- ISH PDAT plot (reproduced below), so I don’t have a ready explanation for it:

Now, 1995 happens to be when the NOAA-11 satellite was replaced by NOAA-14, and those two satellites had to be intercalibrated with NOAA-12, which was going through its own diurnal drift. So, there might be a diurnal drift issue here that has not been sufficiently accounted for. Maybe our new (but unfinished) diurnal adjustement strategy for Version 6 of the UAH dataset will shed light on this.

Of course, it is always possible that a weather regime change around that time led to a change in the tropospheric temperature lapse rate, but that is pure speculation on my part.

The global average lower tropospheric temperature anomaly jumped up in March, 2012, to +0.11 deg. C. as La Nina conditions in the Pacific Ocean waned (click on the image for the full-size version):

The 3rd order polynomial fit to the data (courtesy of Excel) is for entertainment purposes only, and should not be construed as having any predictive value whatsoever.

Here are the monthly stats:

YR MON GLOBAL NH SH TROPICS
2011 1 -0.010 -0.055 +0.036 -0.372
2011 2 -0.020 -0.042 +0.002 -0.348
2011 3 -0.101 -0.073 -0.128 -0.342
2011 4 +0.117 +0.195 +0.039 -0.229
2011 5 +0.133 +0.145 +0.121 -0.043
2011 6 +0.315 +0.379 +0.250 +0.233
2011 7 +0.374 +0.344 +0.404 +0.204
2011 8 +0.327 +0.321 +0.332 +0.155
2011 9 +0.289 +0.304 +0.274 +0.178
2011 10 +0.116 +0.169 +0.062 -0.054
2011 11 +0.123 +0.075 +0.170 +0.024
2011 12 +0.126 +0.197 +0.055 +0.041
2012 01 -0.090 -0.057 -0.123 -0.138
2012 02 -0.112 -0.013 -0.212 -0.277
2012 03 +0.108 +0.128 +0.089 -0.108

As a reminder, the most common reason for large month-to-month swings in global average temperature is small fluctuations in the rate of convective overturning of the troposphere, discussed here.