Over the course of the past few years I have taking a look at cooling stations a few times but never in much depth. A few other people have looked at them, but usually without much rigor. The standard approach is to find some cooling stations and then conclude that somehow global warming is thereby challenged as a theory. I supp that argument will come up in the comments, but my real goal here is to shed some light of the phenomena of stations that have long term cooling trends. The dataset I used here is the Berkeley Earth Data set version 2.0. I used Single value, with Quality control and seasonality left in the dataset.
The data selection process is pretty simple. We read the data in, window it to 1900 to 2011, and remove those stations that have no data
Data <- readAsArray(Directory = choose.dir())
Data <- windowArray(Data, start = 1900, end = 2011.99)
Data <- removeNaStations(Data)
In the next step I take notice of something that Carrick suggested over at Lucia’s and I focus on stations that have a high percentage of complete data. For this exercise I slected those records that had 99% of the data present for the time period. That’s 1332 months of data in this 112 year period (1344 months)
rsum <- rowSums(!is.na(Data))
dex<- which(rsum > 1331)
Data <- Data[dex, , ]
Next we align the station inventory with the data, do inverse weighting and calculate the least squares fit
Data <- intersectInvData(Stations,Data)
weights <- inverseDensity(Data$Inventory,Data$Array)
Temps <- solveTemperature(Data$Array,weights)
TIME <- seq(1900,2011)
plot(ts(Temps, start =1900, frequency = 1))
abline(reg = lm(Temps~TIME))
What we see should come as no surprise. The planet is getting warmer. The slope of the line is roughly .08C/ decade. The stations have the following geographical distribution:
The sum total of all stations is 478. The result, as we expect, is similar to the results we get if we use all stations, regardless of their length. However, we can note that the sample isn’t very representative so their could be bias due to spatial sampling. The overweighting of the northern hemisphere is likely to cause a trend that is slightly higher than a what we would see if the sample included more SH stations.
The next step I will merely describe in words because I will probably change the code I used here when I decide upon a final approach. After calculating the global average of all these long stations I then created trends and confidence intervals for every station. For trends I used a Theil-Sen estimator which may be a bit different than most folks are used to. Its described here: http://en.wikipedia.org/wiki/Theil%E2%80%93Sen_estimator. The zyp package has a quick implementation that spits out confidence intervals as well. I’ll play around with some of Lucia’s approaches and maybe look at Tamino’s approach, but for now, this will do.
Summarizing the trends we see the following:
summary(DF$Slope) Min. 1st Qu. Median Mean 3rd Qu. Max. -0.014450 0.001390 0.006038 0.006151 0.010850 0.036920
Note that the simple meanof all slopes is less than the area weighted global solution. And note that some of the slopes of individual stations are negative. These are so called cooling stations: A crude map below gives you an indication of their locations. Blue is cooling, red is warming. The warming were drawn first so they tend to get covered up.
Well there you have it. Some long stations cool. This is fairly well known but a week doesnt pass on the internet without somebody asking about it or pointing to it as proof of “something.” Before we dive into that there is a bit more work to do.
In order to look a bit deeper into the “cooling stations” one beneficial cut to make through the data is to isolate those stations that are coastal versus those that are inland. That’s interesting in and of itself. The effects of geography on climate are well known. Coastal stations will have a fewer extremes. In mathematical terms the standard deviation of monthly temperatures will be lower. Defining a range for coastal stations is somewhat problematic so I did some sensitivity analysis around that looking at the standard deviation as a function of range from the ocean. More on that in the future but for now Here are the results at 50km
And now for inland stations
An interesting question this raises is could their be a second process at work with a slightly negative mean trend. The first step in looking into that will be separating those stations which have a statistically significant cooling trend from those that do not. That is, before we set out looking for a “cause” behind the cooling stations it’s probably a wise move to look at those where the cooling trend is statistically significant. Of the 478 stations, 54 show a rate of cooling that is statistically significant (95%)
With these 54 stations and the metadata we have we can start to eliminate potential causes of the cooling. One of the first thoughts that occurs to people is “de urbanization” This has been raised several times over the past 5 years I’ve spent looking at the problem. In Our AGU poster I spent a short period of time looking at “de urbanization” and frankly didn’t find much. The number of stations that had decreases in population was very small. Nevertheless, we can look at it again very quickly for these 54 stations:
First, the historgram of year 2000 population density ( people per sq km), second the histogram for 1900, and then finally the histogram for the difference between 2000 and 1900:
As the chart shows there is only one station that had a decrease in population. That station started with 16 people per sq km in 1900 and slowly over time reduced to 14 people per sq km. Lets look at some of its metadata
Slope Upper Lower Id Lat Lon Altitude Months Dst 31137 -0.0088867 -0.005273885 -0.01257143 38048 42.3978 -90.7036 325.55 1343 1212 Sdev Anthrome2000 Anthrome1900 Lcover UrbArea Slope2 Upper2 31137 2.332492 32 32 14 0.0004538235 -0.008970948 -0.005294326 Lower2 popd_1900AD popd_1910AD popd_1920AD popd_1930AD popd_1940AD popd_1950AD 31137 -0.01267606 16.7313 15.5423 15.7447 15.2845 14.9356 15.4139 popd_1960AD popd_1970AD popd_1980AD popd_1990AD popd_2000AD 31137 15.5131 15.3346 15.1406 14.1625 14.3402
The station is 1212 km away from a body of water; and we can see that its standard deviation is high. Stations near the coast have standard deviations close to 1.4. A standard deviation of 2.33, in fact, is quite high in the last quartile of all 487 stations. Other things to note. The anthrome has not changed. Anthrome 32 is residential rainfed croplands. Lcover 14 also indicates rainfed croplands ( 300 meter data). My UrbArea paramater captures the amount of “built” area within 10km, so it looks like there is some human structure in the area. In addition to this metadata, my other sources indicate an airport location nearby, so it looks like it would be an airport built in a farming community. Checking with google earth. There you go:
So, how do we explain the cooling at a station that is located at an airport? Without looking at the station history in detail, you can probably figure that the station has a station move in its history. That move may be documented or undocumented. If there is a move, that might show up in the temperature record as a discontinuity. This is a good lesson in using Berkeley Earth Surface Data. The Single value data has not been through the scalpel. The scalpel performs the analysis to find structural breaks. Series are not homogenized or adjusted, but if there is a break then a single series is replaced with two series. Let’s see what this series looks like:
Looking at the chart the step down post 1950 looks like it may be the cause of the cooling. 1950 would be the start date of the dubuque regional airport series. When I get some time I’ll load up structchange and see if it can find the discontinuity. We can continue this process for the other 53 stations, but the speculation that all the cooling is due to “de population” really doesnt have much merit. The actual number of stations that have suffered de population or de urbanization is small. The more likely explanation for the coolings comes down to undocumented or documented station changes. If you spend some time looking through all the monthly series you’ll begin to see some patterns that usually turn out to be either station moves or station splices. At some point when the “post scalpel” data is released this will be a bit easier to understand.