I analyzed a new global gridded daily weather dataset to see what it would tell us about what climate change, particularly temperature change, we may have already witnessed during my lifetime. What I discovered surprised me very much. I was expecting that some places would show a 1-degree Celsius temperature increases over a 30-year period, with a few additional pixels showing up to a 2-degree increase. Instead, we see a lot of places with 1- and 2-degree increases, and even others with 3- and 4-degree temperature changes.
The dataset I used, AgMERRA, was recently developed by Alex Ruane of NASA and Richard Goldberg of the University of Chicago. The pixels are a half-degree in size, and the data spans the period of 1980 to 2010. The data includes solar radiation, minimum temperature, maximum temperature, rainfall, and wind speed. My primary interest is in the monthly mean daily maximum temperature for the warmest month of the year (for brevity, “tmax”), since this may be the critical limiting factor for agriculture in the future under climate change.
My idea was simply to run a regression for each pixel on tmax using an intercept and the year as explanatory variables. The year parameter would be a measure of the year-to-year change in tmax. I would be able to tell the statistical significance at each point by looking at the z-statistic for the parameter from the regression. Being somewhat a skeptic, my guess was that only a few pixels — maybe less than 10 percent of them — would have temperature changes that would be significant.
Given there has been a global “pause” in warming of surface temperatures for the last decade, I considered cutting off the regression to 2000 or maybe 2004, but in the end decided that the best first pass at the data would be to use it all. If there were a pause, this would have the tendency to make the z-statistic less significant if we estimate it to 2010.
In the map of temperature changes presented here, we show the estimated 30-year change. That is, given the parameter estimated at each point, what did that imply for temperature changes over a 30-year period? By using the parameter times 30 to show change, we effectively smooth out the bumpiness of year-to-year variation. In the map, I only show results that are statistically significant at the 10 percent level. However, I noted that over 60 percent of the pixels that are significant at 10 percent are also significant at 1 percent.
The results are quite dramatic, showing how much change has already taken place. For me, this is a wake-up call. I have been thinking about climate change as something in the future, but something we get hints of now. But these fairly large temperature increases that have already happened in many parts of the world show that hundreds of millions of people are clearly dealing with climate change already. Temperature changes in this range are large enough that they are probably already adversely impacting crop yields.
For example, the map shows a fairly large area in Brazil that has a more than 3-degree Celsius rise in temperature over the 1980 to 2010 period. One of the reasons I took time to explore this data in the first place was to use it to prove that one of the climate models I was using for my research was not believable, since it projected a 7 degree Celsius increase for the period from 2000 to 2050 in the same part of Brazil. In fact, the results that we see in the map suggest that the particular climate model I questioned is likely to be correct, or at least not terribly incorrect, at least for Brazil.
In the global scheme, we see large temperature increases in central Africa and the Horn, northern Canada, eastern Europe, western Asia, Mongolia, northeast China, and parts of Russia. And we actually note falling temperatures in northern Australia, New Guinea, and southern India. It is also important where temperature change has not occurred to a large extent, including the eastern and central United States, southern Canada, and much of Latin America, southern Africa, and south and southeast Asia. For my friends in New Zealand (not shown in the map), the dataset only covered the very northernmost portion of the country; and for my Alaskan friends (also not shown), only 2 small patches of temperature rise were noted.
I would like to emphasize that these are preliminary results that I have not yet reviewed thoroughly, and that no one else has reviewed either. Therefore, they are meant to stimulate discussion and to promote follow-up work. But if confirmed, they clearly say something to policymakers, to each person working in a field related to climate change, and to climate change skeptics.
— Timothy S. Thomas
Addendum, June 18,2014: I corresponded with one of the creators of the dataset, and he suggested that the dataset was not the best to use for trend analysis, mostly because of the change in number of weather stations. Here is a quick example of how this could create a false trend. First, recognize that sometimes weather stations are hundreds of kilometers apart. Let us consider such a case in which two stations are 400 km apart. The temperatures on the grid between the two weather stations are primarily influenced by those two stations, and for simplicity, we might assume that the values are a weighted average of those two station, with the weight being higher for the closest station (perhaps inverse distance weighting). Now, if we added a weather station halfway through the analysis period and directly in between the two stations, values near the middle of the original two stations will be strongly influenced by the new station. If the new station has temperatures two degrees above what would have been predicted at that point using the original two stations, it will appear that there was an increase in temperature at that point (and nearby points), whereas all that happened was that a new station was added that had measured values different (in this case, higher) than the values that had previously been predicted.
In all fairness, the same type of reasoning could remove trends. Let’s say that the original two points were both getting warmer. If the new weather station had measured values lower than predicted, then it would tend to wipe out the trend, even though the trend was real, at least near the two original weather stations.
The creator of the dataset suggested that I think of using weather station data for trend analysis. I hope to do this soon, to see if the weather station data tends to support the analysis I did here with the AgMERRA data.
Follow up article that used a different dataset to confirm these results.
Rainfall analysis using the AgMERRA dataset.
 The climate model I was referring to is one of IPCC’s AR5 (5th Assessment Report, published in 2013) GCMs (i.e., climate models) — the MIROC-ESM-CHEM — under RCP8.5 (an assumption of high greenhouse gas emissions). Another climate model, HADGEM2-ES, had a similar result in the same part of Brazil, though with only around a 4- or 5-degree Celsius increase.
References (as given by the authors who permitted me to use the data):
Ruane, A.C., and R. Goldberg, AgMIP Hybrid Baseline Climate Datasets: Shifted Reanalyses for Gap-filling and Historical Climate Series Estimation (in preparation).
(Tim’s note: I am trying to clarify this, but I believe an even more recent reference and site for downloading the data is at http://data.giss.nasa.gov/impacts/agmipcf/agmerra/).
Elliott, J., M. Glotter, N. Best, J. Chryssanthacopoulos, D. Kelly, M. Wilde, and I. Foster, (2013). The Parallel System for Integrating Impact Models and Sectors (pSIMS). Prepared for a special issue of Environmental Modeling and Software: Agricultural systems modeling & software (in review).
Elliott, J., M. Glotter, N. Best, D. Kelly, M. Wilde, and I. Foster, (2013). The Parallel System for Integrating Impact Models and Sectors (pSIMS). Proceedings of the 2013 XSEDE Conference, in press.