Close window
Close window
This section of the new site isn't ready yet. We've brought you back to the current site.

An overview of global annual average temperature anomalies in 2016

Figure 1: (top) Global annual average temperature anomalies (relative to the long-term average for 1961-1990). Six datasets are shown as indicated in the legend; (bottom) difference from the HadCRUT4 data set on an expanded scale (difference from ERA-Interim not shown due to offset).

January 2017 - Global temperature data from centres around the world show that 2015 and 2016 are clearly the warmest on record with 2016 likely the warmer of the two.


At the end of each year, there is usually interest in where the year sits amongst the global temperature rankings. Sometimes, the answer is straightforward: 2015 was indubitably the warmest year on record. And sometimes, it is not: 2014 was a serious contender for warmest year, but it was not possible to say for certain whether it was or not, although it was clearly one of the top ten.

As the preliminary numbers for the whole of 2016 are finally released, it is a good time to reflect on how the information from different global data centres fits together and what the available evidence tells us about the rankings for the year.

2015 and 2016 saw weak El Niño conditions develop into a strong El Niño event which has now dissipated. El Niño events typically drive a short-lived spike in global temperatures, with the spike usually lagging a little behind the temperature variations in the tropical Pacific. The strong El Niño is the main reason why 2015 and 2016 were warmer than 2014, but anthropogenic influences are necessary to explain the majority of the overall warming since the 1950s.

However, in order to better understand the relative warmth of 2015 and 2016, we need to dig a little deeper into what happened in the global climate over the past two years, how global temperatures are estimated and how the two interact.

Estimating global average temperature

Surface temperature data are gathered by a global network of weather stations, ships and buoys. This network measures air temperatures over land and sea-surface temperatures over the oceans. A number of groups use these data to produce global temperature data sets. The data are carefully processed to account for changing measurement methods and instrumentation, and for the uneven distribution of measurements around the world. The methods that the groups use are different and so the estimates of the global average that they produce are also slightly different. In addition to traditional data sets based purely on surface temperature measurements, there are atmospheric reanalyses, which use a much wider range of observations including satellite data in combination with a weather-forecasting model to produce a globally complete temperature analysis.

Figure 1 shows the estimated global annual average temperature differences from the long-term average, as calculated from six data sets: HadCRUT4 (produced by the Met Office Hadley Centre in collaboration with the Climatic Research Unit at the University of East Anglia), GISTEMP (produced by the Goddard Institute of Space Studies at NASA), NOAAGlobalTemp (produced by the National Centers for Environmental Information at NOAA), Berkeley Earth and Cowtan and Way (CW). It also shows (dotted and at an offset because it begins in 1979) a global temperature series calculated from the ECMWF ERA-Interim reanalysis.

Figure 1: (top) Global annual average temperature anomalies (relative to the long-term average for 1961-1990). Six datasets are shown as indicated in the legend; (bottom) difference from the HadCRUT4 data set on an expanded scale (difference from ERA-Interim not shown due to offset).

In general, the agreement between the data sets is very good. They all display a similar increase in estimated global average temperature over time and the short up ticks and drops in global temperature of around 0.1 to 0.2°C that signify the pulse of El Niño and La Niña can be clearly seen in each of the data sets. However, some differences are to be noted, particularly in the 19th Century where data are sparse and in the mid 20th Century when there were large changes to the way that sea-surface temperatures were measured.

However, we are here focusing on differences in 2015 and 2016 where the data sets fall roughly into two groups. One group, formed of HadCRUT4 and NOAAGlobalTemp shows 2016 as only marginally warmer than 2015. The other group, comprising GISTEMP, Berkeley, CW and ERA-Interim, has 2016 as quite a bit warmer than 2015.

Why the differences?

One major difference between the data sets is the way they deal with geographically-uneven sampling. It leads to a difference in the degree to which they attempt to fill gaps in the station network using information from nearby stations, as illustrated in Figure 2. All the data sets perform some degree of interpolation, but the GISTEMP, Berkeley and CW data sets – in fact those which have 2016 as much warmer than 2015 – do the most interpolation. On the other hand, the HadCRUT4 and NOAAGlobalTemp estimates (in which 2016 is warmer than 2015 by a smaller margin) do less interpolation and, most importantly for the discussion here, they do not interpolate extensively into the polar regions which were particularly warm in November 2016.

Figure 2: Surface temperature anomalies for November 2016 (°C, relative to the 1961-1990 average) from the five traditional global temperature data sets showing the difference in coverage: HadCRUT4, Cowtan and Way, NOAAGlobalTemp, Berkeley Earth and GISTEMP. Colour scales are shown beneath each diagram. The outer edge of the colours scales are ±10°C unless the highest or lowest anomalies for the month lie outside that range

In fact, in 2016, the Arctic was particularly warm throughout the year. Figure 3 shows that temperature anomalies for areas north of the Arctic Circle (approximately 65°N) were nominally the highest on record, with temperatures around 2.0 to 3.5°C above the 1961-1990 average. The Arctic average was considerably higher than the average for the rest of the globe so the data sets with sparser coverage in the Arctic will – all else being equal – tend to underestimate the global average temperature for 2016.

Figure 3: Annual average temperature anomalies (relative to the 1961-1990 average, 2016 is January to November average) for the polar regions: (top) Arctic (North of 65°N) and (bottom) the Antarctic (South of 60°S). Note different temperature scales. Here we have adjusted the anomalies from ERA-Interim to be comparable to those of the other data sets during the 1981-2010 period. The Antarctic series starts in 1958, which is the first year of long-term routine measurements in continental interior.

In October and November 2016, temperature anomalies for the month exceeded 10°C in parts of the Arctic. The unusual Arctic warmth contrasted with large negative temperature anomalies across Asia. This combination was sampled differently by the data sets and it led to a large divergence in estimates of the global average in October and November (Figure 4). Over the same period, the Antarctic was also warmer than average. During October, Arctic sea ice extent recovered slowly from the annual minimum and the monthly extent was the lowest on record. In November, sea ice extent in both the Arctic and Antarctic were at record low levels.

Figure 4: Monthly global average temperature anomalies (relative to the long-term average for 1961-1990). Five datasets are shown as indicated in the legend.

The Arctic was also warmer-than-average in 2015, but the average anomaly here was much lower than in 2016 (by around 1°C). In addition, in the Southern Hemisphere, a strongly persistent pressure pattern of below-average pressure over Antarctica and above-average pressure at lower latitudes  – which defines the positive phase of the Antarctic Oscillation – held sway over the continent during much of 2015. During the positive phase of the Antarctic Oscillation there is a tendency for temperatures across the continent to be below average. From January to September 2015, average temperatures across Antarctica were below the long-term mean, somewhat balancing out higher-than-average Arctic temperatures in the more heavily interpolated data sets. The positive phase continued through much of 2016, but the associated modest cooling across Antarctica was more than balanced out by the extreme warmth in the Arctic.

How much difference does this make?

We can explore the size of the “coverage effect” by successively reducing the coverage of all the data sets to that of HadCRUT4, which performs the least extensive interpolation. In the top panel of Figure 5, we display the global average anomalies estimated from each of the data sets after first converting them to the same grid as HadCRUT4 – the same basic pattern can be observed as in Figure 1. Differences between global averages calculated from the data sets and HadCRUT4 exceed 0.05°C (the shaded region shows the ±0.05°C range) in a number of years, including 2016.

Figure 5: (a) Global average temperature anomalies for the five global temperature data sets at full coverage (HadCRUT4, Cowtan and Way, GISTEMP, NOAAGlobalTemp, Berkeley Earth) relative to the 1961-1990 average. Colours are as in Figure 1. (b) Differences between individual data sets and HadCRUT4. The grey shading indicates the ±0.05°C range in the difference plots. (c) Global average temperature anomalies calculated from data sets matched to HadCRUT4 coverage in the polar regions. (d) Differences between individual data sets and HadCRUT4 after data matched to HadCRUT4 coverage in polar regions. (e) Global average temperature anomalies calculated from data sets matched to HadCRUT4 coverage everywhere. (f) Differences between individual data sets and HadCRUT4 after data matched to HadCRUT4 coverage everywhere. 2016 data are based on January to November.

Next, we reduce the coverage of each of the datasets to be the same as HadCRUT4 north of 60°N and south of 60°S (Figures 5c) This has the effect of reducing the differences (Figure 5d) such that even the largest differences between the five traditional data sets do not exceed 0.05°C. Finally, we reduce the coverage to be equal to that of HadCRUT4 across the whole globe (Figures 5(e) and (f)) which makes only a small additional difference. The residual differences when all the data sets are reduced to their common coverage are around one to four hundredths of a degree, thus indicating excellent agreement between them where they overlap.

The small residual differences reflect the net effects of differing input data, and analysis choices. One recently highlighted difference between the ways the data sets are processed is the choice of how they deal with changes in the marine observing system. HadCRUT4, CW and Berkeley handle it one way (using a method developed at the Met Office) and NOAAGlobalTemp and GISTEMP use another (developed at NOAA). Differences between sea surface temperature data sets are around a few hundredths of a degree when averaged over the global oceans, which is consistent with the size of the residual differences seen between the data sets here. Another difference to be taken into account is the choice of whether to use air temperatures or sea surface temperatures in ocean areas covered by sea ice.

An alternative way to think about the differences is to look at the uncertainty estimates provided with the HadCRUT4 data set (Figure 1). A large component of the uncertainty in the annual average is associated with limited coverage. Other uncertainties associated with local sampling and measurement limitations amount to around four hundredths of a degree, which is the expected size of the residual differences between data sets when disparities in coverage have been accounted for.


Despite the processing differences, the generally good agreement between the data sets demonstrates that estimated long-term global surface temperature change does not depend strongly on the methods used and therefore may be considered a robust scientific result.

2015 and 2016 stand out, in all analyses, as the two warmest years on record. The datasets which have the most complete global coverage indicate that 2016 was warmer than 2015 globally. Although some data sets do not provide temperature information for all areas of the globe, there is good agreement between the data sets in the areas where they do overlap. Differences between the data sets in 2015 and 2016 are mostly due to how data gaps are dealt with in data sparse regions, particularly the Arctic where 2016 was nominally a record warm year in all data sets.

Last updated:

Follow us on

Facebook, Twitter, YouTube, Instagram, Snapchat, or LinkedIn Facebook Follow @metoffice on Twitter YouTube Instagram Snapchat LinkedIn