An anthropogenic climate change signal has been identified in the recent warming of global scale surface temperatures (Stott et al., 2001; Tett et al., 2002) and continental average temperatures (Karoly et al., 2003; Stott, 2003; Zwiers and Zhang, 2003). However, the magnitude of natural climate variability increases relative to any greenhouse-gas-induced warming signal as the area is reduced over which temperatures are averaged, making it more difficult to identify anthropogenic warming at regional scales.
The observational record of Central England temperature (CET) is the longest continuous instrumental surface temperature series available, extending from 1659 to the present (Manley, 1974; Parker et al., 1992; Parker and Horton, 2005). Here, we seek to identify an anthropogenic climate change signal for the first time at a small regional scale, in CET, which has shown a marked annual-mean warming of about 1.0 °C since 1960. We assess whether this warming can be explained by natural internal climate variations and the likely causes of this warming using HadCM3, a coupled ocean-atmosphere general circulation model (Johns et al., 2003). Temperature from a single grid box of size 2.5° latitude by 3.75° longitude over England is used as the model estimate of CET.
A critical issue in the detection and attribution of climate change is that climate model simulations provide a reliable estimate of the unforced natural variability of the climate system at multi-decadal time scales. However, the observational record for temperature is generally too short to provide an estimate of natural temperature variability at 30-year or 50-year time scales and hence to be used to evaluate model simulations of variability at those time scales. The CET record is the longest available instrumental record in the world and is used to demonstrate that HadCM3 can reliably simulate natural climate variability at interannual to 50-year time scales.
In the next sections, the observational record for CET and the climate model data are described briefly. Then the simulated variability of CET is compared with the observed variability at interannual, decadal and 50-year timescales. Finally, the observed warming in CET over the last 50 years is compared with the model-simulated natural internal variations and the model response to changes in anthropogenic and natural external forcing.
2. Observed Central England temperature
The CET record is based on instrument observations at several sites in the midlands of England since 1659 and is the longest available instrumental record of surface air temperature in the world (Manley, 1974; Parker et al., 1992; Parker and Horton, 2005). It is representative of a roughly triangular area of the United Kingdom enclosed by Bristol, Lancashire and London. CET is currently calculated as the average of the surface air temperatures observed at three stations; Pershore, Rothamstead and Stonyhurst, as shown in Figure 1. These stations are chosen to correspond with those used historically from the UK surface station network. The data have been adjusted to ensure consistency with the historical series. Since 1974, the data have been adjusted by 0.1–0.3 °C to allow for urban warming (Parker and Horton, 2005). Monthly mean CET data were obtained from the Hadley Centre for Climate Prediction and Research, UK Met Office. From 1700, the monthly mean values are available in tenths of a degree and are considered to be more accurate. The CET observational data have been used from 1700.
3. HadCM3 model
The HadCM3 climate model (Johns et al., 2003) is a coupled ocean-atmosphere general circulation model with horizontal resolution of 2.5° latitude by 3.75° longitude in the atmosphere and 1.25° latitude by 1.25° longitude in the ocean. The vertical resolution is 19 levels in the atmosphere and 20 levels in the ocean. It includes representations of important physical processes in the atmosphere and the ocean, as well as sea-ice and land-surface processes. It maintains a stable global-mean climate when external forcings are not varied. The surface air temperature from a single model grid box is used to represent CET in the model, as shown in Figure 1. The variations of the model temperatures at interannual and decadal time scales in all 8 surrounding grid boxes shown in Figure 1 have correlations greater than 0.85 with the temperature variations in the CET grid box, associated with the large spatial coherence of temperature variations at these time scales. This specific grid box is chosen to represent CET because it has a larger fraction of land than the neighbouring grid box to its west.
Constant external forcing simulations (control runs) allow the estimation of the natural internal variability of the unforced climate system. A 2400 year control simulation with HadCM3 was available for analysis. Two ensembles of simulations were available to investigate the response of the model to changes in external forcing (Tett et al., 2002; Johns et al., 2003). The first ensemble is a set of four simulations with different initial conditions that include the same observed and estimated changes in concentrations of atmospheric greenhouse gases, ozone and sulphate aerosols over the period 1860–1999 (ANT runs) to represent the human influence on climate. The second ensemble of four simulations represents the climate response to natural external forcings, including estimated changes in total solar irradiance and volcanic aerosol amounts in the stratosphere over the period 1860–1999 (NAT runs).
4. Comparison of variability
The observed annual mean and seasonal mean Central England temperatures were calculated from monthly mean data. For the period from 1700 to 1900, there is likely to be only a very small influence from the climate response due to increasing greenhouse gases, so this period is used to estimate the natural variability of CET. The standard deviations of interannual and decadal variations of observed CET were calculated for the period 1700–1900. The standard deviation of the overlapping 50-year trends in CET during this period was estimated using a sliding 50-year window advancing 20-years for each sample.
CET data for 2400 years are available from the HadCM3 control run. The standard deviations of interannual and decadal variations of CET and 50-year trends were estimated using a sliding 201-year window (the same length as the observational data) advancing 50-years for each sample. The variability of the estimates over the long control run was used to provide 90% confidence intervals for the estimates from a 201-year sample. The comparison of observed and modeled variability of annual mean CET at interannual, decadal and 50-year time scales is shown in Figure 2. A similar comparison of observed and modeled variability of seasonal mean CET is presented in Table I. The model-simulated variability of CET at interannual and decadal timescales agrees very well with the observed CET variability for both annual and seasonal means. The model-simulated variability of annual mean CET at 50-year time scales is slightly larger than the observed variability. For the seasonal means in Table I, the model-simulated variability of 50-year trends is also slightly larger than observed in all seasons except autumn (SON). The observed variability probably includes some response to natural external forcing variations, such as changes in solar irradiance and volcanic aerosols, which are not included in the control model simulations. Hence, the model-estimated low frequency variability provides a slightly conservative estimate when used for detection and attribution studies, as the model-estimated variability of 50-year trends is slightly larger than observed, even though it does not include natural external forcing variations.
Table 1. Variability of Central England temperature at interannual, decadal and 50-year timescales from observations for 1700–1900 and from the HadCM3 control simulation. The 90% confidence intervals for the variability estimated from a 201-year sample are indicated for the HadCM3 values, based on resampling the model control run
Interannual standard deviation: Observed ( °C)
Interannual standard deviation: HadCM3 ( °C)
Decadal standard deviation: Observed ( °C)
Decadal standard deviation: HadCM3 ( °C)
Standard deviation of 50-year trends: Observations ( °C)
Standard deviation of 50-year trends: HadCM3 ( °C)
5. Detection and attribution of the warming trend in CET
The observed low frequency variations of CET from 1700 are shown in Figure 3, together with the ensemble average CET variations from the HadCM3 simulations with changes in anthropogenic (ANT) and natural (NAT) forcings. In the observed CET, there is substantial low frequency variability including rapid warming during 1700–1730, but the recent warming is larger and longer duration than any other period in the record. The ensemble mean CET from the simulations with ANT forcing shows a pronounced warming after about 1960 but this starts from a negative anomaly and does not reach the magnitude of the observed warming in 1999. The ensemble mean CET from the naturally forced simulations shows warming in the first half of the 20th century, in good agreement with the observations, but shows cooling after about 1950.
To assess the significance of the recent observed warming trends in CET, the observed linear trends in annual-mean CET over the period 1950–1999 and 1956–2005 are compared with the frequency distribution of 50-year trends from the HadCM3 control run in Figure 4. The observed warming in annual-mean CET over 1956–2005 of about 1.0 °C is significantly larger than trends due to natural internal climate variations at the 1% level. The observed warming over 1950–99 is slightly smaller and is significant at the 5% level. Hence, the observed annual mean warming trend over the last 50 years is very unlikely to be due to natural internal climate variability alone. The observed 50-year trends in seasonal CET are compared with the simulated trends in Table II. The warming trend in CET over 1956–2005 is largest in winter (DJF) and is significant at the 5% level in all seasons except autumn (SON). The observed annual and seasonal warming trends lie within the range of the model-simulated 50-year trends in CET from the ANT ensemble, except for autumn, and are significantly different from the model-simulated trends from the NAT ensemble. In autumn, the model-simulated trends are larger than in all the other seasons, including winter, in all the ANT ensemble runs. The ensemble-mean warming over 1950–1999 in autumns is significantly larger than observed. This is the only season where there is a significant difference between the observed warming and the ensemble-mean simulated warming from the ANT ensemble. The reason for this difference is unknown.
Table 2. Seasonal and annual mean 50-year warming trends in CET from observations and the ensemble-means from the HadCM3 simulations with changes in anthropogenic (ANT) and natural (NAT) climate forcings. The uncertainty ranges for the ensemble mean for forced trends are estimated to be from the long control run and not from the spread of the ensemble members
Observed trend 1950–1999 ( °C/decade)
Observed trend 1956–2005 ( °C/decade)
5–95% confidence interval for 50-year trends from HadCM3 control run
HadCM3 ensemble mean trend 1950–99 from ANT runs ( °C/decade)
HadCM3 ensemble mean trend 1950–99 from NAT runs ( °C/decade)
The significance of the recent trend of observed warming when considered over different trend lengths is assessed in Table III, where linear trends over 30-year, 50-year and 100-year periods are considered. For all these different trend lengths, the observed warming trends in annual mean CET have accelerated until 2005 and are significant at more than the 5% level. The significance levels of the warming trends do not increase when longer trend lengths are considered, probably because of the observed increase in the rate of warming towards the end of the observational record. It should be noted that there are large 30-year warming trends in the observed CET record in the early 18th and 19th centuries (see Figure 3) that are likely due to natural climate variations and are comparable in magnitude to the observed warming over 1970–1999. However, they are smaller than the observed 30-year warming trend to 2005.
Table 3. Annual mean warming trends in CET over 30-year, 50-year and 100-year intervals from observations and the ensemble-means from the HadCM3 simulations with changes in anthropogenic (ANT) and natural (NAT) climate forcings. The uncertainty ranges for the ensemble mean for forced trends are estimated from the long control run and not from the spread of the ensemble members
30 year trends
Observed trend ending in 1999 ( °C/decade)
Observed trend ending in 2005 ( °C/decade)
5–95% confidence interval for trends from HadCM3 control run
HadCM3 ensemble mean trend ending in 1999 from ANT runs ( °C/decade)
HadCM3 ensemble mean trend ending in 1999 from NAT runs ( °C/decade)
There is good agreement between the variability and the rate of recent warming of observed CET and that from a single grid box from simulations with the HadCM3 climate model. Given that numerical model simulations are not expected to reliably represent variations on the grid box scale due to their limitations in representing sub-grid scale processes and the limitations of numerical algorithms at the grid box scale, this agreement is remarkable. The large spatial coherence of low frequency temperature variations means that the variations at a single grid box are representative of temperature variations over a much larger area. All nine grid cells shown in Figure 1 have similar low frequency variability and recent warming trends, and compare well with the observed CET.
Another issue is that variations of the North Atlantic Oscillation (NAO) and associated changes in thermal advection contribute to a large fraction of the observed variability of CET in winter. The observed increasing trend in the NAO over 1965–1995 may have contributed to some of the warming trend in CET in winter, with one estimate being that 70% of the observed increase in Northern Europe (10W–50E and 50–70N) winter temperatures over that period is due to the strengthening of the NAO (Scaife et al., 2005). Since the observed trend in winter NAO is not simulated well by HadCM3 with changes in ANT forcing, some of the good agreement between the modeled and observed trends in DJF CET shown in Table II could be fortuitous. However, Table I shows that the winter variability of CET in HadCM3 is slightly larger than that observed at all three time scales considered, indicating that there is a greater confidence in the conclusion of detection of a significant warming trend in DJF CET over 1956–2005 (at the 5% level) than in the identification of a human influence in the recent winter warming. Although Table II shows that the ensemble mean model warming in winter over 1950–99 agrees well with the observed warming, this agreement could be misleading if a large part of the observed CET trend was caused by the 1965–1995 increase in the NAO.
In summary, the observed annual mean warming in CET over the last 30, 50 and 100 years is consistent with the model response to increasing greenhouse gases and aerosols and is not consistent with the response to changes in natural external forcing or to natural internal climate variations. Hence, there is evidence for a significant human influence in the recent warming of annual mean in CET, associated with increasing concentrations of greenhouse gases and aerosols in the atmosphere. The model does not simulate the observed increase in the NAO over 1965–1995 and, since the NAO explains a large fraction of the variability in CET in winter, the evidence for a human influence on winter warming is weaker than it is for annual mean warming.
This research was completed in part while DJK was a Visiting Scientist at the Hadley Centre (Reading Unit) in June–July 2005. DJK was supported by the Gary Comer Science and Education Foundation. PAS was funded by the UK Department for Environment, Food and Rural Affairs under contract PECD 7/12/37. The comments from two anonymous reviewers led to significant improvements in this paper.