(Postscript version)
Matthew Collins(1,4), Timothy J. Osborn(2), Simon F. B. Tett(1), Keith R. Briffa(2) and Fritz H. Schweingruber(3)
(1) Hadley Centre for Climate Prediction and Research, The Met. Office, London Road, Bracknell, RG12 2SZ, UK. (2) Climatic Research Unit, University of East Anglia, Norwich, NR4 7TJ, UK. (3) Swiss Federal Institute for Forest and Snow Research, Zurcherstrasse 111, CH-8903 Birmensdorff, Switzerland. (4) Centre for Global Atmospheric Research, Dept. of Meteorology, University of Reading, Reading, RG6 6BB, UK.
Corresponding Author: Matthew Collins Centre for Global Atmospheric Research, Dept. of Meteorology, University of Reading, Reading, RG6 6BB, UK. Tel: +44 118 9875123 x 7871 Fax: +44 118 9318316 matcollins@met.rdg.ac.uk http://www.met.rdg.ac.uk/~mat
In studies of the detection, attribution and prediction of anthropogenic climate change, it is essential to have some estimate of natural fluctuations of climate in order to identify the significance of the human-induced climate change signal. The problem is one of separating the "signal" of climate change from the "noise" of natural variability, and often this involves making quantitative estimates of the confidence limits of the probability density function (PDF) of the natural variability (e.g. Tett et al (1999).
Because of the multi-decadal to century time scales involved in the climate change problem, estimates of this PDF must be robust on such long time scales. The (global) observational record of climate is of little use because of its relatively short length (of a century or so) and because of its probable contamination with the anthropogenic climate change signal. Thus it is common in climate change studies to use an estimate of the natural variability taken from a long (i.e. multi-century) control experiment of a climate model, often a coupled ocean-atmosphere GCM. It is crucial then to validate the variability of such climate models on the multi-decadal to century time scales because of their weighting in the signal detection algorithms of the climate change problem. As the observed global record is inadequate in this respect, our only viable option is to use estimates of climate variability from palaeo records. The validation procedure must be quantitative (because we require quantitative estimates of the PDF of natural climate variability) but there are many problems associated with comparing palaeo estimates of climate variability with numerical models (e.g. Jones et al 1998). For example, palaeo indicators (e.g. those derived from ice-cores) are often measurements of local climate, whereas climate models represent variables on the scale of their grid-boxes which can be many 100s of kilometres. Also palaeo indicators are often expressed in terms of variables which are not predicted by climate models (oxygen isotope ratios, tree-ring widths, etc).
Here we report briefly on a quantitative comparison of the decadal-century time scale variability of a coupled ocean-atmosphere climate model with palaeo-temperature estimates of the last 600 years derived from an extensive network of tree-ring densities. Collins et al (2000b) contains the full details of the study.
The climate reconstructions used here are based on a network of 387 tree-ring density chronologies located over much of the northern hemisphere extra-tropics (fig. 1). The chronologies range in length from 100 to more than 600 years, with each consisting of, on average, data from 25 tree cores from a site close to the present timber-line (i.e., at high elevation or high latitude) to maximise the temperature signal (Briffa et al (2000b)). The dominant climate signal in the data set as a whole is the growing season temperature, the timing of which varies with location, but the mean temperature from April to September provides the best overall correlations with tree-ring density. Briffa et al (2000a,b) aggregate the local tree-ring series into regions (indicated on fig. 1) and calibrate against observed temperatures over the period 1881-1960 using simple linear regression. They also form a Northern Hemisphere (NH) series as a weighted mean of these regional series. The linear regression leaves some unexplained variance and we take account of this in the comparison with the model (see Briffa et al (2000b) and Collins et al (2000b) for more details).
In addition to responding to growing-season temperature, the maximum latewood density of each tree-ring also depends upon the age of the ring (generally showing a downward trend with increasing tree age). Briffa et al (2000a) contrast two approaches for removing this age effect. The "Standardisation" technique involves fitting and removing a generalised exponential function from each tree core and can result in a loss of multi-century variance, the extent of which is dependent on tree longevity. With the Age-Band Decomposition technique (ABD - Briffa et al (2000a)), the age effect is accounted for by only combining in absolute units the density from tree rings whose age falls in a restricted range (or band). There is no artificial loss of multi-century variability, but this is at the cost of greater uncertainty in the earlier part of the record for which there are fewer tree cores. Thus we contrast results obtained using the two methods for processing the tree-ring data which we denote "standard" and "ABD".
We compare the tree-ring temperature estimates with a 1200 year control run of version three of the Hadley Centre Climate Model (HadCM3 - Gordon et al (1999), Collins et al (2000a)). HadCM3 has an atmosphere with a 3.75x2.5 degree longitude-latitude grid and 19 vertical levels, and an ocean with a 1.25x1.25 degree grid and 20 levels in the vertical. The model requires no flux adjustment term and has a stable climate in the global mean when initialised from an observed atmosphere-ocean state. The control simulation has constant concentrations of greenhouse gases and aerosols etc and hence only represents "internal" climate variability. Surface air temperatures (at a height of 1.5m) were extracted from land points in the regions indicated in fig. 1 and during the growing season of the trees (April-September).
A simple yet quantitative way of comparing the variability of HadCM3 with the tree-ring estimates is to compute the variance (or standard deviation) of temperature regionally and over the northern hemisphere as a whole (fig 2a-b). We first average all time series into decades to focus on the decadal-century time scales and we take into account the residual variance from the calibration procedure. The reader is referred to Collins et al (2000b) for a comparison of other diagnostics such as power spectra and spatial patterns.
Comparing HadCM3 with the standard tree-ring reconstructions (fig. 2a) the model captures the regional spatial pattern of variability well and there is no systematic under or overestimation of variability. For the hemispheric variability, the model underestimates the temperature variance, significantly so (by a factor of 1.6) when we include the residual variance from the calibration procedure. Comparing HadCM3 with the ABD tree-ring data (fig. 2b) implies that the model on the whole underestimates regional variance with the maximum disparity between the model and the tree-ring reconstruction being for the Northern Siberian (NSIB) region where the tree-rings have over 6 times the variance of the model. For the Northern Hemisphere (NH) as a whole the the model underestimates the variance by as much as a factor of 3.
Underestimation of the temperature variance by the model is serious as it could lead to false claims of the detection (and attribution) of climate change (e.g. Tett et al (1999)) and to underestimation of the uncertainties in future climate prediction. (Also, regional errors in variance can lead to errors in the relative weights used in the optimal detection algorithm thus making it less powerful.) Hence it is important to consider why the model may be underestimating the temperature variability.
The control simulation of HadCM3 only represents the "internal" variability of the climate system - that which is a consequence of non-linear interactions within (and between) the atmosphere and the ocean. Other "natural" factors such as variations in solar irradiance, volcanic eruptions, natural fluctuations in CO2 etc. can affect climate and the tree-ring data might contain variance attributable to these natural forcings (over the period considered here, the last 600 years, orbital variations are of secondary importance). Hence we should include these factors in our simulation in order to make a correct comparison.
We have been unable to perform such a simulation, partly due to the lack of forcing histories and partly due to constraints on computer time. However, we do have 4 simulations with estimates of solar (Lean et al (1995)) and volcanic (Sato et al (1993)) forcing from 1860-1997. The variance of the regional and NH temperatures from these simulations are shown in fig 2c. The forcings generally do not significantly increase the level of variance on the regional scale (apart from TIBP) but they do enhance the total NH variance by a factor of 1.9. Thus it is possible that the underestimation of variance by the model control simulation is due to the lack of natural climate factors such as solar variations and volcanoes. The conclusion is tentative because we have not run the model over the full 600 year tree-ring period with all the natural forcing factors, however it is consistent with the recent work of Crowley (2000).
We have compared the temperature variability of a coupled climate model with palaeo-temperature estimates from a large network of tree-ring densities. On the hemispheric scale, the model appears to underestimate variance (by as much as a factor of 3) which is serious if the model is used as a surrogate for natural climate variability in studies of the detection, attribution and prediction of climate change. However, we have shown that this underestimation may be due to the lack of natural forcing factors such as solar variations and volcanoes. More detail can be found in Collins et al (2000b).
Palaeo estimates of climate variability are the only way of validating climate models on time scales of many decades to centuries. This study has highlighted many areas where there needs to be more work. Firstly it is important to correctly interpret (e.g. Mann et al (1998)) and quantify the uncertainties in the palaeo data. Secondly it appears that models need to be forced with natural factors in order to make a like-with-like comparison. This in turn requires palaeo-estimates of these forcing factors (e.g. Crowley and Kim (1999)). Finally there is a need for a framework (such as the optimal detection framework (Tett et al (1999)) in which all the uncertainties in the model and in the palaeo data can be taken into consideration when making the comparison.
Tim Johns ran the HadCM3 control integration and Gareth Jones ran the ensemble of simulations with natural forcings. This work was supported by UK Department of the Environment, Transport and the Regions (PECD/7/12/37) and the Public Meteorological Service Research and Development Programme - MC and SFBT; the UK Natural Environmental Research Council (GR3/12107) and the European Union (ENV4-CT95-0127) - TJO, KRB, and FHS.
Briffa, K. R., T. J. Osborn, F. H. Schweingruber, I. C. Harris, P. D. Jones, S. G. Shiyatov, and E. A. Vaganov, 2000a: Low-frequency temperature variations from a northern tree-ring density network. J. Geophys. Res. Submitted.
Briffa, K. R., T. J. Osborn, F. H. Schweingruber, P. D. Jones, S. G. Shiyatov, and E. A. Vaganov, 2000b: Tree-ring width and density data around the northern hemisphere: Part 1, local and regional climate signals. The Holocene. Submitted.
Collins, M., S. F. B. Tett, and C. Cooper, 2000a: The internal climate variability of HadCM3, a version of the Hadley Centre coupled model without flux adjustments. Climate Dynamics. Accepted for publication.
Collins, M., T. J. Osborn, S. F. B. Tett, K. R. Briffa and F. H. Schweingruber, 2000b: A comparison of the variability of a climate model with palaeo-temperature estimates from a network of tree-ring densities. J. Climate. Submitted. Manuscript available from the lead author.
Crowley, T. J., and K.-Y. Kim, 1999: Modeling the temperature response to forced climate change over the last six centuries. Geophys. Res. Letts., 26(13), 1901-1904.
Crowley, T. J., 2000: Causes of climate change over the past 1000 years. Science, 289, 270-277.
Gordon, C., C. Cooper, C. A. Senior, H. Banks, J. M. Gregory, T. C. Johns, J. F. B. Mitchell, and R. A. Wood, 2000: The simulation of SST, sea ice extents and ocean heat transport in a version of the Hadley Centre coupled model without flux adjustments. Climate Dynamics, 16, 147-168.
Jones, P. D., K. R. Briffa, T. P. Barnett, and S. F. B. Tett, 1998: High-resolution palaeoclimatic records for the last millennium: interpretation, integration and comparison with general circulation model control-run temperatures. The Holocene, 8(4), 455-471.
Lean, J., J. Beer, and R. Bradley, 1995: Reconstruction of solar irradiance since 1610: Implications for climate change. Geophys. Res. Letts., 22, 3195-3198.
Mann, M. E., R. S. Bradley, and M. K. Hughes, 1998: Global-scale temperature patterns and climate forcing over the past six centuries. Nature, 392, 779-787.
Sato, M., J. E. Hansen, M. P. McCormick, and J. B. Pollack, 1993: Stratospheric aerosol optical depths (1850-1990). J. Geophys. Res., 98(D12), 22987-22994.
Tett, S. F. B., P. A. Stott, M. R. Allen, W. J. Ingram, and J. F. B. Mitchell, 1999: Causes of twentieth century temperature change. Nature, 399, 569-572.
Figure 1: Locations of individual tree-ring density chronologies (dots) and the definition of the nine regional series used in the calibration against observed temperature. April-September temperatures were extracted from HadCM3 at land points in these regions.
Figure 2: A comparison of the variance of temperature of HadCM3 and palaeo-estimates of temperature from tree-ring data. The black bars correspond to the model and the grey bars correspond to the tree-rings. The white bars show the residual variance from the calibration procedure which must be added to the tree-ring variances. The numbers above the bars are the ratios of tree-ring to model model variance without the residual variance (upper) and with the residual variance (lower). Black numbers are statistically significant at the 95% level using an F-test. (a) is for the standard tree-ring data and (b) is for the ABD tree-ring data. (c) is a similar plot comparing the variance of the HadCM3 control and the HadCM3 simulations with natural forcings.
