![]()
Part I
The status analysis
![]()
1 Uncertainty in climate change projections
Mojib Latif
Introduction
The atmospheric carbon dioxide (CO2) concentration has strongly increased since the start of industrialization (Figure 1.1) in response to anthropogenic emissions and reached a level that is unprecedented in man’s history. The present CO2 concentration amounts to about 390ppm1 as opposed to the pre-industrial concentration of 280ppm. Carbon dioxide is a greenhouse gas and known to warm the Earth’s surface, as it is transparent for the short-wave solar but not for some of the long-wave infrared radiation emitted by Earth’s surface. The globally averaged surface air temperature (SAT) of the planet has warmed by about 0.7°C during the twentieth century (Figure 1.1), global sea level has risen by just under 20cm, and many mountain glaciers and Arctic sea ice have considerably retreated. There is compelling scientific evidence that at least half of twentieth-century warming was forced by the increase of GHG concentrations (IPCC, 2007). They will continue to rise over the next few years and possibly even decades, which together with the inertia of the climate system will support further global warming during this century. A global mean temperature rise implies higher warming over land than over oceans, with the tropical regions warming least and the northern polar region warming the most. But what else do we really know about the climate of the twentieth and twenty-first century?
Figure 1.1 Observed annual globally averaged SAT (°C) taken from HadCRU3 and smoothed atmospheric carbon dioxide concentration (ppm) during 1900–2008.
Natural variability
Climate change predictions are inherently uncertain. This does not come as a surprise to climate scientists. However, the media and public are generally confused by uncertainty. Hawkins and Sutton (2009) using data from a suite of climate models carried out a quantitative assessment of the contributions to uncertainty in predictions of global and regional surface air temperature change for the twenty-first century. They estimated the contributions to the total prediction uncertainty from natural internal variability,2 model uncertainty and scenario uncertainty. For lead times of the next few decades, the dominant contributions are internal variability and model uncertainty. Figure 1.2 displays the contributions to annual globally averaged SAT as a function of lead time. The relative contributions from internal variability and model error dominate at short lead times up to a few decades. At longer lead times of several decades, scenario uncertainty provides the largest share to prediction uncertainty. It is well known that the importance of internal variability increases at shorter (monthly to seasonal) time and smaller space scales, but the analyses of Hawkins and Sutton (2009) suggest that for decadal time scales and regional spatial scales (~2000km), model uncertainty is of greater importance than internal variability. It is important to note, however, that the contributions to prediction uncertainty from internal variability and especially from model uncertainty are potentially reducible through progress in climate science.
Figure 1.2 The fraction of total variance in annual mean surface air temperature predictions explained by the three components of total uncertainty (light grey: internal variability; dark grey: model uncertainty; and medium grey: scenario uncertainty).
Source: After Hawkins and Sutton (2009).
One source of uncertainty in climate change projections is natural variability, as noted above. Surface air temperature during the twentieth century displays a gradual warming and superimposed short-term fluctuations, while carbon dioxide evolved rather smoothly. The warming trend contains the climate response to enhanced atmospheric GHG levels but presumably also a natural component. The temperature ups and downs around the trend, which are particularly pronounced in the Arctic (Figure 1.3), largely reflect natural variability. Natural climate variations are of two types: internal and external. Internal variability is produced by the climate system itself due to its chaotic nature. External fluctuations need a forcing, a change in the boundary conditions. Volcanic eruptions and fluctuations in solar output are examples. The eruption of the Philippine volcano Mt. Pinatubo in 1991, for instance, caused a relatively short-lived drop in global SAT of about 0.2°C in 1992 (Figure 1.1); and an increase of the solar radiation reaching the Earth may have contributed together with other processes to the mid-century warming (MCW) during 1930–1940 (Figure 1.1). The anthropogenic influence on climate is also considered as external.
Figure 1.3 (a) Observed (HadCRU3) Northern Hemisphere averaged and (b) Arctic (60–90°N) SAT (red lines) and the multi-model mean of the CMIP3 models. The shading displays the model spread. Please note that sampling is different in the models than in HadCRU3, which may explain part of the differences.
Source: After Semenov et al (2010).
One way to estimate the external contribution to the twentieth-century SAT change is to run climate models with all (known) observed external (natural and anthropogenic) forcing in ensemble mode with different initial conditions. Figure 1.3 shows such simulations of Northern Hemisphere (upper panel) and Arctic (lower panel) SAT. The average over all (IPCC3) models taken from the CMIP34 database (Meehl et al, 2007) is sometimes referred to as the ‘consensus’ (black lines) and a measure of the externally driven climate change. The ensemble average displays a clear gradual upward trend in both indices that is consistent with the observed trend, but fluctuations are strongly damped by the averaging. Different initial climate states yield different realizations of internal variability even under identical external forcing, one reason for the spread (as denoted by the grey shading), as integrations are performed in ensemble mode with different start conditions.
A well-known example of such internal variability on interannual time scales is El Niño, a warming of the Equatorial Pacific occurring on average about every four years, which is the warm phase of the El Niño/Southern Oscillation (ENSO). The record event of 1997/1998 ‘helped’ to make 1998 the warmest year to date globally (Figure 1.1).5 The year 2009 also happened to be an El Niño year, which supported, for instance, a weak Atlantic hurricane activity, as El Niño causes enhanced upper-level vertical wind shear over the Tropical Atlantic, which is known to hinder hurricane development (e.g. Latif et al, 2007). The event, which was considerably weaker than the 1997/1998 event, persisted into 2010 and was partly responsible for the period January–June 2010 being the warmest first half year on record globally. However, the subsequent major cooling in the Equatorial Pacific referred to as La Niña ‘compensated’ the El Niño warming to some extent so that 2010 did not become the new record year. This discussion illustrates that caution should be exercised when interpreting short-term climate trends that are based on only several months or a few years.
Decadal-scale variability is evident in the data (e.g. Wang and Dong, 2010). In Figure 1.3, the deviation of the observed temperature evolution from the relatively smooth ‘consensus’ reflects the internal variability, assuming the model-mean is a reliable estimate of the externally-forced signal; and multidecadal changes are obvious in the residuals. Such multidecadal or even longer timescale natural variability, internally or externally driven, may mask anthropogenic climate signals that evolve on similar timescales. It has been concluded (e.g. Latif et al, 2006a), for instance, that the expected anthropogenic weakening of the Meridional Overturning Circulation (MOC), a prominent circulation in the Atlantic Ocean transporting large amounts of heat northward thereby contributing to the mild climate of Northern Europe, may not be detectable during the next decades due to the presence of strong internal multidecadal variability. This may not only apply to the MOC itself but also to other potentially related quantities such as surface air temperature in parts of Europe and North America, Sahel rainfall or Atlantic hurricane activity, which are also characterized by pronounced multidecadal variability (e.g. Zhang and Delworth, 2006) and may hamper early detection of an anthropogenic signal.
To some extent, we need to ‘ignore’ the natural fluctuations, if we want to ‘see’ the human influence on climate. Had forecasters extrapolated the MCW into the future, they would have predicted far more warming than actually occurred. Likewise, the subsequent cooling trend, if used as the basis for a long-range forecast, could have erroneously supported the idea of a rapidly approaching ice age. The scientific challenge is to quantify the anthropogenic signal in the presence of the background climate noise. The detection of the anthropogenic climate signal thus requires at least the analysis of long records, because we can be easily fooled by the short-term natural fluctuations, and we need to understand their dynamics to better estimate the noise level. Sophisticated fingerprint methods (Hegerl et al, 1996) maximizing the signal-to-noise ratio were applied to detect the anthropogenic signal in observations. The components of this strategy include observations, information about natural climate variability and a model-derived ‘guess pattern’ representing the expected time–space pattern of anthropogenic climate change. The expected anthropogenic climate change is identified through projection of the observations (say during the twentith century) onto the (model-derived) fingerprint. The latter can be optimized by weighting more strongly those components that are less ‘inflated’ by natural variability. Furthermore, the relative contributions of different external drivers of climate have been quantified. The results appear to be sufficiently robust to conclude that the observed climate change during the last decades cannot be explained solely by natural variability and is consistent with a combined greenhouse gas and aerosol6 forcing, but inconsistent with greenhouse gas or solar forcing alone (Hegerl et al, 1997).
It should be noted that internal natural variability such as El Niño and some decadal phenomena are predictable to some extent (e.g. Latif et al, 1998; Latif et al, 2006b; Smith et al, 2007; Keenlyside et al, 2008), and uncertainty in near-term climate projections can be potentially reduced by initializing climate models with the observed climate state.
Model uncertainty
The model spread seen in Figure 1.3 also reflects model uncertainty, as different models simulate different climate responses even when forced by the same GHG concentration or emission scenario.7 Climate models are grounded on basic physical principles. As such they are fundamentally different to empirical models that are used, for instance, in economic forecasting. Climate models, however, are far away from being perfect. Errors in annual mean SAT, for instance, typically amount to several degrees in some regions. The eastern tropical oceans are a hotspot in this respect, with sea surface temperature (SST) errors amounting to almost 10°C near the immediate vicinity of the coasts in many models. As a consequence the SST gradient along the Equatorial Atlantic is often reversed, yielding erroneous precipitation patterns over parts of South America and Africa. Another region exhibiting large SST errors in virtually all models is the North Atlantic Ocean. An incorrectly simulated path of the North Atlantic Current is at the heart of the problem, leading to errors also of the order of 10°C in some models. Such large biases in certain regions require more attention and coordinated international modelling programmes, as they are common to many models. Furthermore, internal variability in climate models such as ENSO is not always consistent with data (IPCC, 2007).
Limitations in computer resources dictate the use of either reduced or relatively coarse-resolution models. As a consequence many important processes cannot be explicitly simulated; they must be parameterized. Some processes such as cloud formation, cloud-radiation interaction and its influence on the general circulation of the atmosphere, or the role of mesoscale eddies for the large-scale ocean circulation are not completely understood and major sources of model bias. Two important features of MCW, the warming in recent decades and the projected future warming pattern, are known as the so-called ‘Arctic amplification’, a relatively high rate of war...
