previous chapter
Four The Quest For Reliable Regional Scenarios Of Climate Change
next chapter

The Quest For Reliable Regional Scenarios Of Climate Change

W. Lawrence Gates

It is fair to say that the demand for climate scenarios for impact estimation far exceeds the supply of reliable information that climate modelers have in their possession. Climate models certainly have systematic errors; most of these errors are poorly understood and are, in general, poorly documented. Nevertheless, the demand for scenarios and the need to use climate models in applications continue to accelerate. Modelers can only hope that this acceleration of demand and interest will somehow lead to an acceleration in modeling research, and the allocation of the resources that are needed for such work.

Climate models are a derivative of the models that are used to predict the daily weather, the so-called numerical weather-prediction models. Climate models differ from weather-prediction models in that climate researchers do not have to assume any particular initial state such as today's weather in order to go forward, and their models are integrated for much longer periods of time. The evolution of the atmosphere's three-dimensional distribution of variables such as wind, temperature, and pressure are predicted in a climate model, along with the occurrence of convection, clouds, and precipitation, and the structure of the surface boundary layer. Although they predict synoptic events as do weather models, climate models are used to derive the long-term average statistics which we regard as the model climate. Such models offer our only hope of achieving a quantitative understanding of possible future climates; the fields displayed by a climate model are physically consistent, have no missing data (in the sense that the horizontal and vertical distribution of data on the model's resolution are complete), and are objective (in the sense that they depend only on the assumptions or conditions under which the simulation is made). We begin by illustrating


current climate-model performance as a prelude to the consideration of the local climate changes related to impact estimation.

Climate-Model Performance

Figure 1 is an illustration of a typical climate model's performance, in this case a simulation of the February surface-air temperature; this happens to be a model from the Geophysical Fluid Dynamics Laboratory (GFDL) in Princeton, which is widely considered the premier modeling group in the United States. In comparison with the distribution of the observed surface-air temperature, the model's simulation may be regarded as reasonably successful on the large scale. There are, however, errors of the model (versus observation) of the order of 5° C. Such errors, of course, are of great interest to modelers but in general have not been thoroughly diagnosed. In the case of surface-air temperature, the errors are related to the model's inadequate treatment of the exchange of heat and moisture at the surface, which is in turn related to the model's failure properly to simulate the planetary boundary layer and the associated low-level cloudiness.

Model simulations and their errors can be examined for a wide variety of variables. Figure 2 shows the average January precipitation versus latitude for a number of early climate models. There is a lot of scatter amongst these models, as well as substantial errors with respect to the observed distribution shown by the thin full line. Although many of these models have since been improved, this kind of scatter is still characteristic of climate models. In general, the models tend to overestimate precipitation, although the reasons why this is so are not yet well understood. Because it occurs over short spatial scales, precipitation and the associated distributions of convection and cloudiness are notoriously difficult to represent properly in a climate model; cloudiness and precipitation are parameterized in a climate model as sub-grid scale processes that are not directly addressed or resolved. It is ironic to note that it is just these local or small-scale phenomena with which we are often most concerned in the estimation of the impacts of climate change.

Climate-Modeling Strategy

The methodology of climate modeling may be summarized as in figure 3. The climate or long-term averages obtained when a model is integrated under a standard set of conditions is what we refer to as the "control" climate. A model's results can be compared to actual observation, which thereby validate or verify the model. A comparison region by region and variable by variable allows the construction of a matrix of


Figure 1. The mean February surface-air temperature (°K) as simulated by a GCM
at GFDL (above) and as observed (below). (From Manabe and Stouffer 1980)

model errors. The aim of the climate modeler, of course, is to reduce these errors in a systematic way.

Once a control climate is validated, the model may be applied in experiments in which one or another condition is changed; an increase of CO2 is one of the more popular experiments and is the one with which we are most concerned here. When an experiment is made in which CO2 is doubled, the modeler "subtracts" the control climate from the experiment's climate, which then yields the modeled climate change. This is illustrated by the carbon dioxide change experiments given in figure 4, showing two independent models in which the CO2 has either been quadrupled

Figure 2. The mean zonally-averaged January precipitation as simulated
by a variety of GCMs and two estimates of the observed climatological
distribution (thin full and dashed lines). (From Gates 1987)


Figure 3. Elements of a climate-modeling strategy, in which the "true" climate change
to be expected from experimental changes is inferred from a model's experiment (relative
to a control or standard) and the model's verification errors (relative to observed climate).

(upper panel) or doubled (lower panel) in reference to the models' control. Neither the patterns nor the magnitudes of the modeled changes of January average surface-air temperature are similar. Sorting this out and pinpointing which model features are responsible for these differences is a continuing and important research problem.

The Estimation Of Regional Changes

Even though a particular climate model's results may be satisfactory on the large scale, there remains the problem of determining changes on the local or model sub-grid scale, which is everything below several hundred


Figure 4. The distribution of the change in mean December-January-February surface-air
temperature (° C) as a result of quadrupled CO2 with a GCM at GFDL (above) and as a
result of doubled CO2 with a GCM at NCAR (below). (From Washington and Meehl 1984)

kilometers. A strategy for accomplishing this is shown in figure 5, in which the ordinate is useful climate information. In their raw output, climate models certainly have useful information on the large-scale end of this diagram. The first step in using these results is to determine the statistical significance of the modeled climate changes from suitably long control and experimental runs, a step that is not always taken in a systematic way. Assuming this has been done and that at least some of the results have been judged statistically significant, there remains the problem of going from large to small scales.


Figure 5. Elements of a strategy for the application of the results of GCM experiments
to the estimation of the local impacts of climate change. See text for description
of steps 1–5. (From Gates 1985)

The local climate could be estimated by interpolation of the model's solutions. Although there is no theoretical basis for such a procedure (since the solution between model grid-points is not determined), it has often been done. Taking the state of Oregon as an example of an area representing a grid point in a typical climate model, figure 6 illustrates the local geographic information. What has been done here is to take the local or station monthly averaged observation of surface temperature

Figure 6. An illustration of the statistical distribution (over the state of Oregon) of the
change of monthly mean surface-air temperature (above) and precipitation (below) that
may be statistically inferred from a change of the area averages alone.
(From Kim et al. 1984)


and precipitation from thirty years of record and ask the question, How much of the local variation from month to month could be accounted for on the basis of only the statewide area average changes? The answer is shown in terms of the contours of the covariation between the local variability and the variation of the grid-size area average. (Figure 6 actually shows the amplitude of the first eigenfunction of the correlation.) Evidently one can infer an appreciable amount of the local variability of temperature (and similarly for precipitation) from the change of the area mean alone, together with knowledge of the local climatology. By the use of such an empirical transfer function, one could estimate the local allocation of a modeled climate change such as that due to increased CO2 , assuming local climatological observations were available. Such functions clearly show the dominant local influence of features such as mountains and large water bodies.

Returning to figure 5, the approach to local scales could also be achieved by embedding a higher-resolution model within a large-scale climate model, say over the western United States or over California. This is routinely done in weather prediction, but whether or not such a technique will be successful for climate is still an open question. The success of this technique will depend critically upon our knowledge of the resolution-dependence of climate models, a matter to which relatively little attention has been given.

Once estimates of local climate change have been made, the next step is the construction of specific measures or statistics related to the local climate impacts of interest, such as those related to energy, agriculture, water use, or local ecosystems, after which the local impacts may be determined with an impact model. Finally, the large-scale impacts may be constructed by appropriate aggregation.


A number of problems still confront climate modelers if the challenge of providing information that can be used for local impact estimation is to be met. First, the models must be improved. Models continue to show large systematic errors, and the structure and behavior of these errors are not well understood. This suggests the need for more analysis, diagnosis, and intercomparison of model results, in order to understand the reasons for the differences among models and their sensitivity to both parameterization and resolution. It is to be hoped that the resources necessary to do this on a sustained and coordinated basis will be made available.

It should also be recognized that modeled climate changes will inevitably be in terms of frequency distributions rather than categorical results. Ideally, these distributions should be constructed from the statistics of an ensemble of model runs, rather than by guessing or by uncertain


analogs. Such information would permit the determination of the risk or uncertainty of the derived climate impact estimates and would place climate model applications on a firmer scientific basis.




Question: I did not hear any mention of sensitivity testing, which is a common procedure in modeling. Can you identify those parameters which are likely to have the greatest influence on the results of a model? In other words, to what extent have models been subjected to sensitivity testing with respect to their various parameters and processes?


Answer: To a relatively limited extent. Given the number of processes that are represented in a climate model and the number of variables involved, we have only been able to examine what we believe are the most important cases. Perhaps most sensitivity experiments have been made with respect to ocean-surface temperature, although in terms of the internal parameters of a climate model, clouds may be the more important. Depending upon how clouds are parameterized, the change of surface-air temperature in response to CO2 doubling, for example, changes by more than a factor of two.


Question: Since you are interested in the difference between a standard and an elevated CO2 level, how big an influence does the initial starting condition have on the difference?


Answer: If you change the initial condition by a small amount, for example, by changing the last decimal at one point in an otherwise identical calculation, after a few weeks of simulation you will get a different synoptic pattern as a result of the interactions between small and large scales in the model. The time-averaged or climatic conditions, however, will be only slightly changed; such changes are a measure of what is termed natural variability or climatic noise.


Question: How well do the models simulate the position of the subtropical highs, and by how much do they change in the future, presumably under increased CO2 ?


Answer: The models do a fair job of simulating the subtropical highs, in that they are all in about the right positions (if we interpret "right" as being about a thousand kilometers). They tend, however, to be too weak, although their seasonal shift is recognizably realistic. The change of the subtropical highs under increased CO2 appears to be relatively small, as indeed do most changes in the tropics.


Gates, W. L. 1985. The use of general circulation models in the analysis of the ecosystem impacts of climatic change. Climatic Change 7:267–284.


Gates, W. L. 1987. Problems and prospects in climate modeling. In Toward understanding climate change , ed. U. Radok. Boulder, Colo.: Westview Press.

Kim, J. W., J. T. Chang, N. L. Baker, W. L. Gates, and D. S. Wilkes. 1984. The statistical problem of climate inversion: Determination of the relationship between local and large-scale climate. Mon. Wea. Rev. 112:2069–2077.

Manabe, S., and R. J. Stouffer. 1980. Sensitivity of a global climate model to an increase of CO2 concentration in the atmosphere. J. Geophys. Res. 85:5529-5554.

Washington, W. M., and G. A. Meehl. 1984. Seasonal cycle experiment on the climate sensitivity due to a doubling of CO2 with an atmospheric general circulation model coupled to a simple mixed-layer ocean model. J. Geophys. Res. 89:9475-9503.


previous chapter
Four The Quest For Reliable Regional Scenarios Of Climate Change
next chapter