Fhm Australia Models

[Lcs Basinger]

Cumulativeensemble percentile values for (a) P, (b) TMIN, and (c) TMAX. Diamonds represent the 80th percentile, squares the 90th, and triangles the 95th. The observed is shown at the right-hand side of each figure. Note all daily values

The coupled climate models used in the Fourth Assessment Report of the Intergovernmental Panel on Climate Change are evaluated. The evaluation is focused on 12 regions of Australia for the daily simulation of precipitation, minimum temperature, and maximum temperature. The evaluation is based on probability density functions and a simple quantitative measure of how well each climate model can capture the observed probability density functions for each variable and each region is introduced. Across all three variables, the coupled climate models perform better than expected. Precipitation is simulated reasonably by most and very well by a small number of models, although the problem with excessive drizzle is apparent in most models. Averaged over Australia, 3 of the 14 climate models capture more than 80% of the observed probability density functions for precipitation. Minimum temperature is simulated well, with 10 of the 13 climate models capturing more than 80% of the observed probability density functions. Maximum temperature is also reasonably simulated with 6 of 10 climate models capturing more than 80% of the observed probability density functions. An overall ranking of the climate models, for each of precipitation, maximum, and minimum temperatures, and averaged over these three variables, is presented. Those climate models that are skillful over Australia are identified, providing guidance on those climate models that should be used in impacts assessments where those impacts are based on precipitation or temperature. These results have no bearing on how well these models work elsewhere, but the methodology is potentially useful in assessing which of the many climate models should be used by impacts groups.

The importance of examining climate statistics other than climate means is not new (Katz and Brown 1992; Boer and Lambert 2001). For example, recent studies by Frich et al. (2002), Kiktev et al. (2003), and Alexander et al. (2006) used climate indices and probability density functions (PDFs) of indices to explore the frequency and severity of climate extremes. These indices-based analyses provide a clear way forward in using climate model results for society-relevant impact assessments. Dessai et al. (2005) calculated PDFs, using seasonal data from climate models, to assess uncertainty in regional climate change projections. They devised a skill score that accounted for model bias and spatial variation to compare models and data for surface air temperature and precipitation.

There is at least one major advantage of evaluating a climate model based on PDFs. If a climate model can simulate an entire PDF, this demonstrates a capability to simulate values that are currently rare and that may become more common in the future. If the distribution of values represented by a PDF shifts due to climate change, it is likely that significant overlap between the new distribution and the present distribution will remain. If a climate model that has been shown to already simulate this region of the distribution that currently exists and will remain in the future (but becomes more likely to occur probabilistically), then we have identified that the model has skill in simulating these future values. Clearly, this confidence declines as the overlap between the present and future PDFs is reduced further into the future. Until the overlap becomes critically small, however, an impacts modeler could use how well a model simulated the whole PDF of a set of variables as criteria for those models to use in future impacts assessments. Further, establishing the skill of a climate model to simulate whole PDFs is a far harder test of a model than (say) the mean and one standard deviation, and thus by succeeding in such a test, we might have more confidence in projections made with this model. We do not claim that a PDF-based assessment of a model is perfect of course. While performing well in a PDF-based assessment is substantially harder than reproducing the mean, key areas of model performance such as periods of sustained high temperature or rainfall represented by indices (e.g., Kiktev et al. 2003; Alexander et al. 2006) are not assessed. Further, as an event becomes rarer in both the model and the observed data, failure of the model to simulate these events becomes less important to the skill score. While these represent limits to our methodology, we are confident that a PDF-based evaluation of a climate model is substantially preferable to a mean-based assessment and could simply replace the traditional reliance on evaluating the mean performance.

This paper explores the capacity of a large sample of climate models to simulate the PDFs of precipitation (P), minimum temperature (TMIN), and maximum temperature (TMAX). The choice of these three variables was based on available data and on their role in driving many human, industrial, and biological systems (Colombo et al. 1999; Meehl et al. 2000; Christensen and Christensen 2003; Trigo et al. 2005). We utilize the model results submitted to the Program for Climate Model Diagnosis and Intercomparison (PCMDI) at the Lawrence Livermore National Laboratory in the United States ( -pcmdi.llnl.gov/ipcc/about_ipcc.php) as part of the Fourth Assessment Report conducted by the IPCC (AR4).

We have chosen to identify individual models in this paper for two reasons. First, these data are currently being widely used and unless we identify models, users of simulations cannot determine those models with particular strengths or weaknesses. Second, unless model groups know their model performs well/poorly, they cannot build on strengths or address weaknesses in subsequent model development. However, we wish to emphasize that a model that shows skill/weakness over Australia may/may not show comparable skill/weakness in other regions and each user should evaluate the models they choose for their specific region of interest.

Figure 2a presents three PDFs for a 10 10 region of Australia to illustrate the relative insensitivity of a PDF to errors in the observations (compared to means or standard deviations). First, the original observed data (solid line) has a mean of 23.34C and a standard deviation of 6.67C. We then perturb the original dataset by increasing 5% of the values by 5% to obtain the second PDF and to give a revised mean of 23.52C and standard deviation of 6.77C. Finally, we perturb the original dataset by increasing 10% of the values by 10% to obtain the third PDF, giving a revised mean of 23.57C and a standard deviation of 6.75C. While both the mean and standard deviation change in each case, Fig. 2a shows little change in the shape of the resulting PDFs. We will show later that differences between observed and simulated PDFs are commonly too large to be explained by observational error. A second advantage of using PDFs is that we can safely merge data from multiple stations where the data lengths are different and/or a station samples a small amount of a total time series. We can thus use gap-affected observational data with relative ease provided we assume these gaps (in time) are random in terms of their likelihood of contributing to a particular part of the PDF.

In terms of spatial coverage, Fig. 1 indicates that data coverage is biased toward the coast, in particular in southeastern and southwestern Australia. However, excluding regions 8 and 9, data coverage is quite complete with stations widely distributed even in the sparsely populated areas of the continent. In regions 8 and 9, in particular for precipitation, data coverage is clearly incomplete and/or spatially biased. Overall, however, by using a PDF-based analysis that allows all stations to be used to estimate the probability of a given temperature or precipitation event (as distinct from the mean), our comparison of modeled and observed temperature and precipitation can be based on a more data-rich foundation.

3a8082e126