Regional Projections of Future Seasonal and Annual Changes in Rainfall and Temperature over Australia Based on Skill-Selected AR4 Models

2.1. Data

Daily climate model data over Australia for P, T_MIN, and T_MAX were taken from the Program for Climate Model Diagnosis and Intercomparison (PCMDI) archive (http://www-pcmdi.llnl.gov/ipcc/about_ipcc.php). Data from 1981–2000 from the Climate of the Twentieth Century simulations were used as the control (these are fully discussed in Perkins et al. 2007). In this paper we also use results from the B1 (relatively low emissions) and A2 (relatively high emissions) scenarios for two time periods: a 20-yr time period from 2046 to 2065 (hereafter 2050) and a 20-yr period from 2081 to 2100 (hereafter 2100). These time periods were common to all models. By using daily data we retain the maximum time resolution possible and minimize the hiding of biases through averaging.

Some datasets contained erroneous data (gaps, periods of repetitive data), or data were not available at the time this study was undertaken, and were therefore omitted from subsequent analysis. Table 1 lists all models used. When calculating differences in means and yearly returns, only models that had data for both the twentieth-century simulation and the respective twenty-first-century simulation and time period could be used. We used each independent realization directly in the initial analysis rather than average these realizations to produce an ensemble. However, we present our best estimates of the multimodel ensemble over included models. In selected regions, we highlight the differences between models included in the ensemble results.

Daily observed P, T_MIN, and T_MAX were obtained from BOM for the period 1981–2000. The use of the observed data is fully discussed in Perkins et al. (Perkins et al. 2007).

2.2. Skill score

Perkins et al. (Perkins et al. 2007) present a simple skill score to rank the AR4 climate models over Australia. This metric measures the similarity between two PDFs by calculating the cumulative minimum value of two distributions of each binned value. This measurement of the common area between two PDFs provides a skill score that will equal 1.0 where the modeled and observed PDFs are identical and 0.0 where the two PDFs fail to overlap at all. Perkins et al. (Perkins et al. 2007) demonstrate that this measure is robust against uncertainty in the temporal and spatial coverage of the observations. Following Perkins et al. (Perkins et al. 2007) bin sizes were 0.5°C for T_MAX and T_MIN and 1 mm day⁻¹ for P. The ranges of the bins were dependent on the range of the observed data for each region. All daily values of P below 0.2 mm day⁻¹ were set to 0.0 because rates below this amount are not recorded in the observations (Parkinson 1986).

In this paper we use the skill scores obtained by Perkins et al. (Perkins et al. 2007) as the basis for omitting models. We omit models based using a threshold of 0.8. The choice of this was subjective, balancing the desire to only include those climate models with demonstrated skill, while recognizing that the sample size of models needs to be kept reasonable. Had we chosen a skill score of 0.6 virtually no models would be excluded while 0.9 would mean virtually no models were included.

2.3. Calculation of continent ensemble and regional statistics

This study highlights the change in , , P, T_MAX⁹⁹, T_MIN^0.3, P⁹⁹ as well as the change in the frequency of a twentieth-century value for the respective percentiles and the change in the number of no-rain days. The 99.7th and 0.3rd percentiles were chosen as they represent an event that occurs once every year, on average, in the present day. The respective statistics were calculated for A2 and B1 for 2050 and 2100 for all seasons. When calculating the change in frequency of a one-in-one-year twentieth-century value, the number of times an event of the same magnitude or greater was counted in the twenty-first-century dataset and divided by 20 to give the mean yearly occurrence. All statistics were calculated at a regional level for regions 1, 2, 3, 10, and 11 (see Figure 1 and Table 2) and at a continental scale. These regions were chosen because they represent different climatic types (Table 2) and contain most of Australia’s population (6 out of 8 capital cities). When a model provided more than one realization, data were averaged to form a model ensemble prior to statistical calculation since Perkins et al. (Perkins et al. 2007) found very little difference in skill between a model’s realizations.

At a regional level, all model data that were contained by the specified region’s boundaries were concatenated. Models that demonstrated a skill of <0.8 defined by Perkins et al. (Perkins et al. 2007) were then omitted. This means that the number of models in the ensemble varies from region to region and across variables. However, this is kept constant (region specific) throughout the seasons, twenty-first-century time periods, and scenarios. Perkins and Pitman (Perkins and Pitman 2008, manuscript submitted to Climatic Change) identify the amounts the specific projections change given a choice of a specific skill level.

At the continental scale, all statistics were calculated at the models’ native resolution for each scenario and each time period. Model skill masks were also constructed at the native resolution. This comprised using the skill score for each region and fitting the values to a land–sea mask of the model. This maintained the exact skill values by avoiding interpolation. The skill masks were fitted to the model such that grid boxes within regions with a skill <0.8 were changed to missing values. All model data was then interpolated to a common 1° × 1° grid. The continental ensemble was created by simply averaging the interpolated statistics, giving the mean statistic value across the “more skilled” models at each grid box. Continental results for each season, scenario, and time period are presented as maps and regional results as stock plots to illustrate the range of values across the ensemble.

3.1. Mean maximum temperature

Figure 3 shows the results for for 2050 and 2100 for the B1 emission scenario. In 2050, warming exceeding 1.0°C is projected over virtually all the continent in all seasons with the exception of southeastern Australia. In spring, warming exceeds 1.5°C almost everywhere, exceeding 2°C in central regions (Figure 3a). A similar result occurs in summer (Figure 3b). Few areas warm by more than 1.5°C in autumn (Figure 3d). Warming is generally 1.0°–1.5°C in the southern half and 1.5°–2.0°C in the northern half of Australia in winter.

By 2100, the increase in is substantially higher. Warming exceeds 3°C in central regions in spring (Figure 3f) and widespread regions experience 2.0°–2.5°C increases in spring, summer (Figure 3g), and winter (Figure 3i). The least amount of warming occurs in autumn (Figure 3h), but warming still exceeds 1.5°C everywhere except for areas of high orography in the east and southeast. Clearly, the annual average (Figure 3j) hides significant seasonality and geographical variability in the response to warming. There is about 0.5°–1.0°C additional warming between 2050 and 2100.

Under A2 (Figure 4) additional warming in is projected in most regions. By 2050, spring warming (Figure 4a) reaches 3°C in western regions. In summer (Figure 4b), warming reaches about 1.0°–1.5°C over Queensland (this is less than under B1) but 2.0°–2.5°C in the western regions which is higher than under B1. This is duplicated in autumn (Figure 4c). In winter, warming is mainly 2.0°–2.5°C, a full 1°C higher than under the B1 scenario. The lack of large differences in greenhouse gas concentrations by 2050 between the B1 and A2 emission scenarios coupled with natural variability in the models explains how summer warming in some regions is lower under the A2 scenario than B1.

By 2100, under a high-emissions scenario, warming of at least an additional 2°C occurs in comparison to 2050 and in comparison to results for 2100 under the B1 scenario. In spring, large areas of warming exceed 4°C and exceed 5°C in Western Australia. The least warming simulated exceeds 2.5°C (over the northern Queensland coast and the extreme southern area of Victoria). A generally similar result occurs in summer (Figure 4g) and winter (Figure 4i). In autumn parts of Queensland experience slightly less warming in the coastal region, but this still generally exceeds 2°C.

3.2. 99.7th percentile of maximum temperature (T_MAX⁹⁹)

An analysis of the change in T_MAX⁹⁹ illustrates how depending on changes (section 3.1) can mislead. Figure 5a shows the change in T_MAX⁹⁹ in spring under B1. While the northern half of Australia experiences a similar change to the (Figure 3a), the southern half displays increases of about 4°C (double the increase in ). A similar result is seen in summer (Figure 5b) and autumn (Figure 5c). In summer, the region of higher rises in T_MAX⁹⁹ extends northward through New South Wales and into Queensland. In winter, while increases by 1.0°–2.0°C, T_MAX⁹⁹ increases by 2.5°–3.5°C over wide regions.

This basic pattern is reproduced for 2100 under B2. Along the southern coast of Australia warming in T_MAX⁹⁹ exceeds 5°C in spring, summer, and autumn. This is closely related to the rainfall pattern discussed later. Most of the Murray–Darling Basin (MDB) experiences significant warming of more than 3.5°C in spring and winter that is likely to negatively impact on a basin which is already undergoing major challenges with respect to water availability.

Under A2, increases in T_MAX⁹⁹ are higher over southeastern Australia in spring and winter compared to B1 (Figure 6). Warming of 2.0°–2.5°C is widespread. Smaller increases occur elsewhere in spring (Figure 6a). Local increases exceeding 4.0°C occur over Victoria. In summer (Figure 6b) and autumn (Figure 6c) increases in T_MAX⁹⁹ exceeding 1°C are isolated to southeastern Australia and Western Australia (Figure 6b). However, increases exceeding 3.5°C occur over Victoria. Indeed, each season shows anomalously high increases over southern Victoria commonly exceeding 3.5°C and locally exceeding 5°C.

Significant areas of Australia are projected to warm by less than 1°C in T_MAX⁹⁹ by 2050 under A2 (Figures 6a–e). This result is quite noticeable in comparison to B1 (Figures 5a–e). However, a close inspection of these figures shows that the differences are small: while the A2 future does project less warming, it is only slightly less. By 2100, however, the amount of warming in T_MAX⁹⁹ is confronting (Figures 6f–j). Widespread areas warm by more than 5.0°C, particularly over the south-east, south, and western parts of the continent. This occurs in all seasons but is most dramatic in winter and spring.

Figure 7 shows the change in the frequency of the value of T_MAX⁹⁹ that currently occurs once per year on average (1981–2000). Under B1, this value increases in frequency by 2050 to 5–10 times per year in summer over large areas of northwestern Australia and by 2–4 times elsewhere (Figure 7b). In spring (Figure 7a) and winter (Figure 7d) widespread increases of 2–3 times and local increases of 3–5 times are common. By 2100, however, most of the country experiences the current T_MAX⁹⁹ 4–10 times per year in summer (Figure 7g) and 4–5 times over eastern states and 5–10 times over western states in summer and winter (Figures 7g,i).

Under A2, the changes are highly variable by 2050 (Figure 8). The frequency over the northwest increases much more than under B1 in spring, but in many other areas the increase in frequency is either similar or less. Broadly, at least in terms of adaptation to large-scale climate change, the two scenarios for 2050 are similar. The frequency of the current T_MAX⁹⁹ increases quite dramatically by 2100, however. In spring, the current once per year event occurs 5–10 times over widespread areas (Figure 8f) and 20–40 times over the northwest region of the continent. This is 3 times the change projected under B1. Similarly, in summer, the current annual event occurs more than 10 times per year over the entire continent except the extreme southeast (Figure 8g). By winter, the major changes over Western Australia (Figure 8i). Overall, it is the dramatic increase in extreme summer temperatures that is most confronting. At present, is around 36°C in inland New South Wales (NSW). Ten to twenty days a year of this extreme temperature is not something that would be easy to adapt to and would be confronting in terms of fire hazard (Pitman et al. 2007) and the impacts on flora and fauna (Hughes 2000).

3.3. Regional change in the magnitude and frequency of maximum temperature

Our projections are based on only those climate models that can capture over 80% of the observed PDF of a variable over a given region (see Figure 1). Despite this, there is still high intermodel variability in the projections of changes (i.e., skillful models may simulate the twentieth century similarly, but can still project very different future climates). This is problematic since there is no a priori reason to accept the mean change in climate of n models over any individual model when each individual model has been chosen on the basis of good performance.

Figure 9 shows the change in , T_MAX⁹⁹, and the frequency of T_MAX⁹⁹ for each season and for each emission scenario. Each bar shows the range of warming projected by the total sample of the included models. While this is by no means a measure of uncertainty in the projections, it provides a guide as to how variable the good models are in each region. The number of models in each sample varies by region but is the same in each variable shown in Figure 9. Thus, the range in projections cannot be compared across regions, but can be compared between the changes in with respect to the change in T_MAX⁹⁹ or the frequency of T_MAX⁹⁹.

In region 1, for each scenario, the models agree the most on the increase in winter , and agree least on summer warming. There is 2–3 times more variability among the models in summer compared to winter in . Warming is consistently higher in summer and spring. The amount of warming by 2100 under A2 is very clear relative to B1. The amount of uncertainty in the projections does change between scenarios but not in a way that is clearly associated with either time or emissions. In region 2, tends to be highest in winter and spring. While the higher amount of warming by 2100 under A2 is clear, the amount of variability between the models is higher than in region 1. In region 3, there is no clear seasonal pattern to warming and A2 (2100) again shows higher warming. In region 10, there is substantially less variability among the models. Most warming tends to occur in winter and spring. There appears to be a very strong response to the emission scenario with high emissions driving considerably more warming in both 2050 and 2100 than low emissions. Finally, in region 11, warming in is greater in summer and spring, leading to a pattern similar to region 1 but with substantially more warming.

Figure 9 also shows a common pattern of change in T_MAX⁹⁹ across all regions. For a given region, the increase in T_MAX⁹⁹ is therefore similar for both time periods for B1 emissions and for A2 (2050) emissions. There is generally a larger increase in T_MAX⁹⁹ under the A2 scenario by 2100; this is particularly clear in regions 10 and 11. In all regions, the amount of model agreement on the projected change in T_MAX⁹⁹ is noticeably less than for . It is also noteworthy that some of the skillful climate models simulate a reduction in the magnitude of T_MAX⁹⁹, almost always in summer and autumn and despite (Figure 9) always increasing. This reduction occurs through to 2100 under A2 emissions in regions 1 and 2 (see section 6).

The change in the frequency of T_MAX⁹⁹ is also shown in Figure 9. In regions 1, 2, and 3 the differences in the changes in frequency are small until 2100 under the A2 scenario. Within a region, there is no clear seasonality in the changes in the frequency; however, under A2 (2100) a noticeable increase in the frequency of T_MAX⁹⁹ occurs. In region 10, however, there are clear differences in the changes in T_MAX⁹⁹ with emission scenario and time period. The highest changes occur in winter and spring. Substantially larger increases occur under high emissions by 2100, but the variability among the models is also very high. For example, in 2100 (A2) one model projects an increase in the frequency of T_MAX⁹⁹ from 1 time per year to 5 times per year, while another projects over 40 times per year (summer). Region 11 provides a similar pattern of changes to region 1.

4.1. Mean minimum temperature ()

Under B1, warms in a consistent pattern (Figure 10). By 2050, warming is 1.5°–2.0°C in the continental interior and mainly 1.0°–1.5°C near the southern coast. Less than 1.0°C is simulated in winter along the southern coast (Figure 10d) and in the extreme north in autumn (Figure 10c). This warming pattern is largely reproduced in 2100, but the amount of warming in is increased by 0.5°C. The greatest warming is projected in spring (Figures 10a,f) and the least warming in winter (Figures 10d,i).

Under A2, warming in is generally higher by 2050 than under B1, but the additional warming is at most 0.5°C. By 2100, the pattern of warming in is more dramatic. In spring and summer large areas warm by 3.5°–4°C. In autumn and winter, similar increases are identified in the northern half of the continent. Farther south, warming is less, but is at least 2°C along the southern coast. Thus, at best, A2 generates an additional 0.5°C warming in isolated regions but across most of the continent; a comparison of Figures 11f–i identifies large regions of more than 2°C additional warming under A2 in comparison with B1 (Figures 10f–i).

4.2. 0.3rd percentile of minimum temperature (T_MIN^0.3)

Figure 12 shows the change in T_MIN^0.3 that occurs on average once per year (this is the cold event, at the lower end of the distribution). Under the B1 scenario T_MIN^0.3 generally warms by less than 1.5°C. Large areas warm by less than 1°C (e.g., Figure 12c). By 2100, the warming is more distinctive (Figures 12f–i) but is generally less than 2°C. There are local areas of additional warming, but these do not exceed 3°C.

Figure 13 shows the results for the high-emission scenario. By 2050, there is larger warming over many regions, particularly in spring. The patterns show the maximum changes in the northern half of Australia so this is not related to snow feedback (snow is isolated to the southeast over high orography). By 2100 the warming in T_MIN^0.3 is much higher. Large areas show an increase in this value of 3.5°–4.0°C and the smallest increases exceed 1.5°C most seasons. The increase in this event is approximately double for 2100 under A2 in comparison to B1.

Figure 14 shows the change in the frequency of T_MIN^0.3 under B1 emissions. Values less than 1.0 indicate the event becomes less frequent (0.5 means 50% occurrence per year, 0.2 means 20% occurrence per year, 0.01 means 1% occurrence per year, etc.). In general, in summer, autumn, and winter, the frequency of T_MIN^0.3 decreases such that it occurs half as often. In summer (Figure 14a) some regions in Western Australia undergo a larger change such that the current once per year minimum occurs only once per decade. By 2100 this is more confronting with more regions experiencing a change of this scale in spring (Figure 14f) and locally in other seasons.

Under A2 (Figure 15) the results are similar by 2050 to B1 (Figure 14). However, by 2100 there is a large-scale change in the frequency of the current T_MIN^0.3 that is substantially larger than under B1. Figure 15f shows, in spring for example, large areas of Australia no longer experience a daily minimum of the current value that occurs annually. Only over Victoria is the current value experienced more than once per decade. Similarly, in winter, the current annual minimum is only ever experienced by 2100 over the southern half of the continent and it is rare—mainly once every 4 years.

4.3. Regional change in the magnitude and frequency of minimum temperature

The change in shows a clear time-dependent increase in all regions (Figure 16). The substantially higher increase in is particularly visible by 2100 under A2. Given is largely constrained by incoming infrared radiation and radiative cooling at night, a seasonal variation in the amount of warming is not to be expected, and Figure 16 shows that in all regions the amount of increase in is largely independent of season. Figure 16 shows that for the change in T_MIN^0.3, there is little difference between the scenarios until 2100 under A2 where the increase is clearer. While the average increase in T_MIN^0.3 is always positive, some individual models project a reduction in the percentile value through to 2100. As with T_MAX⁹⁹, there is more variability among the models in the percentile than in the mean. This is most clear in region 3 where there is a very high variation between the models’ projection of the change in T_MIN^0.3 in all seasons, but most clearly in autumn. There is also considerable uncertainty in the projected change in the frequency of T_MIN^0.3. Figure 16 shows that models generally project a reduction in the frequency of T_MIN^0.3 as expected. In all regions except region 10 (the tropics), the average result is 0.3–0.5 (i.e., the current annual event occurs once every two to three years) under both B1 and A2 to 2050. In these regions, this decreases to a much rarer event by 2100 under A2. However, there are clear anomalies (see region 3, autumn, A2, 2100). In region 10, the reduction in the frequency of T_MIN^0.3 is less distinct and the mean response is close to 1.0 (no change) in many seasons until 2100 (A2).

5.1. Mean rainfall (P)

Figure 17 shows the change in P under B1. Increasing P over most of the continent is clear in spring (Figure 17a), summer (Figure 17b), and winter (Figure 17d) with more variable changes in autumn (Figure 17c). Given that Australia is the driest inhabited continent and is currently in severe drought, this result might appear encouraging, but the increases are small (rarely in excess of 0.5 mm day⁻¹). Through to 2100, areas of P increase contract to northern regions in spring (Figure 17f) and the eastern half of the continent in summer (Figure 17g). Drying over Victoria in spring also occurs (Figure 17f).

The projections under A2 (Figure 18) in 2050 also show widespread increasing P, particularly over the northern and eastern regions of the continent, in all seasons except winter. Over the tropics and Queensland, this exceeds 0.5 mm day⁻¹ in summer (Figure 18b). Overall, this is seen in the annual average (Figure 18e). A tendency toward drying is apparent along coastal Victoria in spring and winter and clear winter drying is visible in southwest Western Australia in winter (Figure 18d). By 2100 a pattern of spring, autumn, and winter drying has become clearly established along the eastern, southern, and western coasts, extending inland to include the Murray–Darling Basin in winter and spring and including most of Western Australia in autumn and winter. Increased P in the tropics and eastern subtropics occurs in all seasons except winter. The overall effect is visible in Figure 18j which shows coastal drying affecting all significant urban centers in the country except Brisbane. The amount of reduction in P along the coasts is not enormous (mainly a reduction of 0.1–0.5 mm day⁻¹). However, combined with increased evaporative demand and higher temperatures this is likely to exacerbate drought and further stress urban water supply.

5.2. 99.7th percentile of rainfall (P⁹⁹)

The change in P⁹⁹ reflects the change in the event that currently occurs on average once per year in the models. Figure 19 shows that the larger events increase in intensity (i.e., the size of the annual event becomes larger) over most of Australia by 2050 under B1. However, the magnitude of this increase is small and is limited to less than 10 mm day⁻¹ almost everywhere. The exception is the tropics where in summer (wet season, Figure 19) increases of 10–30 mm day⁻¹ occur. By 2100, the increase in P⁹⁹ is similar in most regions (increases of less than 10 mm day⁻¹). Across the tropics, there is again evidence of a larger increase, extending later in the year into winter (Figure 19i). A surprising result occurs in spring. There is a hint of a decrease in 2050 (Figure 19a) over Western Australia that is maintained through to 2100.

Figure 20 shows a very similar result for 2100 under A2 to 2050 under B1. By 2100, there is further intensification of rainfall in summer in the tropics (Figure 20g) exceeding 30 mm day⁻¹. The similarity in the change in P⁹⁹ under A2 is quite striking compared to low emissions except in winter. The increases in intensity seen under B1 in the north and east are not apparent under A2. However, there are also reductions in Western Australia in winter that are not apparent under B1 (cf. Figures 20i and 19i). Thus, the reductions in winter rainfall shown in Figure 18i (A2, 2100) and absent in Figure 17i (B1, 2100) are likely due to declines in higher precipitation events.

5.3. Change in the frequency of P⁹⁹

Figure 21 shows the change in the frequency of the current P⁹⁹. Values higher than 1.0 identify locations where the existing annual rainfall event becomes more frequent. Figures 21a–d show that over most of Australia the current annual event becomes more common such that it occurs 2–4 times per year. Parts of the tropics see a small decrease in the frequency of P⁹⁹. This tendency to an increased frequency of P⁹⁹ becomes clearer in 2100 (Figures 21f–i). Under A2 (Figure 22) the changes by 2050 are similar to those under B1. There are differences in detail, but these are minor. The impact of A2 to 2100 leads to decreasing frequency of the annual event over much of the east and northeast coasts, particularly in summer (Figure 22g) and autumn (Figure 22h). However, these decreases are not particularly strong.

5.4. Change in the frequency of no-rain days

Figure 23 shows a highly regional pattern of changes in rain-free days over Australia. The pattern is strongly seasonal; for example, there are more rain days by 2050 in the inland areas in spring (Figure 23a) and in the tropics in winter (Figure 23d). In summer and autumn the impacts are small and mainly coastal (Figures 23b,c). In spring and winter, there are decreases in rain days over Victoria with significant areas affected by reductions of 5–10 days. Larger changes are visible by 2100 (Figures 23–i). Inland, Queensland seems slightly less affected by 2100 with the reductions in rain days being small. In contrast, by 2100, there is a strong reduction in rain days in the northwest quarter of the continent (by 5–10 days) in summer and autumn. The increase in rain-free days over southwest Western Australia visible in spring and winter over Victoria and southern New South Wales highlights a confronting reduction in rain days, commonly exceeding 5–10 days.

The pattern of changes under A2 (Figure 24) for 2050 is very similar to the changes shown under B1 for 2100 (Figures 23f–i). These intensify further through to 2100 (Figures 24f–i). By 2100, all of Australia experiences a reduction in rain days in summer (5–10 days) and autumn (mainly 5–10 days, except over the south coast where reductions reach 10–20 days). In winter the impact is stronger. The whole southern half of Australia sees reductions of 10–20 days and southwest Western Australia and small regions of Victoria show reductions of 20–30 days. The negligible changes simulated over the tropics in winter in 2050 and 2100 under B1 and in 2050 under A2 changes by 2100 to a small (2–5) reduction in rain days.

5.5. Regional change in the magnitude and frequency of rainfall

Figure 25 shows the range of individual model projections for the mean change in precipitation for the five selected regions. It is useful to reinforce that these projections only include those models able to simulate the twentieth-century PDF of rainfall with a skill exceeding 0.8. The change in P (Figure 25) is a clear reduction in region 1, with autumn rainfall reduced most. However, the amount of reduction is very small and there are models projecting increases in all seasons in all time periods. There is clearly a tendency toward drying across region 1 with the bars illustrating the range of individual model projections largely being negative. There is also a tendency for the larger negatives to be in summer and autumn. There is no evidence of drying intensifying further into the future. In region 2 there are suggestions of an increase in P in most seasons with a hint of reduced autumn rains. Rainfall tends to increase more further into the future in the multimodel average, but both increases and decreases are projected by the models. In region 3, winter drying is common to all models in all time periods and increases in rainfall in summer, autumn, and spring occurring in all models. The apparent lack of disagreement in the projections is not due to sample size, which remains constant in all data points in each panel. In region 10, it is difficult to identify a clear response although most models suggest an increase in rainfall in most seasons. This is most clear in A2 (2100). Finally, in region 11, an increase in summer rainfall occurs in model models in all time periods but in general the changes are small.

Figure 25 shows changes in P⁹⁹. In region 1, in 2050, the change is unclear with the average of the models near zero and the range of individual models equally positive and negative. Under A2, the model means point to an increase in intensity in all seasons, but the uncertainty remains high. In region 2, P⁹⁹almost always increases pointing to an increase in the annual rainfall event by less than 10 mm. The larger increases tend to be in summer suggesting an increase in the intensity of convective events. Generally larger increases occur by 2100 under A2. In summer (A2, 2100) the increase reaches up to 20 mm. In region 3 results appear to be highly uncertain with very different patterns for 2050 depending on emission scenario. No consensus is clear on the sign of the change in P⁹⁹. In contrast, in region 10, summer and autumn events uniformly increase while winter and spring events change little. Under B1 and under A2 to 2050 the increase in P⁹⁹ ranges from 10 to 30 mm in summer and autumn with increasing variability among the models in 2100. Under A2 in 2100 there is a further increase to a model average of about 60 mm. However, it is noteworthy that though the upper range of the increases in summer and autumn rainfall is quite significant (exceeding 160 mm in 2100 in autumn under A2) the low end of each range is near zero. In region 11 increases in summer and autumn is seen in P⁹⁹ with larger increases by 2100 under A2. Changes in winter and spring are small.

The change in days with no rainfall can also be assessed (Figure 26). In region 1, the average across the models is a consistent increase in rain-free days. The increase is around 5 days yr⁻¹ except in A2 (2100) where the increase doubles to 10 days. There is a seasonal pattern to the change with a smaller increase in winter. However, the range of model estimates is considerable. In region 2, in 2050, there is an increase in rain days. This is maintained through 2100, with the exception of winter (2100, B1) and winter and spring (2100, A2). It would be hard to conclude that there was any pattern in region 3 or 4, except for a hint of an increase in rain-free days under high emissions in 2100. Similarly, in region 11, there is no clear picture that points to major changes until 2100 under A2 where in autumn, winter, and spring there is a large and quite consistent increase in rain-free days.

What causes this pattern of changes?

Explaining all of the changes identified in this paper is beyond our capacity. We can, however, link some of the changes identified in this paper with known phenomenon or reasonably established mechanisms identified in earlier studies.

1. The decline in rainfall projected over southwest Western Australia. This appears robust and is consistent with the assessment of others (Hope 2006; Christensen et al. 2007). This is an important region in Australia (IOCI 2002; Power et al. 2005), and it is useful to show that earlier findings are not affected significantly by the inclusion of weak climate models.

   In a thorough analysis of the rainfall decline in this region, Cai and Cowan (Cai and Cowan 2006) explored the extent to which the observed decline in rainfall could be explained by anthropogenic forcing and was congruent with the southern annular mode (SAM). They showed that in winter there was a relationship between the AR4 models’ simulation of the SAM and the decline in rainfall over this region. They suggested that anthropogenic forcing contributes to about 50% of the observed rainfall decline. The key connection is a southward migration of the Southern Hemisphere storm tracks (Miller et al. 2006) that cause rain-bearing frontal systems to miss the southern coastline of Australia. The AR4 models appear to show a continuation of the trends observed over this region since about 1970 (IOCI 2002) and point to ongoing regional drying through the twenty-first century (Hope 2006). The changes in the synoptic systems were explored by Hope (Hope 2006) who indicates that most models project decreases in troughs and the emergence of more high pressure systems across the region. However, the Goddard Institute for Space Studies Model E-R (GISS-ER) produced a different change in the behavior of the synoptic patterns. Perkins et al. (Perkins et al. 2007) shows that GISS-ER is relatively weak over this region in rainfall simulation and is not included in our projections over Western Australia.

2. The decline in rainfall along the southern coast of Australia. A similar mechanism involving the latitude of the storm tracks appears to cause coastal drying in the models. Miller et al. (Miller et al. 2006) explored changes in the AR4 models and showed that the multimodel average of sea level pressure over the Southern Hemisphere exhibited a decreasing trend over the pole and a compensating increase in midlatitudes. The trend in the multimodel average was associated with a poleward shift of the storm track in both hemispheres and a strengthening of the upper-level westerly flow. There are alternative mechanisms that might explain the coastal decline in rainfall. For example, Hope et al. (Hope et al. 2006) noted that changes in the hemispheric circulation could affect the frequency of storms in the region, and Fyfe (Fyfe 2003) showed that increasing greenhouse gases decreased the number of subantarctic surface cyclones. Finally, Yin (Yin 2005) found a southward shift in the strong baroclinic zones under higher greenhouse gas concentrations. All these mechanisms have the potential to decrease the likelihood of rainfall along the southern coast of Australia and all point to further declines in rainfall in the future.

3. The amplification of warming in the T_MAX⁹⁹ is associated with changes in rainfall. Because of the increases in greenhouse gases, and T_MAX⁹⁹ increase. However, T_MAX⁹⁹ increases by an additional increment because of the decline in P in some regions. Reduced rainfall tends to dry a surface and suppress the cooling effect of evaporation. Thus, there is a correlation between those regions that show the highest increases in T_MAX⁹⁹ and decreases in P (e.g., along the southern coast of Australia). This coupling of the surface energy and water balance is expected, but it highlights the need to capture changes in regional rainfall in order to project changes in regional temperatures.

4. The reduction in rain days and the increase in rainfall intensity. It is no surprise that we find an increase in the intensity of rainfall over many regions. Trenberth et al. (Trenberth et al. 2003) noted that increases in greenhouse gases leads to increases in the water-holding capacity of the atmosphere by about 7% K⁻¹ (Trenberth 1998). Soden et al. (Soden et al. 2002) indicate that models suggest that changes in relative humidity are small, hence the actual moisture content of the atmosphere should increase by roughly 7% K⁻¹ (Trenberth et al. 2003). However, because of the nonlinear dependence of moisture with temperature, the largest absolute increase in moisture would be expected in the tropics. Trenberth et al. (Trenberth et al. 2003) use this insight, coupled with changes in moisture convergence, to argue that in a warmer climate, heavy precipitation intensity should also increase by about 7% K⁻¹. However, since heavy rainfall intensity increases by 7% K⁻¹ while total rainfall amounts increase by 1%–2% K⁻¹ (Cubasch et al. 2001), this implies a reduction in low intensity and/or a reduction in the frequency of rainfall. This is what we find, and indeed Sun et al. (Sun et al. 2007) found similar results at a global scale. Thus, our results over Australia—an increase in intense rainfall, a reduction in light rainfall, and a decrease in the frequency of precipitation events—are consistent with expectations, and consistent with global analyses (Sun et al. 2007).

5. The increase in the amount and intensity of rainfall in the tropics in summer. Meehl et al. (Meehl et al. 2007) note that many climate models simulate an increase in monsoonal precipitation under warming scenarios. However, the assumption that this increase is related to the more rapid warming over land compared to the oceans and a stronger summer land–sea contrast is too simplistic. Rather, the intensification may be related to weakening of the Walker and monsoon circulations (Tanaka et al. 2005). Despite weakening of the dynamical monsoon circulation, atmospheric moisture buildup due to increased greenhouse gases and consequent temperature increase results in a larger moisture flux and a higher probability of intense rainfall (Meehl et al. 2007). Part of the monsoonal intensification may also relate to the mechanisms discussed by Trenberth et al. (Trenberth et al. 2003) and noted above.