This paper assesses the relative uncertainties from GCMs and from hydrological models in modeling climate change impact on runoff across southeast Australia. Five lumped conceptual daily rainfall–runoff models are used to model runoff using historical daily climate series and using future climate series obtained by empirically scaling the historical climate series informed by simulations from 15 GCMs. The majority of the GCMs project a drier future for this region, particularly in the southern parts, and this is amplified as a bigger reduction in the runoff. The results indicate that the uncertainty sourced from the GCMs is much larger than the uncertainty in the rainfall–runoff models. The variability in the climate change impact on runoff results for one rainfall–runoff model informed by 15 GCMs (an about 28%–35% difference between the minimum and maximum results for mean annual, mean seasonal, and high runoff) is considerably larger than the variability in the results between the five rainfall–runoff models informed by 1 GCM (a less than 7% difference between the minimum and maximum results). The difference between the rainfall–runoff modeling results is larger in the drier regions for scenarios of big declines in future rainfall and in the low-flow characteristics. The rainfall–runoff modeling here considers only the runoff sensitivity to changes in the input climate data (primarily daily rainfall), and the difference between the hydrological modeling results is likely to be greater if potential changes in the climate–runoff relationship in a warmer and higher CO2 environment are modeled.
Climate change can significantly impact on all aspects of the hydrologic cycle. Changes in climate will be amplified as a larger change in runoff (Boughton and Chiew 2007; Sankarasubramanian et al. 2001; Wigley and Jones 1985). Numerous modeling studies have been carried out in nearly every part of the world to explore climate change impact on runoff (e.g., Schaake 1990; Fowler et al. 2007b; Nyenje and Batelaan 2009). In almost all of these studies, the hydrological models are first calibrated against observed streamflow data and then driven with future climate data downscaled from GCM outputs with the same optimized parameter values, and the modeled future and historical runoff are compared to estimate the climate change impact on runoff. Understanding the uncertainty associated with the modeling results will help decision makers to interpret the reliability of the results and to use the results in a probabilistic manner.
Essentially, there are three main sources of uncertainty in climate change impact on runoff studies: uncertainty in GCM future projections, in downscaling techniques, and in hydrological modeling. The uncertainty in GCM projections comes from uncertainty in future greenhouse gas emission scenarios, as well as in the way GCMs respond to changes in atmospheric forcing, which is associated with model structure, parameterization, and spatial resolution. The differences between GCMs often result in different climate outputs, especially at regional scales. Several studies have looked at the effect of different sources of uncertainty on climate change impact on runoff. Wilby and Harris (2006) looked into the components of uncertainty in a climate change impact study on low flows in the River Thames basin using four GCMs, two emission scenarios, two statistical downscaling techniques, two hydrological model structures, and two sets of hydrological model parameters. Kay et al. (2009) compared various sources of uncertainty for flood frequency estimation under climate change for two catchments in England. Prudhomme and Davies (2009) investigated the bias in GCMs and downscaling methods propagating to river flow in four catchments in the United Kingdom using one lumped hydrological model. All of these studies concluded that the uncertainty from GCMs is generally larger than that from other sources.
Downscaling techniques are used to transform the low-resolution GCM outputs to a finer scale suitable for catchment modeling. The systematic biases associated with the different downscaling methods also add uncertainty to climate projections. There are two types of downscaling methods: statistical downscaling models and dynamic downscaling models. In addition, many impact studies rely on a simple empirical perturbation of the observations to obtain local climate change information (usually referred to as “perturbation method” or “delta-change approach”). Each method has its advantages and limitations (Fowler et al. 2007a; Wilby and Wigley 1997). Chiew et al. (2010) compared the future runoff predicted using five different downscaling methods on eight unimpaired catchments in Australia and van Roosmalen et al. (2010) compared different dynamic downscaling models in predicting the hydrological impact in Denmark. They suggested that the differences between the modeled future runoff using the different downscaled rainfall can be significant but less significant than using different GCM projections.
The uncertainty in hydrological modeling normally arises from model structure and parameterization. Chiew (2006) and Jones et al. (2006) showed that several hydrological models estimate that a 1% change in mean annual rainfall in Australia resulted in a 2%–3% change in mean annual streamflow. Jiang et al. (2007) found that the hydrological models have similar capabilities in reproducing historical runoff. However, the difference between predicted runoff from hydrological models increases if driven by hypothetical climate change scenarios. Dibike and Coulibaly (2007) also found that using the same downscaled series with different hydrological models leads to different changes in mean river flow magnitude. Recent studies comparing the sources of uncertainty suggest that the hydrological models have a relatively minor impact on the results of hydrological simulations driven with climate projections (Kay et al. 2009; Wilby and Harris 2006) but can vary significantly between catchments (Prudhomme and Davies 2009). Unlike the GCMs, which attempt to model changes to large-scale circulation in response to higher CO2 and temperature, the hydrological modeling in most studies only reflect the sensitivity of runoff to changes in the climatic inputs without considering potential changes in the rainfall–temperature–runoff relationship in a warmer and higher CO2 environment.
The majority of the uncertainty studies reported in the literature are based on relatively small study areas with only a couple of catchments, few GCMs, and a couple of hydrological models, and the results are usually catchment specific. This paper presents modeling results for the large and important agricultural region of southeast Australia. The high-resolution modeling (~5-km gridded) is carried out using lumped conceptual rainfall–runoff models for each grid cell in this large area (1.3 million km2). The study uses 15 GCMs, one empirical perturbation downscaling method, and five widely used hydrological models. The study looks at relative differences between GCM projections and the difference between runoff estimates from five hydrological models in response to these projections across the study area. The main objective of this paper is to compare and quantify the relative uncertainties (or range of results) in modeling the climate change impact on streamflow, in terms of several salient runoff characteristics, from two major sources: GCMs and hydrological models, in the way they are generally applied in most large-scale climate change impact studies.
2. Study area and data
a. Study area
The study area in southeast Australia is about 1.3 million km2 and covers about 20% of continental Australia (Fig. 1). The region generates more than half of Australia’s agricultural income and more than half of Australia’s population lives in the southern and eastern parts of the region. Australia’s two largest cities, Sydney and Melbourne, and the national capital, Canberra, are located in the region. The study area includes the whole of Victoria and the Australian Capital Territory and parts of Queensland, New South Wales, and South Australia. The entire Murray–Darling River basin and the Southeast Coast drainage divisions are in the study area, with the Great Dividing Range separating the two drainage divisions.
There is a clear east-to-west rainfall and runoff gradient and most of the runoff is generated from the upland catchments in the southeast and along the east coast. The climate near the coast is temperate and becomes semiarid and arid farther inland toward the northwest and west. Summer runoff dominates in the northern half of the region and winter runoff dominates in the southern half.
b. Climate data
1) Historical climate data
In this study, ~50 000 0.05° cells representing gridded daily time series of rainfall and potential evapotranspiration (PET) for 114 yr from July 1895 to June 2009 are used as observed historical climate data. The source of this dataset is the SILO Data Drill (http://www.longpaddock.qld.gov.au/silo) of the Queensland Department of Natural Resources and Water (Jeffrey et al. 2001). The SILO Data Drill provides surfaces of daily rainfall and other climate variables for 0.05° (~5 km) grids across Australia, interpolated from point measurements made by the Australian Bureau of Meteorology. Daily potential evapotranspiration is calculated from the 0.05° incoming solar radiation, maximum and minimum temperature, and actual vapor pressure using Morton’s wet environment (or equilibrium evaporation or areal potential evaporation) algorithms (Chiew and McMahon 1991; Morton 1983). The PET used here is conceptually the upper limit of actual evaporation in the hydrological models.
2) Future climate data
The future climate data are sourced from 15 GCMs in the Fourth Assessment Report (AR4) of the Intergovernmental Panel on Climate Change (IPCC) (Solomon et al. 2007; CSIRO and Australian Bureau of Meteorology 2007). The GCM outputs are obtained from the World Climate Research Programme’s (WCRP’s) phase 3 of the Coupled Model Intercomparison Project (CMIP3) archived at Program for Climate Model Diagnosis and Intercomparison (PCMDI) (http://www-pcmdi.llnl.gov). The 15 GCMs are listed in Table 1 and described further on the PCMDI Web site (http://www-pcmdi.llnl.gov/ipcc/model_documentation/ipcc_model_documentation.php). The 1870–2100 monthly GCM outputs for the twentieth century and the multirun ensembles driven by the Special Report on Emissions Scenarios (SRES) A1B and A2 scenarios are used to obtain the long-term trend for pattern scaling [see section 3b(2)]. The daily rainfall series for two 20-yr periods—(i) 1981–2000 [1980–99 for the two National Center for Atmospheric Research (NCAR) GCMs because of the data availability] in the twentieth century and (ii) 2046–65 for the A1B scenario—are used to perturb the daily rainfall distribution [see section 3b(2)].
Daily observed streamflow data from 232 catchments (50–2000 km2) are used in this study to calibrate the hydrological models. The locations of the catchments are shown in Fig. 1. All the catchments are largely unregulated and are located mainly in the upland areas in the southeast and along the east coast where most of the runoff is generated. The streamflow data are sourced from another study (Vaze et al. 2011a), which compiled, checked, and cleaned the original data from the state water agencies. More than half of the streamflow gauges have been recording since the early 1960s, and the daily streamflow data have been compiled for the period of available data. For the hydrological modeling undertaken in this study, streamflow data from 1975 to 2008 are used in the model calibration. The large majority of the gauges have less than 10% missing data over this period.
The hydrological models used here are five lumped conceptual daily rainfall–runoff models [described in section 3a(1)]. They are first calibrated against observed streamflow data [section 3a(2)] and then driven using historical and future climate data with the same optimized parameter values to model the historical and future runoff [section 3a(3)]. The future climate series are obtained by scaling the historical data, informed by outputs from 15 GCMs (section 3b). The modeled future and historical runoff are then compared to estimate the climate change impact on runoff (section 3c). The range of results from the 15 GCMs and from the five hydrological models is then compared to assess the relative uncertainties.
a. Modeling runoff
1) Hydrological models
The five lumped conceptual daily rainfall–runoff models used are the Australian Water Balance Model (AWBM; Boughton 2004); Identification of Unit Hydrographs and Component Flows from Rainfall, Evaporation, and Streamflow Data (IHACRES; Croke et al. 2006); Sacramento (Burnash et al. 1973); Simplified Hydrolog (SIMHYD; Chiew et al. 2002); and Soil Moisture Accounting and Routing Model (SMARG; Vaze et al. 2004). The model versions used here are very similar to those described in the above references. The models are typical of lumped conceptual rainfall–runoff models, with interconnected storages and algorithms that mimic the hydrological processes used to describe movement of water into and out of storages. For the application here, the numbers of parameters calibrated are 6 for AWBM, 7 for IHACRES, 14 for Sacramento, 6 for SIMHYD, and 8 for SMARG. All five models have been widely used in Australia, Europe, and the United States, including for regionalization studies to predict runoff in ungauged catchments and for climate impact and land-use change studies (Boughton and Chiew 2007; Gan and Burges 2006; Oudin et al. 2008; Post et al. 1996, Reichl et al. 2009; Tuteja et al. 2007, Vaze et al. 2011b).
2) Model calibration
The hydrological models are calibrated against 34 yr of observed daily streamflow data from 1975 to 2008 in the 232 catchments. In the model calibration, the hydrological model parameters are optimized to maximize the Nash–Sutcliffe efficiency (NSE)-bias objective function, which is a weighted combination of daily Nash–Sutcliffe (Nash and Sutcliffe 1970) efficiency and a logarithmic function of bias given by
where NSE is the Nash–Sutcliffe efficiency of daily streamflow and B is the bias (the sum of daily modeled error divided by the sum of observed daily streamflow) (Viney et al. 2009). The coefficients of this equation control the severity and shape of the bias constraint penalty.
The NSE is commonly used in hydrological modeling to describe the agreement between the modeled streamflow and observed streamflow time series. An NSE value of 1.0 indicates that all the modeled daily streamflow are the same as the observed daily streamflow, and an NSE value of less than zero indicates that the model simulations are poorer than simply using the mean observed daily streamflow as the streamflow estimate for every single day. The bias, B, is included in the objective function to ensure that the total modeled streamflow is similar to the total observed streamflow. The shuffled complex evolution global optimization method (Duan et al. 1993) followed by a local optimization method (Rosenbrock 1960), with multiple starting parameter sets, is used to calibrate the models.
Table 2 summarizes the calibration results for the five hydrological models in the 232 catchments. The relative difference (second to sixth column in Table 2) is the percentage difference between the modeled and observed runoff characteristic relative to the observed runoff characteristic. The median (value at the top) and the range of 10th to 90th percentile values (the values in brackets below the median value) for each of the characteristics from the 232 catchments are shown in Table 2. The relative differences are small for mean annual runoff (Table 2, column 2, below 5% for all the models in all the 232 catchments) but larger for other runoff characteristics. This is partly because of the objective function used in this study, which ensures that the total modeled streamflow is similar to the total observed streamflow. In general, all the five models perform better in estimating high-flow characteristics (column five in Table 2) compared to estimating low-flow characteristics (column six in Table 2). The calibration NSE values are generally highest for the Sacramento model (most likely because it has the most number of parameters) followed by IHACRES and SMARG models, and then the AWBM and SIMHYD models. However, the differences between the results from the five models are small, and the calibration NSE values from the different models generally differ by less than 0.1. The model calibration results are generally satisfactory, with NSE values greater than 0.6 in about 90% of the catchments and greater than 0.8 in about 40% of the catchments for all the five models.
3) Modeling runoff across the region
Each of the five hydrological models is applied separately with the historical observed climate to estimate 114-yr daily runoff from July 1895 to June 2009 (referred to as “historical runoff”) for the ~50 000 grid cells across southeast Australia and then driven with the 15 future climate variants to estimate future runoff (referred to as “future runoff”).
The same set of optimized parameter values [described in section 3a(2)] are used for all the grid cells within a gauged (calibration) catchment. The runoff for the other grid cells that do not lie within a calibration catchment is modeled using optimized parameter values from the geographically closest grid cell that lies within a calibration catchment (donor catchment). Many studies have shown that selecting a donor catchment based on spatial proximity gives similar or better modeling results in the target ungauged catchment compared to the more involved process of selecting a donor catchment based on other similarity measures (Oudin et al. 2008; Zhang and Chiew 2009). This is partly because neighboring catchments are likely to have similar climate and catchment physical characteristics, and partly because it is difficult to define meaningful catchment-average attributes and derive these attributes accurately. As most of the drier and large western part of the study region is ungauged, the calibrated parameters from the gauged catchments along the western fringe of the gauged areas are used to simulate runoff for this ungauged region (as seen in the straight line artefacts in the runoff map in Fig. 1). Nevertheless, the runoff contribution from this large western part is very small compared to the high runoff generated over the well-gauged areas in the coastal regions in the east and upland areas in the southeast. The spatial proximity regionalization method used here is suitable for this study as the main interest is in estimating the relative differences between the hydrological model results when using future projections from 15 GCMs. However, given the large regionalization distances, all the statistical results and plots are based on the modeling in the 232 gauged catchments while the spatial maps are shown for the entire study region (to provide a complete spatial picture across this important region for which most of the policy and management decisions are made as a whole). The spatial extent of the area within 100 km from any gauged catchments, where we have more confidence in the regionalization results, is also shown on all the runoff maps.
4) Historical runoff
Figure 2 and Table 3 compare the historical runoff characteristics modeled by the five hydrological models for all the grid cells within the 232 gauged catchments. The key runoff characteristics considered here are mean annual runoff, mean summer runoff [December–February (DJF)], mean winter runoff [June–August (JJA)], high daily runoff (Q1, shown as daily runoff that is exceeded 1% of the time), and a low-flow characteristic [DLF, shown as number of days per year when runoff is less than 0.1 mm—many metrics have been used to describe low-flow characteristics (McMahon and Finlayson 2003; Smakhtin 2001; Wilby and Harris 2006) and the criteria used here simply provides one example]. The values estimated by each of the five hydrological models for all the grid cells within the gauged catchments are plotted against each other as scatterplots. The closeness of the points to a 1:1 line indicates that there is a good agreement between the runoff estimates from the hydrological models.
The scatterplots comparing the modeled mean annual runoff for all grid cells within the 232 gauged catchments by the five hydrological models are shown in Fig. 2 [scatterplots for seasonal runoff (Q1 and DLF not shown) and statistics provided in Table 3]. The mean annual and seasonal runoff simulated by the five models is in good agreement with each other, with most of the points falling on or very close to the 1:1 line. However, the IHACRES results are slightly different to those from the other four models. This is likely because IHACRES is a simpler model based on unit hydrograph principles and it uses a rainfall scaling factor to convert observed rainfall to effective rainfall. There is a bigger difference between the models in the Q1 estimates compared to mean annual and seasonal runoff (see Table 3), but the results from the five models are generally in good agreement with each other. There are considerable differences in DLF values simulated by the five models. This is likely because the hydrological models are not calibrated to low-flow characteristics in this study and, therefore, it is not surprising to find poorer performance in reproducing low flows compared to their capability in reproducing mean runoff and high flows.
b. Climate projections
1) Global climate models
GCMs are widely used and arguably the best available tools for generating future climate projections. Several studies have assessed GCMs for their ability to reproduce observed historical rainfall within Australia based on different criteria (Perkins et al. 2007; Smith and Chandler 2010; Suppiah et al. 2007; Vaze et al. 2011c; Watterson 2008). These studies demonstrate that although the GCMs can generally reproduce the observed spatial mean annual rainfall patterns in southeast Australia, the rainfall amounts can be very different from the observed values. This study uses the result from Vaze et al. (2011c), which ranked the GCMs based on the combination of RMSE of mean annual rainfall across southeast Australia and NSE of daily rainfall distribution. The skill score of each GCM is presented as a number between 0 and 1 shown in the last column in Table 1 (a higher score represents better-performing GCMs). These scores are used in section 4 to investigate whether eliminating “poorer” GCMs or weighting to favor “better” GCMs can reduce the magnitude of uncertainty in climate change impact on runoff results.
2) Scaling daily historical climate sequence to obtain future daily climate sequence
Fifteen future climate variants, each with 114 yr of daily rainfall and PET data, are developed by scaling the observed historical daily rainfall and PET sequence for the ~50 000 0.05° grid cells [described in section 2b(1)], based on the analyses of 15 GCM outputs [described in section 2b(2)] for a 1°C global warming (median of the projected increase in global average surface air temperature by 2030 relative to 1980–99) (Solomon et al. 2007; CSIRO and Australian Bureau of Meteorology 2007). The daily climate series for the historical and future climate have the same length of data (114 yr) and the same sequence of daily climate.
The changes from the GCMs are expressed as the “change per degree global warming” by fitting a linear regression through the local GCM variables against global average temperature simulated from the same GCM (Chiew et al. 2009). The “change in climate variables per degree global warming” approach adopted in this study decouples the model’s response from the greenhouse gas emissions scenarios. The resultant change per degree of global warming can be rescaled by a given amount of global warming to produce a pattern of change that would apply for a given future date and global warming scenario. However, it will not capture any systematic model responses that are not varying linearly with global temperature. This pattern scaling approach has been shown to be robust and gives similar results as comparing a GCM simulation for a future period relative to a historical period (Mitchell 2003; Suppiah et al. 2007).
An empirical perturbation method (daily scaling method), together with seasonal pattern scaling, is used to scale the historical daily climate series, informed by the GCMs, to obtain a future climate series. This method is used to scale the rainfall series, whereas only seasonal pattern scaling is used to scale the PET series. This approach has been widely used in a number of climate impact assessment projects such as Harrold and Jones (2003), Prudhomme et al. (2002), Commonwealth Scientific and Industrial Research Organisation (CSIRO) Sustainable Yields projects (http://www.csiro.au/partnerships/SYP.html), the South Eastern Australian Climate Initiative (SEACI; http://www.seaci.org; Chiew et al. 2009), and New South Wales future climate and runoff projections (Vaze and Teng 2011), and has been described in detail in Chiew et al. (2009). This approach takes into account the changes in daily distribution as well as the changes in seasonal means. This is important for rainfall (and runoff) because many GCMs indicate that future high rainfall is likely to be more intense, even in some regions where projections indicate a decrease in mean seasonal or annual rainfall. As the high-rainfall events generate large runoff, the use of the pattern scaling method alone, which assumes the entire rainfall distribution changes in the same way, would lead to an underestimation of the high runoff as well as the mean annual runoff.
Like all downscaling methods, the daily scaling method has its advantages and limitations. The daily scaling method uses the historical rainfall sequence (but different amounts) to represent the future. It does not consider potential changes to other rainfall characteristics, including the sequencing and timing of rainfall events (Chiew et al. 2009; Mpelasoka and Chiew 2009). On the other hand, the advantage of using the daily scaling method is that it is robust and relatively simple to implement and can be used to scale long-term (114 yr in this study) observed rainfall directly (which takes into account the natural variability associated with the historical record). The daily scaling method can be used for hydrological impact assessment studies over very large regions, particularly when the main considerations are changes to seasonal and annual catchment water yield. Additionally, runoff impact studies comparing the results when using downscaled rainfall from different downscaling methods (Chiew et al. 2010; Quintana Segui et al. 2010) suggest that there is no consistently superior downscaling method when considering a fuller range of runoff characteristics.
Figure 3 shows the future rainfall projections from the 15 GCMs. The projections are expressed as the percentage change in mean annual rainfall per degree global warming. In Fig. 3, the GCMs are shown based on future mean annual rainfall (averaged across the region) with the driest GCM at the top-left corner and the wettest GCM at the bottom-right corner. The same sequence is used to present the results for runoff (next section). There is a relatively large variability in the results from the 15 GCMs with the values varying from −8% to +4% when averaged across the whole region. There is also a large difference in the spatial pattern of percentage change in future rainfall predicted by the 15 GCMs. However, there is strong agreement between the GCMs in the southern part of the region with the majority of GCMs projecting a decrease in rainfall in this region.
c. Estimating climate change impact on runoff
The 15 future climate variants (described in section 3b) are used to drive the five hydrological models (described in section 3a), using the same optimized parameter values, to estimate 114 yr of future daily runoff. This commonly used approach for modeling future runoff contributes to the uncertainty from hydrological models. The underlying assumption of this approach is the stability of the dominant hydrological processes under a changed climatic condition (Vaze et al. 2010). Although most of the hydrological impact studies are based on a similar assumption (e.g., Fowler et al. 2007b; Maurer and Duffy 2005; Prudhomme and Davies 2009; Wilby 2005), more research is needed because of the possible nonstationarity of the rainfall–runoff relationship (Milly et al. 2008).
The 114 yr of modeled future runoff is then compared with the 114 yr of modeled historical runoff to assess future climate impact on runoff. The climate change impact on runoff is assessed here as change in key runoff characteristics described earlier.
4. Results and discussions
The results for change in future mean annual runoff, Q1, and DLF are presented in Figs. 4, 5, and 6, respectively. The spatial patterns of variability in future mean annual runoff (for a 1°C global warming) modeled by the five hydrological models, informed by projections from the 15 GCMs, are shown in Fig. 4. Each row of maps in Fig. 4 shows the percentage change in future mean annual runoff estimated by the five hydrological models for rainfall projections from the 15 GCMs and each column shows the same results for each of the hydrological models when using rainfall projections from the 15 GCMs. Two thirds of the modeling results show a reduction in future runoff averaged across the study area. Although there is variability in the runoff reduction or increase across different parts of the region, a large majority of the results show a decrease in mean annual runoff in the southern part of the region. As expected, the variability in future mean annual runoff across the study area is much larger between the 15 GCMs (for any of the five hydrological models) compared to the variability in runoff between the five hydrological models (for any of the 15 GCMs).
The plots in Fig. 7 show the range (averaged across all the calibration catchments) for mean annual runoff (top row), mean summer runoff (second row), mean winter runoff (third row), Q1 (fourth row), and DLF (bottom row). The percentage changes in annual and seasonal rainfall are also shown on the plots. The left-side plots show the difference in the results between the 15 GCMs for each hydrological model and the right-side plots show the difference in the results between the five hydrological models for each GCM (the GCMs are shown in the same sequence as in Fig. 3 from the driest projection to the wettest projection averaged across entire study area). The numbers in these plots show values averaged across all the grid cells in the 232 catchments. For Q1 and DLF, the values are estimated by weighting the change for each grid cell by the amount of mean annual runoff in that grid cell. As a result, the values are influenced by the wetter catchments where the rainfall and runoff are higher.
The range of results from the five hydrological models informed by a particular GCM is relatively small (less than 7% difference between the minimum and maximum values; see top-right plot in Fig. 7). In contrast, the range of results for any one of the hydrological models informed by the 15 GCMs is considerably larger, with a 28%–35% difference between the minimum and maximum values. The summer and winter percentage change in runoff results (left-side plots in second and third rows in Fig. 7) are similar to the annual results with considerably larger variability between the GCMs than between the hydrological models. The range of the results is also generally higher for the extremely dry future rainfall projections [e.g., Institut Pierre Simon Laplace (IPSL) and Centre National de Recherches Météorologiques (CNRM) results in Fig. 7].
The plots in Fig. 7 also show that the range of results for summer is larger than the range of results for winter (spatial maps not shown). There are almost as many GCMs projecting a decline in summer rainfall (and therefore summer runoff) as the GCMs projecting an increase in summer rainfall (and therefore runoff). In contrast, there is a strong agreement between the GCMs on a decline in future winter rainfall (and therefore runoff, particularly in the southern parts of the region where most of the runoff occurs in winter). This projection of winter rainfall decline is also consistent with the patterns observed during the past several decades of the expansion in the Hadley cell and the poleward shift of autumn and winter storm tracks, and with the expected changes in the large-scale atmospheric and oceanic drivers of rainfall in the southern parts of the region (Cleugh et al. 2011).
The range in the climate change impact on Q1 results from the 15 GCMs (29%–34%) compared to the range of the results from the five hydrological models (less than 3%) are similar to that of the mean annual and seasonal runoff (Figs. 5 and 7). This is to be expected as the high-flow events have the largest contribution to the annual and seasonal totals. It is interesting to note that the decline in future Q1 is smaller than the projected decline in the mean annual runoff. This is because although most of the GCMs project a drier future for the region, a large majority of the GCMs indicate that the high rainfall will become more intense (Fig. 7).
The majority of the results also indicate that the number of days with runoff less than 0.1 mm (DLF) will increase. The range of results for this DLF low-flow metric is generally larger than the range for the other runoff characteristics, and the relative difference between the results from the 15 GCMs (21–32 days) and the five hydrological models (2–12 days) is smaller than that for the other runoff characteristics. This is most likely because it is difficult to model low flow and the hydrological models are conceptualized and calibrated mainly to model the medium and high flows.
Figure 8 shows the range of projected changes in future mean annual runoff estimated by the five hydrological models for each of the 15 GCMs. The change in future runoff (y axis) is plotted against the GCM skill scores (x axis) based on their ability to reproduce the historical rainfall (Vaze et al. 2011c) shown in Table 1. The consistency in the future rainfall (and therefore runoff) projections between the better GCMs is no better than when results from all the GCMs are pooled together. This analysis was repeated with skill scores from three other studies, which assessed the GCMs using criteria that are slightly different (Smith and Chandler 2010; Suppiah et al. 2007; Watterson 2008) and those results also indicate little agreement between the GCMs with higher skill scores. The results here therefore indicate that using only the GCMs that best reproduce the historical observed climate or weights to favor these GCMs do not reduce the range of results, and the uncertainty associated with the GCM projections will still be much larger than that from hydrological models. Nevertheless, using different skill metrics to rank the GCMs, particularly with the improvements in GCMs through CMIP Phase 5, may reduce the uncertainty in the future climate projections (Mote et al. 2011; Knutti et al. 2010).
This paper assesses the relative uncertainties sourced from GCMs and from hydrological models in the way they are generally used in large-scale climate change impact on runoff studies. The modeling is carried out across a large 1.3-million-km2 region in southeast Australia. Five lumped conceptual daily rainfall–runoff models are used to model runoff using historical daily rainfall and PET series and using future climate series obtained by empirically scaling the historical climate series informed by simulations from 15 GCMs. The majority of the GCMs project a drier future for this region, particularly in the southern parts, and this is amplified as a bigger reduction in the runoff.
As expected, the modeling results indicate that the uncertainty sourced from the GCMs is much larger than the uncertainty in the rainfall–runoff models. The variability in the climate change impact on runoff results for one rainfall–runoff model informed by 15 GCMs (about 28%–35% difference between the minimum and maximum results for mean annual, mean seasonal, and high runoff) is considerably larger than the variability in the results between the five rainfall–runoff models informed by one GCM (less than 7% difference between the minimum and maximum results). The difference between the rainfall–runoff modeling results is larger in the drier regions for scenarios of large declines in future rainfall and in the low-flow characteristics.
Like most climate change impact studies, the rainfall–runoff modeling here considers only the runoff sensitivity to changes in the input climate data (primarily daily rainfall). The difference between hydrological modeling results may be greater if potential changes in the climate–runoff relationship and land–vegetation–atmosphere feedbacks in a warmer and higher CO2 environment are modeled. Nevertheless, the uncertainty in the hydrological modeling is still likely to be considerably smaller than the uncertainty in future climate projections from GCMs. The biggest improvement in predictions of future water availability and runoff characteristics will therefore need to come from better future rainfall projections from global and regional climate models and downscaling models.
This work is carried out within the CSIRO Water for a Healthy Country National Research Flagship and is funded by the South Eastern Australian Climate Initiative (SEACI; www.seaci.org). The hydrological modeling is undertaken using the Catchment Water Yield Estimation Tool (CWYET) modeling framework. The authors thank Dewi Kirono, Janice Bathols, and David Kent for their help in preparing climate projections. The authors would also like to thank David Post and Cuan Petheram for reviewing the paper and providing useful comments.