Potential Improvements To Statistical Downscaling of General Circulation Model Outputs To Catchment Streamflows With Downscaled Precipitation and Evaporation
ABSTRACT: An existing streamflow downscaling model (SDM(original)), was modified with the outputs of a precipitation downscaling model (PDM) and an evaporation downscaling model (EDM) as additional inputs, for improving streamflow projections. For this purpose, lag 0, lag 1 and lag 2 outputs of PDM were individually introduced to SDM(original) as additional inputs, and then it was calibrated and validated. Performances of the resulting modified models were assessed using Nash-Sutcliffe efficiency (NSE) during calibration and validation. It was found that the use of lag 0 precipitation as an additional input to SDM(original) improves NSE in calibration and validation. This modified streamflow downscaling model is called SDM(lag0_preci). Then lag 0, lag 1 and lag 2 evaporation of EDM were individually introduced to SDM(lag0_preci) as additional inputs and it was calibrated and validated. The resulting models showed signs of over-fitting in calibration and under-fitting in validation. Hence, SDM(lag0_preci) was selected as the best model. When SDM(lag0_preci) was run with observed lag 0 precipitation, a large improvement in NSE was seen. This proved that if precipitation produced by the PDM can accurately reproduce the observations, improved precipitation predictions will produce better streamflow predictions.
Título original
Potential Improvements to Statistical Downscaling of General Circulation Model Outputs to Catchment Streamflows With Downscaled Precipitation and Evaporation
ABSTRACT: An existing streamflow downscaling model (SDM(original)), was modified with the outputs of a precipitation downscaling model (PDM) and an evaporation downscaling model (EDM) as additional inputs, for improving streamflow projections. For this purpose, lag 0, lag 1 and lag 2 outputs of PDM were individually introduced to SDM(original) as additional inputs, and then it was calibrated and validated. Performances of the resulting modified models were assessed using Nash-Sutcliffe efficiency (NSE) during calibration and validation. It was found that the use of lag 0 precipitation as an additional input to SDM(original) improves NSE in calibration and validation. This modified streamflow downscaling model is called SDM(lag0_preci). Then lag 0, lag 1 and lag 2 evaporation of EDM were individually introduced to SDM(lag0_preci) as additional inputs and it was calibrated and validated. The resulting models showed signs of over-fitting in calibration and under-fitting in validation. Hence, SDM(lag0_preci) was selected as the best model. When SDM(lag0_preci) was run with observed lag 0 precipitation, a large improvement in NSE was seen. This proved that if precipitation produced by the PDM can accurately reproduce the observations, improved precipitation predictions will produce better streamflow predictions.
Potential Improvements To Statistical Downscaling of General Circulation Model Outputs To Catchment Streamflows With Downscaled Precipitation and Evaporation
ABSTRACT: An existing streamflow downscaling model (SDM(original)), was modified with the outputs of a precipitation downscaling model (PDM) and an evaporation downscaling model (EDM) as additional inputs, for improving streamflow projections. For this purpose, lag 0, lag 1 and lag 2 outputs of PDM were individually introduced to SDM(original) as additional inputs, and then it was calibrated and validated. Performances of the resulting modified models were assessed using Nash-Sutcliffe efficiency (NSE) during calibration and validation. It was found that the use of lag 0 precipitation as an additional input to SDM(original) improves NSE in calibration and validation. This modified streamflow downscaling model is called SDM(lag0_preci). Then lag 0, lag 1 and lag 2 evaporation of EDM were individually introduced to SDM(lag0_preci) as additional inputs and it was calibrated and validated. The resulting models showed signs of over-fitting in calibration and under-fitting in validation. Hence, SDM(lag0_preci) was selected as the best model. When SDM(lag0_preci) was run with observed lag 0 precipitation, a large improvement in NSE was seen. This proved that if precipitation produced by the PDM can accurately reproduce the observations, improved precipitation predictions will produce better streamflow predictions.
Potential improvements to statistical downscaling of general
circulation model outputs to catchment streamflows with downscaled precipitation and evaporation D. A. Sachindra & F. Huang & A. Barton & B. J. C. Perera Received: 12 December 2013 / Accepted: 18 September 2014 #Springer-Verlag Wien 2014 Abstract An existing streamflow downscaling model (SDM (original) ), was modified with the outputs of a precipitation downscaling model (PDM) and an evaporation downscaling model (EDM) as additional inputs, for improving streamflow projections. For this purpose, lag 0, lag 1 and lag 2 outputs of PDM were individually introduced to SDM (original) as additional inputs, and then it was calibrated and validated. Performances of the resulting modified models were assessed using Nash- Sutcliffe efficiency (NSE) during calibration and validation. It was found that the use of lag 0 precipitation as an additional input to SDM (original) improves NSE in calibration and valida- tion. This modified streamflow downscaling model is called SDM (lag0_preci) . Then lag 0, lag 1 and lag 2 evaporation of EDM were individually introduced to SDM (lag0_preci) as addi- tional inputs and it was calibrated and validated. The resulting models showed signs of over-fitting in calibration and under- fitting in validation. Hence, SDM (lag0_preci) was selected as the best model. When SDM (lag0_preci) was run with observed lag 0 precipitation, a large improvement in NSE was seen. This proved that if precipitation produced by the PDMcan accurately reproduce the observations, improved precipitation predictions will produce better streamflow predictions. 1 Introduction General Circulation Models (GCMs) are considered as the most widely used tools for projection of global climate into the future (Bastola and Misra 2013). GCMs use the funda- mentals of physics for describing the global climate. When forced with plausible scenarios of future greenhouse gas (GHG) concentrations, they are capable of projecting the global climate hundreds of years into the future (Tripathi et al. 2006). The spatial resolution of a current GCM is in the order of a few hundred kilometres and GCMs are capable of correctly simulating the global and continental climate. However, since GCMs coarsely represent the topography and land use, they are unable to correctly simulate the catch- ment scale climate. Therefore, statistical and dynamic down- scaling techniques are used for translating the coarse scale information in the GCM outputs to catchment scale hydroclimatic information (e.g. Hertig and Jacobeit (2008); Can et al. (2011); Samadi et al. (2012); Flint and Flint (2012)). In dynamic downscaling, a Regional Climate Model (RCM) is nested in a GCM for simulation of regional climate (Murphy 1998). In this process, initial and lateral boundary conditions to the RCM are provided by the GCM at multiple vertical and horizontal levels (Wilby and Fowler 2011). In the RCM, the information provided by the GCM is processed using the fundamentals of physics of the atmosphere, and hence regional patterns of climate variables are generated (Rummukainen 2010). When the difference of the spatial resolutions between the GCM and the RCM is high, multiple nesting which involves downscaling starting from a larger domain with coarser spatial resolution and progressively mi- grating to smaller domains with finer spatial resolution (until the desired spatial resolution is attained) is performed (Rojas 2006). RCMs are also to simulate the climate over a catchment at a spatial resolution of a few kilometres (e.g. 550 km). Dynamic downscaling techniques can produce spatially con- tinuous fields of climatic variables while maintaining the spatial coherence (Maurer et al. 2008). Since RCMs operate at relatively higher spatial resolutions, the improved D. A. Sachindra (*) : F. Huang : A. Barton : B. J. C. Perera College of Engineering and Science, Victoria University, Footscray Park Campus, P.O. Box 14428, Melbourne, Victoria 8001, Australia e-mail: sachindra.dhanapalaarachchige@live.vu.edu.au A. Barton Federation University, PO Box 663, Ballarat, Victoria 3353, Australia Theor Appl Climatol DOI 10.1007/s00704-014-1288-7 representation of topographic features such as mountains, water bodies and other land use features aid in the simulation of local climate more accurately (Rummukainen 2010; Horvath et al. 2012). However, RCMs suffer from high com- putational costs associated with their complex physics-based structure (Haas and Pinto 2012). The computational cost of a dynamic downscaling exercise largely increases with the spa- tial resolution of the RCMand the extent of the domains (Qian and Zubair 2010). Owing to the high computation costs asso- ciated with dynamic downscaling, use of multiple GCMs and GHG emission scenarios in a dynamic downscaling study may not be always feasible. Also, the simulations produced by a RCM during its spin-up period (time taken to attain climate equilibrium) which can be in the order of a few months or a few years (Denis et al. 2002) are discarded. Statistical downscaling techniques are dependent on the statistical relationships developed between the GCM outputs and the catchment-scale hydroclimatic variables (Fowler et al. 2007). Owing to the simplicity, statistical downscaling tech- niques are associated with much less computational costs (Bedia et al. 2013). Statistical downscaling techniques can be used to produce projections of hydroclimatic variables such as streamflows, leaf wetness etc. which are not simulated by GCMs. Also, unlike dynamic downscaling, statistical met hods enabl e downscal i ng of GCM out put s t o hydroclimatic variables at specific points in the catchment. On the other hand, statistical downscaling techniques are not able to produce spatially continuous fields of hydroclimatic variables. For proper calibration and validation of a statistical downscaling model, long series of observations are preferred. This is because a long series of observations can possibly expose the downscaling model to the full variance of the local scale climate and make the downscaling model more robust (Sachindra et al. 2014a). In statistical downscaling, it is as- sumed that the statistical relationships determined between the large-scale atmospheric variables (e.g. GCM outputs, reanal- ysis outputs) and catchment scale climatic variables (e.g. precipitation) for the past climate are also valid for the chang- ing climate in the future (Benestad et al. 2008). This assump- tion is similar to that of the validity of parameterisation schemes in the RCMs for future climate. According to Wilby et al. (2004), statistical downscaling techniques can be classi- fied into three categories; (1) regression techniques (e.g. Tareghian and Rasmussen 2013), (2) weather classification techniques (e.g. Gutirrez et al. 2013) and (3) weather gener- ation techniques (e.g. Wilks 1999). Regression based statisti- cal downscaling techniques are regarded as the most widely used statistical downscaling techniques (Nasseri et al. 2013). Reliable forecasts of streamflows are useful in the manage- ment of water resources in a catchment. These management activities include flood control, water supply, hydroelectricity generation and also the maintenance of environmental flows. Therefore, it is realised that the projection of catchment scale streamflows into the future under changing climate is of high importance. Landman et al. (2001) used bias-corrected moisture and circulation outputs of a GCM in canonical correlation analysis (CCA)-based statistical downscaling model for simulation of seasonal streamflows at 12 locations in South Africa. Cannon and Whitfield (2002) used multi-linear regression (MLR) and artificial neural networks (ANN) for downscaling GCM out- puts to 5-day average streamflows at 21 locations in Canada. They used variables representative of the atmospheric circula- tions at the Earths surface and the mid troposphere, tempera- ture at lower troposphere and boundary layer moisture vari- ables. In that study, it was commented that ANN was more capable compared to MLR in downscaling GCM outputs to streamflows. Hence, it was realised that non-linear techniques are able to better capture the complex association between the atmospheric variables and streamflows. Ghosh and Mujumdar (2008) used support vector machine (SVM) and relevance vector machine (RVM) regression techniques for downscaling GCM outputs to monthly streamflows at a site in India. They used principal component analysis (PCA) for derivation of PCs from GCM outputs and applied fuzzy c-mean clustering to classify the PCs into classes before introducing them to the downscaling models. It was found that the RVM-based down- scaling model was less prone to over-fitting unlike the SVM- based downscaling model which suffered severe over-fitting. Tisseuil et al. (2010) employed ANN, generalized additive models, aggregated boosted trees and generalized linear models for downscaling NCEP/NCAR reanalysis and GCM outputs to daily streamflows, at 51 stations in France. In that study, using hierarchical ascending cluster (HAC) analysis, predictor vari- ables were separated into several clusters based on their simi- larity, and then the first PCof each cluster was used as inputs to the downscaling model. It was found that aggregated boosted trees (non-linear technique) were more efficient in downscaling large-scale atmospheric variables to streamflows. Sachindra et al. (2013) used least square support vector machine regres- sion (LS-SVM-R) and MLR to downscale NCEP/NCAR re- analysis outputs to monthly streamflows at a station in north- western Victoria, Australia. Unlike in the former streamflow downscaling studies, they used soil moisture content as an input to the downscaling model and found that it is highly influential on streamflows. Furthermore they commented that both LS-SVM-R and MLR produced very comparable perfor- mances. However, both techniques were unable to simulate the extremes of streamflow. The main advantage of downscaling GCM outputs directly to streamflows in a catchment is that it allows the quick estimation of streamflows without the need of a hydrologic model. However, direct downscaling of GCM outputs to streamflows can only be used for simulation of unregulated flows in a catchment. Since many complex hydrological pro- cesses (e.g. infiltration) are not explicitly modelled in a direct D.A. Sachindra et al. streamflow downscaling exercise, errors to streamflow simu- lations may be introduced. Precipitation and evaporation are two highly influential climatic variables on the availability of water resources in a catchment. Precipitation is regarded as the main driver of streamflow in many catchments. In past literature, there are many case studies on downscaling GCM outputs to precipi- tation at the catchment scale (e.g. Tripathi et al. (2006), Anandhi et al. (2008), Timbal et al. (2009), Jeong et al. (2012), Sachindra et al. (2014a)). Evaporation is one of the many processes responsible for the loss of water from a catchment. Some examples for downscaling GCM outputs to evaporation are found in the studies of Timbal et al. (2009), Yang et al. (2012), Goyal and Ojha (2012). As stated earlier, Sachindra et al. (2013) detailed the de- velopment of a statistical model for downscaling NCEP/NCAR reanalysis outputs directly to monthly streamflows at a location in north-western Victoria, Australia. The current paper provides the details of further potential improvements to that streamflowdownscaling model and also the streamflow projections produced into the future using the improved model. As stated previously, precipitation and evap- oration are influential on the streamflows in a catchment. In this study, outputs of downscaling models for precipitation and evaporation at a station located close to the streamflow site were used as additional inputs to the original streamflow downscaling model, for improving its performances. The performances of the newstreamflowdownscaling model were compared with those of the original streamflow downscaling model graphically and numerically. Following the improve- ments to the streamflowdownscaling model, using the outputs of HadCM3, ECHAM5 and GFDL2.0 pertaining to A2 and B1 GHG emission scenarios, the streamflow at the site of interest was projected into the future period 20002099. Smith and Chandler (2010) stated that HadCM3, ECHAM5 and GFDL2.0 are capable in correctly simulating the precipita- tion over Australia and also able to produce accurate simulations of El Nio Southern Oscillation (ENSO). They also argued that a GCM which can correctly simulate precipitation should be able to simulate other climatic variables with a good degree of accuracy. Therefore, for the present study, the outputs of HadCM3, ECHAM5 and GFDL2.0 were used. In this paper, Section 2 details the study area and the data used in this study. Section 3 provides the details of the generic methodology. The application of this methodology is de- scribed in Section 4 with results. Finally, Section 5 presents the summary of the study and conclusions. 2 Study area and data The operational area of Grampians Wimmera Mallee Water Corporation (GWMWater) is located in the north-western region of Victoria, Australia. The streamflow site considered in this study (same as in Sachindra et al. 2013) is located close to the southern boundary of the operational area of GWMWater. The projection of streamflow in the operational area of GWMWater (about 62,000 km 2 ) into the future is an important task, as the water supply system located in this area provides water to a large number of domestic and industrial customers, and to the surrounding environment. Figure 1 shows the location of the streamflow site and its catchment in the operational area of GWMWater. The streamflow site (Lat. 37.17, Lon. 142.54) consid- ered in this study is located on the Fyans Creek at Lake Bellfield (refer to Fig. 1). In other words, this site represents the inflow to Lake Bellfield (GWMWater 2011a). Lake Bellfield provides water to nearby towns of Halls Gap and Pomonal, and also many recreational activities such as swim- ming, fishing and boating that take place at this lake. The catchment area demarcated by this streamflow site is about 96 km 2 (GWMWater 2011a) and it is located within the Grampians national park (GWMWater 2011b). Since the catchment area is located within a national park, the land use has remained stationary over the last few centuries and also it is expected to remain constant in the future. The catchment of Lake Bellfield contains soils composed of mainly sandstones and some mudstones (Cayley and Taylor 1997). The inflowto Lake Bellfield is usually of good quality with low salinity and turbidity (Sachindra et al. 2013). The inflow to Lake Bellfield provided by the Fyans Creek is influenced by some diversions into the catchment and also some extractions out of it. The unregulated inflow (naturalised) to Lake Bellfield have been computed by Sinclair (2004). There is no weather observation station located within the catchment of the Bellfield Lake. The closest weather observation station at Halls Gap post office (Lat. 37.14, Lon. 142.52), which is situated about 10 km outside the boundary of the catchment, was selected as the station representative of the climate of this catchment (refer to Fig. 1). In order to provide inputs to the statistical downscaling models (precipitation, evaporation and streamflow) for their calibration and validation, monthly reanalysis outputs of Na- tional Centers for Environmental Predictions/National Center for Atmospheric Research (NCEP/NCAR) were obtained from the website of National Oceanic and Atmospheric Administration/Earth System Research Laboratory (NOAA/ESRL) Physical Sciences Division at http://www. esrl.noaa.gov/psd/, for the period 19502010. For calibration and validation of the precipitation and evaporation downscaling models, observations of monthly precipitation and evaporation of the station at Halls Gap post office were obtained from the SILO database of Queensland Climate Change Centre of Excellence at http://www.longpaddock. qld.gov.au/silo/ for the period 19502010. For calibration and validation of the streamflow downscaling model, the Potential improvements to statistical downscaling of streamflows unregulated monthly streamflow data of the site (inflow to Lake Bellfield) considered in this study were obtained from GWMWater for the period 19502010. In order to identify the bias in the outputs of downscaling models and for correction of bias, past observations of month- ly precipitation, evaporation and streamflows have to be reproduced. For this purpose, the twentieth century climate experiment (20C3M) outputs of HadCM3, ECHAM5 and GFDL2.0 were extracted from the website of the Programme for Climate Model Diagnosis and Inter-comparison (PCMDI) at https://esgcet.llnl.gov:8443/index.jsp, for the period 1950 1999. For validation of the bias-correction, HadCM3 outputs for COMMIT GHG emission scenario were also obtained from the same web site for the period 20002099. For projection of precipitation and evaporation, and hence streamflowinto the future, the outputs of HadCM3, ECHAM5 and GFDL2.0 pertaining to the A2 and B1 GHG emission scenario were also obtained from the PCMDI (https://esgcet. llnl.gov:8443/index.jsp), for the period 20002099. 3 Generic methodology In this study, a statistical model that was developed for down- scaling reanalysis outputs to monthly streamflows was mod- ified for improving its performances. For this purpose, two additional separate statistical downscaling models (one for precipitation and the other for evaporation) were developed with reanalysis outputs for downscaling them to monthly precipitation and evaporation at a station which is located close to the streamflow site. The outputs of these two down- scaling models were introduced to the streamflow downscal- ing model to enhance its performances. The development of the original model for downscaling large-scale atmospheric variables to monthly streamflows using the MLR technique was detailed in Sachindra et al. (2013). This original streamflow downscaling model is referred to as SDM (original) throughout this paper. 3.1 Downscaling models for precipitation and evaporation Sachindra et al. (2014a) found that the MLR technique was capable of successfully capturing the relationships between the reanalysis outputs and precipitation observations. Hence, in this study, downscaling models for monthly precipitation and evaporation were built using the MLR technique. The precipitation downscaling model and evaporation downscal- ing model used in this study are referred to as PDM and EDM throughout this paper. The procedure employed in this study for the development of PDM and EDM was quite similar to that used in developing SDM (original) using MLR. For extrac- tion of large-scale climate information, the same atmospheric domain used in the development of SDM (original) was selected. Fig. 1 Streamflow site and its catchment within operational area of GWMWater D.A. Sachindra et al. Two pools of probable predictors were selected for PDM and EDM by considering the past literature and fundamentals of hydrology. These probable predictors were the variables that are likely to influence precipitation and evaporation. Potential predictors are the subsets of probable predictors which highly influence the predictand of interest. The poten- tial predictors were extracted from the pools of probable predictors for each calendar month for each predictand sepa- rately. This was performed because the set of predictors influ- ential on a certain predictand can vary with the seasonal changes in the atmosphere (Karl et al. 1990). To extract potential predictors from the pools of probable predictors, the Pearson correlation coefficient (Pearson 1895) was used. For each predictand, the record of observations and the reanalysis data pertaining to the probable predictors were split into 20-year time slices. Then the Pearson correlation coefficients between the observations of the predictand of interest and the reanalysis data pertaining to probable predic- tors were computed for all 20-year time slices and the whole period of the records. This procedure was performed at each grid point in the atmospheric domain and for each calendar month separately. The probable predictors which showed statistically significant (p0.05) good correlations with the observations in all time slices and the whole period of the study were selected as the potential predictors for each calen- dar month for each predictand. The first two thirds of the observations of each predictand and the reanalysis data pertaining to the potential predictors, for each calendar month, were allocated for calibration of the downscaling models. The rest of these data sets were used for validation of the downscaling models. The reanalysis data for both calibration and validation phases of the downscaling models were standardised with their means and the standard deviations corresponding to the calibration period. Initially, the standardised reanalysis data of the potential predictor which showed the best correlation with the observations of the predictand over the whole period of the study was intro- duced to the downscaling model. Then, by minimising the sum of squared errors between the model outputs (e.g. pre- cipitation) and the observations, the optimum values of the coefficient and the constant of the linear regression equation were determined. The downscaling model was then validated by introducing the rest of the reanalysis data of that potential predictor to the model. During the validation, the optimum values of the coefficient and the constant determined in the calibration phase of the downscaling model were kept con- stant. The model performances in the calibration and valida- tion phases were measured using the Nash-Sutcliffe efficiency (NSE) (Nash and Sutcliffe 1970). Thereafter, the next best potential predictors were introduced to the downscaling mod- el, one at a time, and the model calibration and validation were performed as stated previously. This stepwise addition of potential predictors was continued until the model performance in validation reached a maximum in terms of the NSE. The maximum performance in the validation was considered in selecting the best model for each calendar month for each predictand, as it aided in avoiding any models which displayed over-fitting in calibration and under-fitting in validation. This procedure yielded the best sets of potential predictors and the optimum downscaling models for each calendar month of each predictand. Note that all downscaling models detailed in this paper had common calibration and validation periods. 3.2 Potential improvements to streamflowdownscaling model Initially, the precipitation and evaporation outputs of the PDM and the EDM pertaining to both calibration and validation phases of the models were standardised for each calendar month using the means and the standard deviations of the observations of the calibration period. Then, for each calendar month, the standardised lag 0 precipitation output of the PDM pertaining to its calibration phase was introduced to the SDM (original) . Following the introduction of the standardised lag 0 precipitation, this model was calibrated, and the opti- mum values for its parameters (coefficients and constants of MLR equations) were found for each calendar month. The SDM (original) modified with lag 0 precipitation is called SDM (lag0_preci) . In the validation of SDM (lag0_preci) , the opti- mum values of the constants and the coefficients of MLR equations determined in the calibration phase of SDM (lag0_preci) were kept fixed. Then, the standardised lag 0 precipitation outputs of the PDM pertaining to the validation phase were introduced to SDM (lag0_preci) for reproducing the observed streamflow for that period. The performance of SDM (lag0_preci) was monitored using the NSE and scatter plots in the calibration and validation phases. Similarly, SDM (original) was again calibrated and validated by separately introducing the standardised lag 1 and lag 2 precipitation outputs of the PDM, yielding modified streamflow downscaling models SDM (lag1_preci) and SDM (lag2_preci) , respectively. Then, considering the NSEs in validation, the best model out of SDM ( l ag0_pr eci ) , SDM (lag1_preci) and SDM (lag2_preci) was selected for the next step of the model improvement. This best model is called SDM (lagi_preci) . In the next step, SDM (lagi_preci) was further modified by individually introducing the standardised lag 0, lag 1 and lag 2 evaporation outputs of the EDM pertaining to the calibration phase. This yielded SDM (lagi_preci_ & _lag0_evap) (i.e. SDM (lagi_- preci) with lag 0 evaporation), SDM (lagi_preci_ & _lag1_evap) (i.e. SDM (lagi_preci) with lag 1 evaporation) and SDM (lagi_- preci_ & _lag2_evap) (i.e. SDM (lagi_preci) with lag 2 evaporation). The optimum values of the coefficients and constants of MLR equations were found for each calendar month in the calibra- tion of above models, and then they were validated using the Potential improvements to statistical downscaling of streamflows standardised lag 0, lag 1 and lag 2 evaporation outputs of the EDM pertaining to the validation phase as described previ- ously. The performances of these three SDMs were monitored using NSE and scatter plots, in their calibration and validation phases. The best SDM was selected based on the NSE in the validation phase. This best model is referred to as SDM (lagi_- preci_ & _lagj_evap) , where i and j refer to lags of precipitation and evaporation. The above stepwise modification process applied to SDM (original) is graphically illustrated in Fig. 2. Another SDM was developed (calibrated and validated) by introducing observed lag i precipitation and lag j evaporation to SDM (original) corresponding to the best SDM identified previously. The lags i and j refer to lags of precipitation and evaporation used in the best SDM. This downscaling model developed using the observed lag i precipitation and lag j evaporation as additional inputs to SDM (original) is called SDM (original+OBS) . It allowed the quantification of the maxi- mum potential improvement to SDM (original) with observa- tions of precipitation and evaporation. Numerical and graph- ical performance comparisons between the SDM (original) , SDM (lagi_preci_ & _lagj_evap) and SDM (original+OBS) were per- formed. This enabled the quantification of the actual and potential improvements to SDM (original) following the modifi- cations made in this study. 3.3 Reproduction of precipitation, evaporation and hence streamflow using GCM outputs and bias-correction Once the PDM and the EDM were developed, they were run with the 20C3M outputs of each GCM. For this purpose, the 20C3M outputs of each GCM corresponding to potential predictors used in the PDM and EDM were standardised with the means and the standard deviations of reanalysis data relevant to the calibration phase of the downscaling models. The reproduction of precipitation and evaporation using the 20C3M outputs of each GCM enabled the quantification of bias introduced to precipitation and evaporation simulations by each GCM and its subsequent correction. Once precipita- tion and evaporation observations for the past were reproduced by the PDM and the EDM using 20C3M outputs of the GCMs, the statistics of reproduced precipitation and evaporation were compared with those of observations for the quantification of bias. Then, the equidistant quantile mapping (EQM) technique (Li et al. 2010) was used to correct the bias in precipitation and evaporation reproduced by PDM and EDM. For the application of the EQM technique, empirical cu- mulative distribution functions (CDFs) were derived from the outputs of PDM and EDM pertaining to the past climate (using 20C3M outputs of each GCM) and also from the precipitation and evaporation observations, for each calendar month separately. Then, for each calendar month, the CDFs derived from the outputs of PDM and EDM for the past climate were mapped onto the CDFs developed from the precipitation and evaporation observations respectively. In this mapping process, first, for a given value of the predictand of interest (e.g. precipitation) downscaled by the model (e.g. PDM), the corresponding CDF value was found from the CDF derived from the values of that predictand of interest downscaled by the model. Then pertaining to that CDF value, the value of the predictand of interest was found fromthe CDF derived from the past observations. This value of the predictand of interest obtained from the CDF derived from the past observations is the bias-corrected value of the predictand downscaled by the model. The above process was applied to all values of the predictand of interest down- scaled by the model and it yielded the bias-corrected CDF. The application of the EQM technique allowed the correction of bias in all statistical moments of precipitation and evapo- ration downscaled by the PDM and the EDM for the past climate with GCMoutputs. Adetailed application of the EQM technique is found in Li et al. (2010), Salvi et al. (2011) and Sachindra et al. (2014b). The time series of precipitation and evaporation were de- rived from their bias-corrected CDFs and for each calendar month precipitation and evaporation data were standardised using the means and the standard deviations of their observa- tions relevant to the calibration period of the downscaling model. The 20C3M data relevant to the potential predictors used in the SDM (original) were also standardised with the means and the standard deviations of reanalysis data of the calibration phase of the downscaling models. Then, the SDM (lagi_preci_ & _lagj_evap) was run with the bias-corrected out- puts of the PDM and the EDM produced in the above step along with 20C3M outputs of each GCM pertaining to the inputs to SDM (original) . The EQM technique was used to correct the bias in the streamflow downscaled by SDM (lagi_- preci_ & _lagj_evap) , run with the bias-corrected outputs of the PDM and the EDM along with 20C3M outputs of each SDM (original) SDM (lagi_preci) Select best model Lag i (i = 0, 1, 2) precipitation SDM (lagi_preci_&_lagj_evapo) Lag j (j = 0, 1, 2) evaporation Select best model Fig. 2 Flow chart for modification of SDM (original) D.A. Sachindra et al. GCM. For this purpose, the same procedure used for the correction of bias in the outputs of PDM and EDM was employed. For validating the performances of the EQM technique for precipitation and evaporation, statistics of bias-corrected pre- cipitation and evaporation downscaled using the outputs of a GCM pertaining to the COMMIT GHG emission scenario for the future climate were compared with those of observed precipitation and evaporation. For validating the performances of the EQM technique for streamflow, statistics of bias- corrected streamflow downscaled using the outputs of the GCM and the bias-corrected precipitation and evaporation relevant to the COMMIT GHG emission scenario were com- pared with those of observed streamflow. The COMMIT GHGemission scenario assumes that the GHG concentrations observed at the end of the twentieth century (CO 2 concentra- tion in the atmosphere370 ppm) will remain the same throughout the twenty-first century. Assuming that the rise in the GHG concentrations in the atmosphere in the latter half of the twentieth century is small, it can be argued that the climate simulated by a GCM over the latter half of the twen- tieth century is in close agreement with the climate projected into the future by the same GCM under the COMMIT GHG emission scenario (argument proven by Sachindra et al. (2014b)). The GCM outputs pertaining to the COMMIT GHG emis- sion scenario were standardised using the means and the standard deviations of reanalysis outputs. Then, PDM and EDM were run with the COMMIT GHG emission scenario outputs of the GCM. Thereafter, the empirical CDFs were derived from the precipitation and evaporation time series projected into the future under the COMMIT GHG emission scenario. Then for each calendar month, the difference be- tween the CDF of the predictand of interest (precipitation or evaporation) produced into the future under the COMMIT GHG emission scenario and the CDF simulated for the past corresponding to 20C3M was added to the CDF of the obser- vations of the predictand. This process yielded the bias- corrected CDFs of precipitation and evaporation for the future climate characterised by the COMMIT scenario. Following this bias-correction, for the validation of the EQM technique, the statistics of precipitation and evaporation downscaled from GCM outputs for COMMIT GHG emission scenario were compared with those of past observations. The bias-corrected time series of precipitation and evapo- ration were standardised with the means and the standard deviations of observations relevant to the model calibration period for each calendar month. The GCM outputs of the COMMIT GHG emission scenario for SDM (lagi_preci_ & _lagj_- evap) were also standardised with the corresponding means and standard deviations of reanalysis outputs of the model calibra- tion period for each calendar month. Then, SDM (lagi_- preci_ & _lagj_evap) was run with these bias-corrected outputs of the PDM and the EDM and the GCM outputs of COMMIT GHG emission scenario. The streamflow produced in the above step was bias-corrected using the EQM technique, following the same procedure used for the correction of bias in precipitation and evaporation produced by the downscaling model for the COMMIT GHG emission scenario. Then, the statistics of bias-corrected streamflow produced by SDM (lagi_- preci_ & _lagj_evap) for the COMMIT GHG emission scenario were compared with those of past observed streamflow. This comparison of statistics enabled the assessment of effective- ness of the EQM technique (or validation of EQM) in correcting the bias in streamflow projections. Figure 3 shows the main steps involved in the application and validation of the bias-correction for precipitation, evapo- ration and streamflows in a flow chart. In that flow chart, the steps shown above the dashed line (in dark printblack) refer to the application of the bias-correction to precipitation, evap- oration and streamflows for the past climate. The steps shown below the dashed line (in light printblue) refer to the vali- dation of the bias-correction with COMMIT GHG emission scenario and the same set of steps were followed for the bias- correction of projections of precipitation, evaporation and streamflows produced into the future with another GHGemis- sion scenario (e.g. A2) as described in the next section. 3.4 Projection of precipitation, evaporation and hence streamflow into the future, and bias-correction The outputs of each GCM for the PDM and EDM pertaining to future climate (e.g. A2, B1 GHG emission scenario) were standardised using the corresponding means and standard devi- ations of reanalysis outputs relevant to the model calibration period for each calendar month. Then, the PDM and the EDM were run with these standardised outputs of each GCM for the future climate. Following the procedure described in Section 3.3 used for the validation of effectiveness of the EQM technique (refer to Fig. 3), it was used to correct the bias in the precipitation and evaporation projected into the future. The bias-corrected time series of precipitation and evaporation for future were standardised with the means and the standard deviations of observations relevant to the model calibration period for each calendar month. The outputs of each GCM for the SDM (lagi_- preci_ & _lagj_evap) for future climate were standardised using the corresponding means and standard deviations of reanalysis out- puts relevant to the model calibration period. The bias-corrected standardised outputs of PDM and EDM for future climate along with other standardised outputs of each GCM were introduced to SDM (lagi_preci_ & _lagj_evap) for the projection of streamflow at the point of interest into the future. This procedure was performed for each calendar month sepa- rately. Following the same procedure used in correcting the bias in the projections of precipitation and evaporation for the COMMIT GHG scenario (refer to Section 3.3), bias in the Potential improvements to statistical downscaling of streamflows streamflow projected into the future by SDM (lagi_preci_ & _lagj_- evap) was corrected using the EQM technique. Then, the sea- sonal statistics of observed streamflow were compared with those of streamflow projected into the future. Also, the ensem- ble average streamflow time series was computed from the streamflow outputs of SDM (lagi_preci_ & _lagj_evap) when it was run with the outputs of each GCMand the outputs of PDMand EDM. The probability exceedance curves for projected ensem- ble average streamflow and observed streamflows were com- pared for each season for the assessment of possible changes in the streamflow regime in the future. 4 Application 4.1 Downscaling models for precipitation and evaporation 4.1.1 Domain and predictor selection for downscaling models Before developing the downscaling models, an atmospheric domain spanning over the study area was defined. This atmo- spheric domain spanned over the longitudes 135150 E and latitudes 3042.5 S. It included seven and six grid points in the longitudinal and latitudinal direction respectively. The PDM/EDM 20C3M GCM outputs Application of EQM to correct bias SDM (lag0_preci) 20C3M GCM outputs CDF of precipitation/ evaporation for past climate Bias-corrected CDF of precipitation/ evaporation for past climate CDF of streamflow for past climate Bias-corrected CDF of streamflow for past climate Application of EQM to correct bias Bias-corrected time series of precipitation/ evaporation for past climate PDM/EDM COMMIT GCM outputs/any future scenario Application of EQM to correct bias SDM (lag0_preci) COMMIT GCM outputs/any future scenario CDF of precipitation/ evaporation for COMMIT/any future scenario Bias-corrected CDF of precipitation/evaporation for COMMIT/any future scenario CDF of streamflow for COMMIT/ any future scenario Bias-corrected CDF of streamflow for COMMIT/ any future scenario Application of EQM to correct bias Bias-corrected time series of precipitation/ evaporation for COMMIT/any future scenario Difference between two CDFs (past and future) Bias-corrected time series of streamflow for past climate CDF of observed precipitation/ evaporation for past climate CDF of observed streamflow for past climate Difference between two CDFs Bias-corrected time series of streamflow for COMMIT/ any future scenario Application of bias- correction Validation of bias- correction Fig. 3 Main steps involved in application and validation of bias- correction D.A. Sachindra et al. spatial resolution of the atmospheric domain was selected as 2.5 in both directions in order to be consistent with the spatial resolution of NCEP/NCAR reanalysis outputs. This atmo- spheric domain is shown in Fig. 4. The predictor selection, the calibration and validation, the bias-correction and the projection of precipitation into the future using a precipitation downscaling model at the Halls Gap post office was detailed in the studies by Sachindra et al. (2014a, b). That precipitation downscaling model was used in the current study as the PDM. Therefore, for more details about the PDM used in this study, readers are referred to Sachindra et al. (2014a, b). The details on the EDMdeveloped in this study are provided in this section. Once the atmospheric domain was defined, a pool of prob- able predictors for evaporation was selected. Timbal et al. (2009) used meteorological analogues for downscaling GCM outputs to daily precipitation, pan evaporation, mini- mum temperature, maximum temperatureand dew point tem- perature over the southern half of Australia. Since the present study area is also located in the southern region of Australia, the predictors used in that study for downscaling evaporation were also used in the present study. The NCEP/NCAR reanalysis data pertaining to the probable predictors and the observations of evaporation at the Halls Gap post office for the period 19502010 were split into three time slices; 19501969, 19701989 and 19902010. Following the procedure detailed in Section 3.1, potential predictors for evap- oration for each calendar month were identified fromthe pool of probable predictors. Then, the NCEP/NCAR reanalysis data pertaining to the potential predictors and the observations of evaporation were separated into two chronological groups; 19501989, 19902010. The former group of data was used for the calibration, while the latter group was used for the validation of the downscaling model. As described in Sec- tion 3.1, the MLR-based statistical downscaling models were developed for each calendar month for evaporation, and the best sets of potential variables were identified. The same procedure was practised by Sachindra et al. (2013, 2014a) in developing the SDM (original) and PDM, respectively. The best sets of poten- tial predictors identified for precipitation and evaporation for each calendar month are shown in Table 1. Table 2 shows the best sets of potential predictors used in SDM (original) for each calendar month with their gird locations. Note that in December, volumetric moisture content in soil layer 10200 cm at grid points (4,3) and (4,4) used as inputs to SDM (original) were removed from the set of input. This was because the volumetric moisture content in soil layer 10 200 cm at grid points (4,3) and (4,4) simulated by HadCM3 indicated presence of large bias in its variance (in comparison to NCEP/NCAR reanalysis data). SDM (original) was re-calibrated and validated only for December without the volumetric mois- ture content in soil layer 10200 cmat grid points (4,3) and (4,4), and it was found that the removal of the above two variables from the model does not change its performance significantly. 4.1.2 Performances of precipitation and evaporation downscaling models The performances of the PDM and the EDM were monitored during their calibration and validation periods using statistical Fig. 4 Atmospheric domain for downscaling Potential improvements to statistical downscaling of streamflows and graphical techniques. Table 3 shows the statistics of the observed and EDMreproduced evaporation for the calibration and validation period. Performances of PDM were detailed in Sachindra et al. (2014b, c). According to Table 3, it was seen that the EDM was able to reproduce the statistics of observed evaporation with good degree of accuracy in both calibration and validation periods. EDM showed NSEs of 0.97 and 0.98 in its calibration and validation periods respectively, and PDM displayed NSEs of 0.74 and 0.70 for the same periods (Sachindra et al. 2014b). In general, precipitation shows high levels of fluctuations than evaporation (which usually displays smooth regular variations over time). This allows downscaling large-scale atmospheric variables to evaporation with higher degree of accuracy in comparison to precipitation. Overall, both PDM and EDM showed good per- formances in the calibration and validation phases. 4.2 Potential improvements to streamflowdownscaling model As described in Section 3.2, the SDM (original) was modified using the precipitation and evaporation reproduced by PDM Table 1 Best sets of potential predictors for each calendar month of PDM and EDM Month Potential variables used in the PDM and EDM with grid locations Precipitation Evaporation January Surface precipitation rate {(3,3),(4,4)} 1000 hPa air temperature {(3,4)} 1000 hPa specific humidity {(3,3),(3,4),(4,4)} 850 hPa meridional wind {(2,6),(3,5),(3,6)} 850 hPa relative humidity {(1,2)} 2 m specific humidity {(3,3),(3,4)} February Surface precipitation rate {(3,4),(4,4),(4,5)} 1000 hPa relative humidity {(3,3),(4,4)} 925 hPa relative humidity {(3,3),(3,4),(4,4),(4,5)} March Surface precipitation rate {(3,3),(3,4),(3,5),(4,3),(4,4),(4,5),(4,6)} 1000 hPa air temperature {(2,5)} April 850 hPa relative humidity {(4,3),(4,4)} 925 hPa relative humidity {(2,1)} Surface precipitation rate {(4,3)} May Surface precipitation rate {(4,4),(5,5)} 1000 hPa relative humidity {(4,4)} 850 hPa geopotential height {(4,3)} June Surface precipitation rate {(3,2),(3,3),(4,2),(4,3),(4,4),(4,5)} 1000 hPa relative humidity {(4,4)} Mean sea level pressure {(4,3),(5,3)} 850 hPa zonal wind {(2,4)} Surface pressure {(4,3),(5,3),(5,4)} July 850 hPa zonal wind {(1,3),(1,4)} 1000 hPa relative humidity {(2,2),(2,3),(3,4)} 850 hPa geopotential height {(4,3),(4,4),(4,5)} 925 hPa relative humidity {(2,3),(3,4)} Surface precipitation rate {(1,4),(1,5),(1,6),(2,5)} August Surface precipitation rate {(4,3),(5,4),(5,5)} 1000 hPa relative humidity {(3,1)} 925 hPa relative humidity {(2,3),(3,3)} September Surface precipitation rate {(2,1),(2,2),(3,2),(3,3), (3,5),(4,2),(4,3),(4,4),(4,5)} 1000 hPa relative humidity {(2,2),(3,2),(3,3)} 850 hPa relative humidity {(3,3)} 700 hPa relative humidity {(3,4)} October Surface precipitation rate {(3,2),(4,2),(4,3),(4,4)} 700 hPa geopotential height {(1,1)} 850 hPa relative humidity {(4,3)} 700 hPa geopotential height {(1,1)} November 850 hPa relative humidity {(3,2),(3,3)} 850 hPa air temperature {(2,5),(3,5)} Surface precipitation rate {(4,3),(4,5)} 700 hPa geopotential height {(2,3),(2,4),(2,5)} December Surface precipitation rate {(2,1),(3,2),(4,3),(4,4),(5,5)} 1000 hPa air temperature {(2,2)} 850 hPa relative humidity {(3,2)} 925 hPa relative humidity {(3,2)} Surface skin temperature {(2,1),(2,2),(2,3),(3,5)} The locations are given within brackets (see Fig. 1) hPa atmospheric pressure in hectopascal D.A. Sachindra et al. and EDM. Table 4 provides the statistics of observed and model simulated streamflows for the calibration and validation periods before and after the modification of SDM (original) . As can be seen from Table 4, when the SDM (original) was modified using the lag 0 and lag 1 precipi- tation produced by the PDM, NSEs in both calibration and validation period increased equally (refer to SDM (lag0_preci) and SDM (lag1_preci) in Table 4). However, when lag 2 precip- itation was introduced to SDM (original) , the model performance in terms of NSE increased in calibration, but decreased in validation. This indicated that with the introduction of lag 2 precipitation, SDM (lag2_preci) tends to show signs of over- fitting in calibration and under-fitting in validation. Consider- ing the improvement in NSE in validation, SDM (lag0_preci) was selected for further modification with evaporation produced by the EDM. After the modification of SDM (lag0_preci) with lag 0, lag 1 and lag 2 evaporation, Table 4 shows that further improvement to this model was not possible. Furthermore, when lag 0 and lag 1 precipitation together, lag 0, lag 1 and lag 2 precipitation together, lag 1 precipitation and lag 0 evapo- ration together, lag 1 precipitation and lag 1 evaporation together and lag 1 precipitation and lag 0 and lag 1 evapora- tion together were used as additional inputs to SDM (original) , no further improvement to the model was seen. Hence SDM (lag0_preci) was selected as the best streamflow downscal- ing model. As seen in Table 2, for SDM (original) , in all seasons volumetric soil moisture content and humidity variables (relative or specific humidity at different pressure levels) have been selected as best potential variables. The rate of evaporation from a catchment is dependent on factors such as: wind speeds, air temperature, atmospheric humidity, sunshine hours (Wang et al. 2012) and also the soil mois- ture content (Shang et al. 2007). Therefore, it can be argued that the inclusion of variables representative of atmospheric humidity and soil moisture content in the downscaling model can compensate the exclusion of evap- oration to a certain degree. In Table 4, it was seen that with the introduction of lag zero evaporation to SDM (lag0_preci) , the performance of the SDM (lag0_preci_ & _lag0_evap) in terms of NSE does not show any significant change. In fact, with the introduction of lag zero evaporation the performance of SDM (lag0_preci_ & _lag0_evap) in terms of NSE has slightly increased in model calibration period and slightly decreased in model validation period with respect to that of SDM (lag0_preci) . The SDM (lag0_preci) already contains atmospheric humidity and soil moisture variables (see Table 2) which can indirectly explain the influence of evapo- ration on streamflow. Therefore, when evaporation is intro- duced to SDM (lag0_preci) it brings redundant information to the model. This redundant information fed into the model does not cause any significant improvement to the model performance. In other words, the inclusion of evaporation in the SDM (lag0_preci) does not lead to any significant improvement in model performance. Table 2 Best sets of potential predictors for each calendar month of SDM (original) Month Potential variables used for MLR model with grid locations January 1000 hPa relative humidity {(1,2),(2,2),(2,3),(3,4)} February Volumetric soil moisture content 010 cm {(4,3)} Volumetric soil moisture content 10200 cm {(2,2),(3,1),(3,2)} March 1000 hPa relative humidity {(5,6),(6,6),(6,7)} April 700 hPa relative humidity {(4,2),(4,3)} 850 hPa relative humidity {(4,2)} May Volumetric soil moisture content 010 cm {(4,3),(4,4),(4,5)} June Volumetric soil moisture content 010 cm {(4,3),(4,4),(5,2),(5,3),(5,4),(6,3)} 500 hPa geopotential height {(4,2)} July 700 hPa geopotential height {(4,4)} 850 hPa geopotential height {(4,3),(4,4),(4,5)} August 700 hPa geopotential height {(5,4),(5,5)} 850 hPa geopotential height {(5,4),(5,5),(5,6)} September 700 hPa geopotential height {(2,1),(3,2),(3,3)} October Volumetric soil moisture content 010 cm {(4,3),(4,4),(5,3)} November 700 hPa geopotential height {(2,2),(2,3),(2,4),(3,2),(3,3),(3,4)} December 700 hPa relative humidity {(4,3)} 850 hPa relative humidity {(3,1)} 1000 hPa relative humidity {(6,6)} 850 hPa specific humidity {(5,5)} Volumetric soil moisture content 010 cm {(4,4),(4,5),(5,4)} Volumetric soil moisture content 10200 cm {(3,2),(4,3) a ,(4,4) a } The locations are given within brackets (see Fig. 1) hPa atmospheric pressure in hectopascal a Variables were removed from the model due to large bias in variance in HadCM3 outputs Table 3 Performances of EDM in calibration and validation Statistic Calibration (19501989) Validation (19902010) Observations EDM Observations EDM Avg 110.3 110.3 109.7 109.4 Std 67.3 66.5 65.7 65.6 Min 21.8 22.2 23.8 22.8 Max 277.7 259.3 262.0 251.3 NSE 0.97 0.98 R 2 0.97 0.98 Avg average of monthly evaporation in mm, Std standard deviation of monthly evaporation in mm, Min minimum of monthly evaporation in mm, Max maximum of monthly evaporation in mm, NSE Nash-Sutcliffe efficiency, R 2 coefficient of determination Potential improvements to statistical downscaling of streamflows Then, in order to quantify the maximumpotential improve- ment to SDM (original) , observed lag 0 precipitation was intro- duced to SDM (original) as an additional input and the model was calibrated and validated; this model is called SDM (original+ OBS) . SDM (original+OBS) showed a significant increase in NSE in both calibration and validation periods as seen in Table 4. Also, the average, the standard deviation and the maximum of streamflow reproduced by SDM (original+OBS) was in better agreement with those of observations in both calibration and validation periods. The rise in NSE of SDM (original+OBS) was much higher than that of SDM (lag0_preci) . This indicated that if the PDM is able to perfectly mimic the observed precipitation, then a significant improvement to SDM (original) can be intro- duced by using the lag 0 precipitation of PDM as an input. It was concluded that any significant improvement to PDM will improve the performances of SDM (lag0_preci) . Figure 5 shows the scatter of streamflow produced by SDM (original) , SDM (lag0_preci) and SDM (original+OBS) during the calibration (19501989) and the validation (19902010) pe- riods. Though a rise in NSEs in both calibration and validation periods was seen, the reduction of scatter in streamflow down- scaled by SDM (lag0_preci) in its calibration and validation phases was minimumin comparison to the scatter of streamflowdown- scaled by SDM (original) . However, the streamflow output of SDM (original+OBS) indicated a clear reduction of its scatter in both calibration and validation periods in comparison to scatter of streamflow of SDM (original) and SDM (lag0_preci) . 4.3 Reproduction of precipitation and hence streamflow using outputs HadCM3, ECHAM5 and GFDL2.0 4.3.1 Reproduction of precipitation and bias-correction As stated in Section 3.3, the PDM was run with the 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 for reproduction of observed precipitation over the period 19501999. This allowed the quantification of bias in the precipitation downscaled by the PDM and its subsequent correction. Since SDM (lag0_preci) does not need evaporation as an input, the EDM was not used for the rest of the study. The statistics of precipitation reproduced by the PDM with NCEP/NCAR reanalysis and 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 and the statistics of observed precipitation (for the observation station located at Halls Gap post office) were shown in Sachindra et al. (2014c). In that study, it was seen that with 20C3M outputs HadCM3 and GFDL2.0, the PDM was able to reproduce the standard deviation, the minimum and the maximum of precipitation with good accuracy during the period 19501999. However, with 20C3M outputs of ECHAM5, the PDM over-estimated (presence of bias) the average, the standard deviation, and the minimum and the maximum of precipitation. Also, the average of precipitation was largely over-estimated by the PDM with the 20C3M outputs of HadCM3 and GFDL2.0. For cor- rection of this bias, in the study by Sachindra et al. (2014c), the EQM technique was applied to the precipita- tion simulated by the PDM run with the 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 for the period 19501999 as described in Section 3.3. In that study, after the application of the EQM technique, the average, the standard deviation and the maximum of precipitation were perfectly corrected, though the scatter of downscaled pre- cipitation still remained large indicated by small R 2 and NSE. In Sachindra et al. (2013c, d), for validating the effective- ness of the EQM technique for precipitation, statistics of bias- corrected precipitation downscaled using the outputs of HadCM3 pertaining to COMMIT GHG emission scenario for the period 20002099 were compared with those of Table 4 Performances of SDMs in calibration and validation Model Calibration (19501989) Validation (19902010) Avg Std Min Max NSE Avg Std Min Max NSE Observations 2129.3 2387.3 0.0 12,427.0 N/A 1318.6 1788.1 0.0 9387.0 N/A SDM (original) 2132.4 1984.5 0.0 9773.5 0.69 2022.3 2123.7 0.0 11,255.2 0.33 SDM (lag0_preci) 2131.0 2019.1 0.0 10,770.9 0.72 2048.4 2157.7 0.0 11,539.8 0.36 SDM (lag1_preci) 2133.7 2023.4 0.0 9231.9 0.72 2011.4 2169.0 0.0 11,759.2 0.36 SDM (lag2_preci) 2134.3 2021.2 0.0 9739.9 0.72 2024.1 2176.7 0.0 11,915.1 0.28 SDM (lag0_preci_ & _lag0_evap) 2131.3 2044.0 0.0 10,256.5 0.73 2039.7 2179.0 0.0 11,239.6 0.35 SDM (lag0_preci_ & _lag1_evap) 2131.9 2048.9 0.0 10,816.6 0.74 2112.0 2244.6 0.0 12,117.5 0.30 SDM (lag0_preci_ & _lag2_evap) 2130.8 2029.1 0.0 10,930.7 0.72 2071.0 2195.7 0.0 11,851.7 0.33 SDM (original+OBS) 2133.5 2145.1 0.0 12,072.7 0.81 1867.8 2006.0 0.0 10,741.8 0.60 Avg average of monthly stream flow in 10 3 m 3 , Std standard deviation of monthly streamflow in 10 3 m 3 , Min minimum of monthly streamflow in 10 3 m 3 , Max maximum of monthly streamflow in 10 3 m 3 , NSE Nash-Sutcliffe efficiency D.A. Sachindra et al. observed precipitation for the period 19501999. This proce- dure was briefly described in Section 3.3 of this paper. Sachindra et al. (2014c) commented that in validation, EQM was able to successfully reduce the bias in precipitation downscaled with HadCM3 outputs of the COMMIT GHG emission scenario for the period 20002099 and also it was assumed that EQM is able to reduce the bias in the precipita- tion downscaled with the outputs of ECHAM5 and GFDL2.0. 0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.72 (c) Calibration 0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.36 (d) Validation 0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.81 (e) Calibration 0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.60 (f) Validation 0 2000 4000 6000 8000 10000 12000 14000 0 2000 4000 6000 8000 10000 12000 14000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.69 (a) Calibration 0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000 12000 P r e d i c t e d
s t r e a m f l o w
( x
1 0 3 m 3 / m o n t h ) Observed flow (x 10 3 m 3 /month) N-S = 0.33 (b) Validation SDM (original) SDM (lag0_preci) SDM (original+OBS) Fig. 5 Scatter plots for calibration and validation phases of SDMs Potential improvements to statistical downscaling of streamflows 4.3.2 Reproduction of streamflow and bias-correction For reproduction of past observed streamflow, SDM (lag0_preci) was run with the 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 and the bias-corrected precipitation simulated by the PDM, as described in Section 3.3. This allowed the quan- tification of bias in the streamflow downscaled by the SDM (lag0_preci) and its subsequent correction. The statistics of observed streamflow and those of streamflow simulated by SDM (lag0_preci) during the period 19501999 are shown in Table 5. It can be seen from Table 5 that when the SDM (lag0_preci) was run with the 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 and outputs of PDM, it tended to over-estimate the average, the standard deviation, and the maximum of streamflow. This indicated the influence of bias in 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 on downscaled streamflow. Once the EQM technique was ap- plied to the streamflow produced by SDM (lag0_preci) , the bias in the average, the standard deviation and the maximum of streamflow was corrected to a good degree. However, the improvement in NSE and R 2 were small. This indicated that though the statistics of streamflow were well corrected, the scatter of the downscaled streamflow has not reduced much. This leads to the conclusion that projections of streamflow produced into the future should not be used as time series, instead they should be used in terms of probability distribution and statistics such as seasonal average, standard deviation and minimum/maximum etc. 4.3.3 Validation of bias-correction for streamflow After the EQM technique was applied to the streamflow outputs of SDM (lag0_preci) for past climate, a validation of this bias-correction was performed. For validation of ef- fectiveness of the EQM-based bias-correction for streamflow, the SDM (lag0_preci) was run with the outputs of HadCM3 pertaining to COMMIT GHG emission sce- nario. Then, the EQM technique was applied to the streamflow projection produced into the future period 20002099 under COMMIT GHG emission scenario, as described in Section 3.3. In Tables 6 and 7, the statistics of bias-corrected streamflow downscaled using COMMIT HadCM3 outputs for the period 20002099 were compared with those of observed streamflow of the period 19501999. As shown in Tables 6 and 7, it was seen that prior to the application of the EQM bias-correction, the average of streamflow was over-estimated by the down- scaling model in all seasons. However, with the application of the EQM technique, this over-estimation has largely reduced. The standard deviation and the maximum of streamflow in summer and autumn were also corrected with good accuracy by EQM. Following the EQM bias-correction, the mismatch between the minimum of observed streamflow and that of streamflow projected into the future also reduced largely. As the influence of bias in 20C3M outputs of HadCM3, ECHAM5 and GFDL2.0 on downscaled streamflow were similar in nature (see Table 5), it was assumed that the influ- ence of bias in COMMIToutputs of HadCM3, ECHAM5 and GFDL2.0 on downscaled streamflow were also similar in nature. Hence, it was realised that the EQM technique should also be able to correct the bias in the streamflows projected into the future by the downscaling model run with the outputs of ECHAM5 and GFDL2.0. 4.4 Projection of precipitation and hence streamflow into the future 4.4.1 Projection of precipitation into the future In order to produce streamflow projections into the future using SDM (lag0_preci) , projections of precipitation are required. In the study by Sachindra et al (2014c), by introducing the Table 5 Performances of SDM (lag0_preci) with NCEP/NCAR and HadCM3 outputs before and after bias-correction Statistic Period (19501999) Observations With NCEP/NCAR outputs With HadCM3 outputs With ECHAM5 outputs With GFDL2.0 outputs Before bias- correction After bias- correction Before bias- correction After bias- correction Before bias- correction After bias- correction Avg 2048.7 2144.5 3515.5 2035.4 3507.4 2034.9 3506.6 2035.2 Std 2339.4 2031.3 2724.1 2350.1 2719.5 2350.5 2706.0 2350.3 Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Max 12,427.0 11,255.2 13,791.1 12,427.0 13,997.3 12,427.0 12,873.7 12,427.0 NSE 0.69 0.79 0.20 0.77 0.15 0.80 0.25 R 2 0.69 0.17 0.16 0.17 0.18 0.16 0.14 Avg average of monthly streamflow in 10 3 m 3 , Std standard deviation of monthly streamflow in 10 3 m 3 , Min minimum of monthly streamflow in 10 3 m 3 , Max maximum of monthly streamflow in 10 3 m 3 , NSE Nash-Sutcliffe efficiency, R 2 coefficient of determination D.A. Sachindra et al. standardised outputs of HadCM3, ECHAM5 and GFDL2.0 pertaining to the A2 and B1 GHG emission scenarios to the PDM, precipitation projections were produced over the period 20002099, and those projections were bias-corrected using the EQM technique. For a more detailed description on the application of the EQM technique in correcting bias in the precipitation output of a downscaling model, readers are re- ferred to Sachindra et al (2014b). The A2 and B1 GHG emission scenario refer to high and low levels of GHG emissions in the twenty-first century respectively (A2/B1 CO 2 concentration of about 850/ 550 ppm at the end of the twenty-first century) (IPCC 2000). The use of A2 and B1 GHG emission scenarios in a statistical downscaling study allows the quantification of impacts of relatively high and low levels of GHG emissions on the catchment scale hydroclimatology, respectively (Sachindra et al. 2014b). In the study by Sachindra et al (2014c), it was found that when the HadCM3 outputs corresponding to A2 and B1 GHG emission scenarios were used as inputs to the PDM, the average of precipitation at the observation station at Halls Gap post office in the period 20002099 showed an increase in autumn and winter, and a decrease in summer and spring in comparison to the corresponding seasonal averages of ob- served precipitation of the period 19501999. In the same study, when the ECHAM5 outputs for the A2 and B1 GHG emission scenarios were used as inputs to the PDM, the average of precipitation in the period 20002099 showed an increase in summer and a decline in winter and spring in comparison to the corresponding seasonal averages of ob- served precipitation of the period 19501999. With the use of outputs of GFDL2.0 pertaining to A2 and B1 GHG emis- sion scenarios to the PDM, the average of precipitation in the period 20002099 showed a rise in spring and a decline in autumn in comparison to the corresponding seasonal averages of observed precipitation of the period 19501999. 4.4.2 Projection of streamflow into the future Once precipitation was projected into the future (20002099) and bias-corrected, it was standardised for each calendar month using the mean and the standard deviation of observed Table 6 Statistics of observed streamflow of the period 19501999 and streamflow downscaled using HadCM3 COMMIT outputs for 20002099 (summer and autumn) Statistic Summer Autumn Observed SDM (lag0_preci) with HadCM3 COMMIT outputs (20002099) Observed SDM (lag0_preci) with HadCM3 COMMIT outputs (20002099) Before bias-correction After bias-correction Before bias-correction After bias-correction Avg 836.0 997.6 822.0 603.2 1424.9 754.5 Std 850.4 793.8 865.6 936.2 1473.8 925.9 Min 0.0 0.0 0.0 0.0 0.0 0.0 Max 8041.0 3357.0 8714.9 8017.0 4640.6 8207.0 Avg average of monthly streamflow in 10 3 m 3 , Std standard deviation of monthly streamflow in 10 3 m 3 , Min minimum of monthly streamflow in 10 3 m 3 , Max maximum of monthly streamflow in 10 3 m 3 Table 7 Statistics of observed streamflow of the period 19501999 and streamflow downscaled using HadCM3 COMMIT outputs for 20002099 (winter and spring) Statistic Winter Spring Observed SDM (lag0_preci) with HadCM3 COMMIT outputs (20002099) Observed SDM (lag0_preci) with HadCM3 COMMIT outputs (20002099) Before bias-correction After bias-correction Before bias-correction After bias-correction Avg 3394.3 3888.0 3448.8 3361.4 4620.9 2453.6 Std 2496.1 1930.3 3294.9 2651.4 2398.5 3045.4 Min 0.0 510.4 0.0 0.0 0.0 0.0 Max 11,162.0 13,180.6 17,540.7 12,427.0 13,933.8 16,162.3 Avg average of monthly streamflow in 10 3 m 3 , Std standard deviation of monthly streamflow in 10 3 m 3 , Min minimum of monthly streamflow in 10 3 m 3 , Max maximum of monthly streamflow in 10 3 m 3 Potential improvements to statistical downscaling of streamflows precipitation of period 19501989. Then, the HadCM3, ECHAM5 and GFDL2.0 outputs pertaining to the A2 and B1 GHG emission scenarios for SDM (lag0_preci) corresponding to future climate were also standardised for each calendar month using the means and the standard deviations of NCEP/NCAR reanalysis outputs of period 19501989. Using the above standardised precipitation of the PDM and outputs of HadCM3, ECHAM5 and GFDL2.0 pertaining to A2 and B1 GHG emission scenarios as inputs to SDM (lag0_preci) , pro- jections of streamflowfor each calendar month was produced. Table 8 shows the seasonal statistics of observed and projected streamflows (inflow to Lake Bellfield in north- western Victoria, Australia) corresponding to A2 and B1 GHG emission scenarios for periods 19501999 and 2000 2099, respectively. Also, the ensemble average streamflow time series was computed from the streamflow outputs of SDM (lag0_preci) when it was run with the outputs of HadCM3, ECHAM5 and GFDL2.0 and the precipitation outputs of PDM. Statistics of streamflow derived from the ensemble average streamflowtime series are also shown in Table 8. Note that in Table 8, changes in the average of precipitation pro- duced by PDM over the period 20002099 with respect to the average of observed precipitation in the period 19501999 are also provided (for more details, refer to footnote of Table 8). According to Table 8, it was seen that the average of streamflowcorresponding to both A2 and B1 emission scenarios, indicated a decline in summer, autumn and spring and showed an increase in winter, for all GCMs and for the ensemble. However, as shown in Table 8, the precipitation outputs of the PDMdid not showsuch decline in summer, autumn and spring and an increase in winter for all three GCMs and for both GHG emission scenarios. Instead, average of precipitation projected by PDM indicated mixed results in above seasons, depending on the GCM and the GHG emission scenario of interest. It was realised that an increase (or decrease) in precipitation in a season does not guarantee an increase (or decrease) in the streamflow in that season. This is possibly because though precipitation is the main driver of streamflow, variables such as soil moisture content and atmospheric moisture content (influence the evaporation rate) can significantly influence the streamflow generation process. As an example, a dry catchment (i.e. low soil moisture content and atmospheric moisture content) can absorb large amount of the precipitation and cause either small increase or even some decrease in the average of streamflow. Similar to precipitation, streamflow also showed a rise in its standard deviation and the maximum, in all seasons corre- sponding to all three GCMs and both GHG emission scenar- ios. This indicated that there will be more fluctuations in the Table 8 Seasonal statistics of observed and projected streamflow Season Statistic Observed (195099) HadCM3 (200099) ECHAM5 (200099) GFDL2.0 (200099) Ensemble average (200099) A2 B1 A2 B1 A2 B1 A2 B1 Summer Avg 835.9 421.7 a 758.9 a 767.1 b 764.3 b 775.3 b 772.4 a 654.7 c 765.2 a Std 850.4 808.7 988.6 969.8 962.3 928.6 923.2 641.0 905.9 Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Max 8041.0 9275.7 11,055.1 11,230.5 11,000.3 10,911.7 10,716.8 7380.7 10,924.0 Autumn Avg 603.2 524.8 b 499.8 a 512.0 c 504.9 b 484.8 a 491.6 a 507.2 a 498.8 a Std 936.1 1054.1 1047.2 1046.4 1047.7 1046.2 1046.0 1030.4 1026.6 Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Max 8017.0 9260.3 9268.0 9258.7 9258.7 9261.1 9257.8 9260.0 9261.5 Winter Avg 3394.3 3797.1 b 3823.7 b 3487.4 a 3588.7 a 3736.7 c 3869.1 b 3673.7 a 3760.5 b Std 2496.1 3184.5 2891.5 2739.8 2804.7 3023.2 3017.3 2612.9 2474.2 Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 29.2 Max 11,162.0 19,964.9 15,435.5 14,753.4 14,509.1 14,229.4 13,849.2 15,494.7 13,426.9 Spring Avg 3361.4 2509.2 a 2375.5 a 2051.7 a 2173.2 a 2812.3 b 2965.9 b 2457.8 a 2504.9 a Std 2651.4 3042.3 2734.5 2579.5 2533.1 2922.6 2949.3 2384.5 2242.8 Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Max 12,427.0 15,863.1 13,841.3 13,722.9 12,768.3 14,133.1 14,944.9 12,500.1 11,424.7 Avg average of monthly streamflow in 10 3 m 3 , Std standard deviation of monthly streamflow in 10 3 m 3 , Min minimum of monthly streamflow in 10 3 m 3 , Max maximumof monthly streamflowin 10 3 m 3 , Ensemble average average time series computed fromthe outputs of SDM (lag0_preci) when it was run with the outputs of HadCM3, ECHAM5 and GFDL2.0 and precipitation outputs of PDM a Average of precipitation showed a decrease with respect to that of observations of period 195099 b Average of precipitation showed an increase with respect to that of observations of period 195099 c Average of precipitation did not show any change with respect to that of observations of period 195099 D.A. Sachindra et al. streamflow regime in the future and the magnitude of the extremely high streamflows tend to increase in all seasons. Also, it was noticed that in general, the maximum of streamflow was higher for the A2 GHG emission scenarios compared to that of B1, for the majority of GCMs in all seasons. Figure 6 shows the probability exceedance curves for ob- served streamflow for the period 19501999 and the proba- bility exceedance curves for streamflow projected into the future period 20002099 derived using the multi-model en- semble average streamflowtime series, for each season. It was clear that in all seasons, except spring, the magnitude of the extremely high streamflow tends to show a rise for both A2 and B1 emission scenarios. In summer, for most of the ex- ceedance probabilities (particularly for smaller exceedance probabilities), streamflow shows a decrease under both sce- narios; however still the extremely high streamflow tends to show a rise. This decrease in streamflow was probably due to the dryness of the catchment in summer caused by the decline in precipitation in spring (see Table 8 for decline in average precipitation in spring). In autumn, for smaller exceedance probabilities (high streamflows), streamflowshowed a rising trend and for higher exceedance probabilities (lowstreamflows) a decreasing trend was seen. In winter, streamflow for both A2 and B1 GHG emission scenarios showed a rising trend for the majority of the exceedance probabilities (low to high streamflows). In spring, for the majority of exceedance probabilities streamflow showed a decrease corresponding to both A2 and B1 GHG emission scenarios. Furthermore, according to the exceedance curves in Fig. 6, it was realised that in the future there will be more months with zero flows particularly in summer and autumn. The long-term seasonal statistics (e.g. average, standard deviation, minimum and maximum) of monthly streamflows obtained from the bias-corrected time series are useful for the management of water resources in a catchment. The average of streamflow in each season determined in this study pro- vides an idea of the availability of water in Lake Bellfield in the future, hence the availability of water to Halls Gap and Pomonal towns and also for the recreational activities which take place at this lake can be determined. This allows the appropriate allocation of water to various needs depending on the availability. The standard deviation of streamflow shows the degree of fluctuations in inflow to Lake Bellfield in the future. In this study, a rise in the standard deviation of inflowto Lake Bellfield was seen in all seasons. The rise in the standard deviation of inflow to Lake Bellfield indicated more fluctuations in the streamflow regime in the future. These fluctuations in the inflow to the lake should be taken into account in its future operations and any modification, as they impact the reliability of water supply to customers. The knowledge of the extremes in the streamflowregime is helpful in the management of droughts and floods. Since the number of months with zero inflow and the magnitude of the peak inflow to Lake Bellfield in all seasons have increased, new drought and flood mitigation measures may be needed in the management of droughts and floods in the catchment in the future. 4.5 Uncertainties associated with streamflow projections It should be noted that the projections of streamflows pro- duced in this study using statistical downscaling are subject to a cascade of uncertainties originating from a number of sources. These sources of uncertainties include; GHG emis- sion scenarios, GCMs, the downscaling technique, methodol- ogy followed in developing the downscaling model (e.g. predictor selection and pre-processing, selection of calibration period) and predictor-predictand stationarity assumption (Sachindra et al. 2014d). The actual amounts of GHG emissions in the future world are unknown as they are dependent on a number of factors such as population, technological development and govern- ment policies, which can largely change in the future over time. Therefore, several equally likely but different GHG emission scenarios have been defined for the future (e.g. SRES GHG emission scenarios (IPCC 2000), Representative concentration pathways (van Vuuren et al. 2011)). The use of several equally likely but different GHGemission scenarios in producing catchment scale hydroclimatic projections into the future enables the quantification of uncertainties introduced by GHG emissions to the projections. In this study, the pro- jections of streamflows were produced corresponding to the A2 and B1 GHG emission scenarios for the quantification of relatively high and low impacts of GHG emissions on streamflows in the future. The streamflow projections pro- duced in this study should be treated as plausible rather than definite as the GHG emission scenarios on which they are based are on are plausible realisations of future GHG emissions. The projections produced using a downscaling model can vary from GCM to GCM as the approximations and assump- tions employed in the structures of GCMs vary from one another. In the present study, outputs of HadCM3, ECHAM5 and GFDL2.0 were used to produce inputs to the downscaling models. Using sets of outputs from different GCMs on the downscaling models and hence deriving an ensemble of pro- jections can better explain the uncertainties (e.g. visualize the upper and lower uncertainty bounds of the projections of the predictand) introduced by different GCMs to the projections of the predictand of interest. Furthermore, combination of ensemble of projections using an ensemble modelling tech- nique (e.g. averaging) can reduce the dependence of projec- tions of one specific GCM. In this study, for the derivation of the ensemble projections, the simple averaging technique was Potential improvements to statistical downscaling of streamflows used. In simple averaging, each projection produced by the statistical downscaling model is assigned the same weightage. In other words, it is assumed that all GCMs perform equally. However, this assumption can be coarse and assigning weightages to each projection based on the performance of each GCM for each calendar month and hence deriving an ensemble projection can be regarded as a better approach (Zhang and Huang 2013). However, such procedure is com- putationally expensive. Since different downscaling techniques can represent the predictor-predictand relationships differently, the downscaling technique used in a study can also introduce a certain degree of uncertainty to the projections. However, in comparison to the uncertainties introduced by the GHG emission scenarios and GCMs, the uncertainties introduced by different downscaling techniques are negligible. Sachindra et al (2013) used LS-SVM and MLR for statistically downscaling monthly GCM outputs to monthly streamflows and commented that both techniques yielded similar results. Also, Tripathi et al. (2006) found that LS-SVM is marginally better than ANN in downscaling GCM outputs to precipitation. Furthermore, they commented that both tech- niques failed to correctly capture the extremes of precipitation. The methodology used in the development of a downscal- ing model can also introduce some uncertainty to the projec- tion. The overall methodology of a downscaling exercise includes the selection of potential predictors, standardisation of predictor data, selection of the calibration period and model calibration and validation. Each of the above steps can be performed in a number of different manners. As an example, in a downscaling exercise, instead of using traditional calibra- tion and validation approach, a cross-validation approach can be adopted. Therefore, depending on the methodology employed, the outputs of a downscaling model can vary. Almost all statistical downscaling models are dependent on the assumption that the relationships derived between the predictors and predictands for the past climate will be valid for the future, under changing climate. In other words, the validity of the stationarity of predictor-predictand relation- ships under non-stationary climate is assumed. However, the validity of the above assumption is questionable under non- stationary climate. A few attempts to handle the non- 0 2000 4000 6000 8000 10000 12000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Observed streamflow in summer 1950-1999 Multi-model ensemble projected streamflow in summer A2 2000-2099 Multi-model ensemble projected streamflow in summer B1 2000-2099 Exceedance probability S t r e a m f l o w / ( x
1 0 3 m 3 / m o n t h ) (a) Summer 0 2000 4000 6000 8000 10000 12000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Observed streamflow in autumn 1950-1999 Multi-model ensemble projected streamflow in autumn A2 2000-2099 Multi-model ensemble projected streamflow in autumn B1 2000-2099 Exceedance probability S t r e a m f l o w / ( x
1 0 3 m 3 / m o n t h ) (b) Autumn 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Observed streamflow in winter 1950-1999 Multi-model ensemble projected streamflow in winter A2 2000-2099 Multi-model ensemble projected streamflow in winter B1 2000-2099 Exceedance probability S t r e a m f l o w / ( x
1 0 3 m 3 / m o n t h ) (c) Winter 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Observed streamflow in spring 1950-1999 Multi-model ensemble projected streamflow in spring A2 2000-2099 Multi-model ensemble projected streamflow in spring B1 2000-2099 Exceedance probability S t r e a m f l o w / ( x
1 0 3 m 3 / m o n t h ) (d) Spring Fig. 6 Seasonal probability exceedance curves for observed streamflow and bias-corrected streamflow for future climate under A2 and B1 GHG emission scenarios D.A. Sachindra et al. stationarity of predictor-predictand relationships under non- stationary climate are documented in Raje and Mujumdar (2010), Duan et al (2012) and Hertig and Jacobeit (2013). In this study, like most of the other downscaling studies, it was assumed that the predictor-predictand relationships will be stationary under changing climate (non-stationary climate). Further investigation is needed for the confirmation of the impact of the above assumption on the streamflowprojections produced in this study. Furthermore, it should be noted that in direct downscaling of GCM outputs to catchment scale streamflows, the changes in the land use patterns and water diversions in and out of a catchment are not considered as such catchment scale changes are not characterised in the GCM outputs. 5 Conclusions The following conclusions were drawn from this study: 1. When the precipitation downscaled from large-scale at- mospheric variables is used as an additional input to a statistical model developed for downscaling large-scale atmospheric variables to streamflows in a catchment, its performances improved by a small amount. However, the use of observed precipitation as an additional input to that streamflow downscaling model increased its perfor- mances significantly. Then, it was realised that any im- provement to the precipitation downscaling model will introduce an improvement to the streamflow downscaling model, when it is used with improved precipitation predictions. 2. If the statistical model developed for downscaling large- scale atmospheric variables to streamflows in a catchment already contains atmospheric humidity and soil moisture variables, inclusion of evaporation as an additional input to that downscaling model may not lead to any improve- ment in model performance. This is because atmospheric humidity and soil moisture variables can indirectly ex- plain the influence of evaporation on streamflow. There- fore, when evaporation is introduced to a streamflow downscaling model, it brings redundant information to the model, leading to no significant improvement to the model performance. 3. A mismatch was seen between the statistics of observed streamflow for the past climate and those of streamflow reproduced by the downscaling model when run with the 20C3M of GCMs and the precipitation downscaled using the 20C3M outputs of GCMs as inputs. It indicated the presence of bias in the 20C3M outputs of GCMs, and the need of a correction to bias was identified. 4. The equidistant quantile mapping technique corrected the average, the standard deviation, the minimum and the maximum of streamflow reproduced by a downscaling model pertaining to the past climate with high degree of accuracy. This is because the cumulative distribution function of the model reproduced streamflow is mapped directly onto the cumulative distribution function of the corresponding observations. The validation of the equi- distant quantile mapping technique also showed reduction in the bias in the statistics of streamflow produced by the downscaling model. 5. Though the statistics of streamflow downscaled using the GCM outputs can be successfully corrected using the equidistant quantile mapping technique, the scatter of downscaled streamflow was reduced only by a small amount. This indicated that the improvement to the time series of the variable of interest (e.g. streamflow) is small. Therefore, after the bias-correction based on the equidis- tant quantile mapping technique, the streamflows projected into the future should be interpreted in terms of statistics and probability distributions rather than a time series. Acknowledgments The authors wish to acknowledge the financial assistance provided by the Australian Research Council Linkage Grant scheme and Grampians Wimmera Mallee Water Corporation for this project. References Anandhi A, Srinivas VV, Nanjundiah RS, Kumar DN (2008) Downscaling precipitation to river basin in India for IPCC SRES scenarios using support vector machine. Int J Climatol 28:401420 Bastola S, Misra V (2013) Evaluation of dynamically downscaled reanal- ysis precipitation data for hydrological application. Hydrol Process. doi:10.1002/hyp.9734 Bedia J, Herrera S, Martn DS, Koutsias N, Gutirrez JM (2013) Robust projections of fire weather index in the Mediterranean using statis- tical downscaling. Clim Chang 120:229247 Benestad R, Hanssen-Bauer I, Chen D (2008) Empirical-statistical down- scaling. World Scientific, Singapore, p 228 Cannon AJ, Whitfield PH (2002) Downscaling recent streamflow condi- tions in British Columbia, Canada using ensemble neural network models. J Hydrol 259:136151 Can J, Domnguez F, Valds JB(2011) Downscaling climate variability associated with quasi-periodic climate signals: a new statistical approach using MSSA. J Hydrol 398:6575 Cayley RA, Taylor DH (1997) Grampians special map area and geolog- ical report. Geol Surv Victoria Rep 107:42130 Denis B, Laprise R, Cay D, Ct J (2002) Downscaling ability of one- way nested regional climate models: the big-brother experiment. Clim Dynam 18:627646 Duan J, McIntyre N, Onof C. (2012) Resolving non-stationarity in statistical downscaling of precipitation under climate change sce- narios. British Hydrological Society Eleventh National Symposium, Hydrology for a changing world, Dundee, UK Potential improvements to statistical downscaling of streamflows Flint LE, Flint AL (2012) Downscaling future climate scenarios to fine scales for hydrologic and ecological modelling and analysis. Ecol Process 1:2 Fowler HJ, Blenkinsop S, Tebaldi C (2007) Linking climate change modelling to impacts studies: recent advances in downscaling tech- niques for hydrological modelling. Int J Climatol 27:15471578 Ghosh S, Mujumdar PP (2008) Statistical downscaling of GCM simula- tions to streamflow using relevance vector machine. Adv Water Resour 31:132146 Goyal MK, Ojha CSP (2012) Downscaling of surface temperature for lake catchment in an arid region in India using linear multiple regression and neural networks. Int J Climatol 32:552566 Gutirrez JM, San-Martn D, Brands S, Manzanas R, Herrera S (2013) Reassessing statistical downscaling techniques for their robust ap- plication under climate change conditions. J Clim 26:171188 GWMWater (2011a) Lake Bellfield operating rules Available online at http://www.gwmwater.org.au/information/publications/ground-and- surface-water/west-wimmera-gma/cat_view/163-reservoir- operating-rules GWMWater (2011b) Storage management rules for the Wimmera Mallee system headworks. Available online at http://www.gwmwater.org. au/information/publications/ground-and-surface-water/west- wimmera-gma/cat_view/163-reservoir-operating-rules. 56 Haas R, Pinto JG (2012) A combined statistical and dynamical approach for downscaling large-scale footprints of European windstorms. J Geophys Res Lett 39:L23804 Hertig E, Jacobeit J (2008) Assessments of Mediterranean precipitation changes for the 21 st century using statistical downscaling tech- niques. Int J Climatol 28:10251045 Hertig E, Jacobeit J (2013) A novel approach to statistical downscaling considering nonstationarities: application to daily precipitation in the Mediterranean area. J Geophys Res Atmos 118:520533 Horvath K, Koracin D, Vellore RK, Jiang J, Belu R (2012) Sub-kilometre dynamical downscaling of near-surface winds in complex terrain using WRF and MM5 mesoscale models. J Geophys Res 117:D11 IPCC (2000) IPCC special report on emissions scenariossummary for policymakers. Online report available at http://www.ipcc.ch/pdf/ special-reports/spm/sres-en.pdf. 4-8. Jeong DI, St-Hilaire A, Ouarda TBMJ, Gachon P (2012) Multisite statistical downscaling model for daily precipitation combined by multivariate multiple linear regression and stochastic weather gen- erator. Clim Chang 114:567591 Karl TR, Wang WC, Schlesinger ME, Knight RW, Portman D (1990) A method of relating general circulation model simulated climate to the observed local climate part I: seasonal statistics. J Clim 3:1053 1079 Landman WA, Mason SJ, Tyson PD, Tennant WJ (2001) Statistical downscaling of GCM simulations to streamflow. J Hydrol 252: 221236 Li H, Sheffield J, Wood EF (2010) Bias correction of monthly precipita- tion and temperature fields from intergovernmental panel on climate change AR4 models using equidistant quantile matching. J Geophys Res-Atmos 115:120 Maurer EP, Hidalgo HG (2008) Utility of daily vs. monthly large-scale climate data: an intercomparison of two statistical downscaling methods. Hydrol Earth Syst Sci 12:551563 Murphy J (1998) An evaluation of statistical and dynamical techniques for downscaling local climate. Int J Climatol 12:22562284 Nash JE, Sutcliffe JV (1970) River flow forecasting through conceptual models, part 1a discussion of principles. J Hydrol 10:282290 Nasseri M, Tavakol-Davani H, Zahraie B(2013) Performance assessment of different data mining methods in statistical downscaling of daily precipitation. J Hydrol 492:114 Pearson K (1895) Mathematical contributions to the theory of evolution. iii. Regression heredity and panmixia. Philos T Roy Soc A187:253 318 Qian JH, Zubair L (2010) The effect of grid spacing and domain size on the quality of ensemble regional climate downscaling over South Asia during the north-easterly monsoon. Mon Weather Rev 138: 27802802 Raje D, Mujumdar PP (2010) Constraining uncertainty in regional hy- drologic impacts of climate change: nonstationarity in downscaling. Water Resour Res 46:W07543 Rojas M (2006) Multiply nested regional climate simulation for Southern America: sensitivity to model resolution. Mon Weather Rev 134: 22082223 Rummukainen M (2010) State-of-the-art with regional climate models. Wiley Interdiscip Rev Clim Chang 1:8296 Sachindra DA, Huang F, Barton AF, Perera BJC (2013) Least square support vector and multi-linear regression for statistically downscal- ing general circulation model outputs to catchment streamflows. Int J Climatol 33:10871106 Sachindra DA, Huang F, Barton AF, Perera BJC (2014a) Statistical downscaling of general circulation model outputs to precipitation part 1: calibration and validation. Int J Climatol 34:32643281 Sachindra DA, Huang F, Barton AF, Perera BJC (2014b) Statistical downscaling of general circulation model outputs to precipitation part 2: bias-correction and future projections. Int J Climatol 34: 32823303 Sachindra DA, Huang F, Barton AF, Perera BJC (2014c) Multi-model ensemble approach for statistically downscaling general circulation model outputs to precipitation. QJ Roy Meteor Soc 140:11611178. doi:10.1002/qj.2205 Sachindra DA, Huang F, Barton AF, Perera BJC (2014d) Statistical downscaling of general circulation model outputs to catchment scale hydroclimatic variables issues, challenges and possible solutions. J Water Clim Change. doi:10.2166/wcc.2014.056 Salvi K, Kannan S, Ghosh S (2011) Statistical downscaling and bias- correction for projections of Indian rainfall and temperature in climate change studies. Proceedings of 4 th International Conference on Environmental and Computer Science. 1618 September 2011, Singapore, pp.7-11 Samadi S, Carbone GJ, Mahdavi M, Sharifi F, Bihamta MR (2012) Statistical downscaling of climate data to estimate streamflow in a semi-arid catchment. Hydrol Earth Syst Sci Discuss 9:48694918 Shang KZ, Wang SG, Ma YX, Zhou ZJ, Wang JY, Liu HL, Wang YQ (2007) A scheme for calculating soil moisture content by using routine weather data. Atmos Chem Phys 7:51975206 Sinclair KM (2004) Wimmera-Malle simulation modelAnnual update methodology. A report by Sinclair, Knight Merz, Melbourne, Australia. 2179 Smith I, Chandler E (2010) Refining rainfall projections for the Murray Darling Basin of south-east Australiathe effect of sampling model results based on performance. Clim Chang 102:377393 Tareghian R, Rasmussen PF (2013) Statistical downscaling of precipita- tion using quantile regression. J Hydrol 487:122135 Timbal B, Fernandez E, Li Z (2009) Generalization of a statistical downscaling model to provide local climate change projections for Australia. Environ Model Softw 24:341358 Tisseuil C, Vrac M, Lek S, Wade AJ (2010) Statistical downscaling of river flows. J Hydrol 385:279291 Tripathi S, Srinivas VV, Nanjundiah RS (2006) Downscaling of precip- itation for climate change scenarios: a support vector machine approach. J Hydrol 330:621640 van Vuuren PD, Edmonds JA, Kainuma M, Riahi K, Thomson AM, Hibbard K, Hurtt GC, Kram T, Krey V, Lamarque JF, Masui T, Meinshausen M, Nakicenovic N, Smith SJ, Rose S (2011) The repre- sentative concentration pathways: an overview. Clim Chang 109:531 Wang P, Yamanaka T, Qiu GY (2012) Causes of decreased reference evapotranspiration and pan evaporation in the Jinghe River catch- ment, northern China. Environmentalist 32:110 D.A. Sachindra et al. Wilby RL, Fowler HJ (2011) Regional climate modelling in modelling the impact of climate change on water resources. Modelling the impact of climate change on water resources. F. Fung, A. Lopez and M. New, eds., Wiley-Blackwell, Chichester, UK. 209 Wilby RL, Charles SP, Zorita E, Timbal B, Whetton P, Mearns LO (2004) Guidelines for use of climate scenarios developed fromstatistical down- scaling methods, supporting material to the IPCC. Available online at http://www.ipcc-data.org/. 321. (Accessed 28 th January 2012) Wilks DS (1999) Multisite downscaling of daily precipitation with a stochastic weather generator. Climate Res 11:125136 Yang T, Li H, Wang W, Xu CY, Yu Z (2012) Statistical downscaling of extreme daily precipitation, evaporation, and temperature and con- struction of future scenarios. Hydrol Process 26:35103523 Zhang H, Huang GH (2013) Development of climate change projections for small watersheds using multi-model ensemble simulation and stochastic weather generation. Clim Dynam 40:805821 Potential improvements to statistical downscaling of streamflows