Impact of GCM structure uncertainty on hydrological processes in an arid area of China

The Intergovernmental Panel on Climate Change (IPCC 2014) stated that the precipitation and temperature patterns would significantly change by the end of the 21st century. Climate change has a wide and profound impact on hydrological processes. Changes in hydrological processes and increased extreme events (e.g., drought and flood) have been detected and have exerted significant impacts on ecological and social systems (IPCC 2014). Therefore, understanding the hydro-climatic effects of future climate change is critical to local water management, especially for arid regions, where hydrological changes are more sensitive to climate change than those humid regions.

When accessing the impact of future climate change on hydrology, climate models are often coupled with hydrological models (HMs) to predict future changes in hydrological processes (Liu et al. 2010; Ficklin et al. 2013; IPCC 2014). Climate change projected by state-of-the-art general circulation models (GCMs) suggested a warming temperature trend along with seasonally and spatially varying precipitations for the 21st century (Reyers et al. 2013). As uncertainty is inherited in modelling, it is necessary to consider the uncertainties from GCMs, climate variable downscaling and hydrological modelling (Graham et al. 2007; Jiang et al. 2007; Chen et al. 2011, 2012) in local impact studies. Previous studies have investigated different uncertainty sources, and most of them noted that GCMs are one of the greatest sources of uncertainty in assessing climate change impact on hydrological processes (Horton et al. 2006; Wilby & Harris 2006; Graham et al. 2007; Chen et al. 2011, 2016; Dobler et al. 2012; Bosshard et al. 2013; Lung et al. 2013; Exbrayat et al. 2014), although uncertainty from HMs was also important over many areas of the world (Hagemann et al. 2013). More specifically, major contributors to uncertainty depend on the assessed hydrological variables (Booij 2005; Chen et al. 2011; Gampe et al. 2016; Shrestha et al. 2016) and the assessed watersheds (Finger et al. 2012; Lutz et al. 2013; Ragettli et al. 2013; Addor et al. 2014; Huss et al. 2014; Vidal et al. 2016). Methodology used for uncertainty decomposition includes qualitative visual interpretation of the prediction spread (Wilby & Harris 2006; Graham et al. 2007; Chen et al. 2011) and the quantitative standard deviation method (Xu & Xu 2012) and analysis of variance (ANOVA) approach (Bosshard et al. 2013; Addor et al. 2014; Duethmann et al. 2016). Although these studies have been conducted in many regions throughout the world, the differences between these uncertainty decomposition methods have been seldom compared.

The Tienshan Mountains, ‘water tower’ of central Asia, are the main water sources and ecological barriers, and very typical in terms of the dry and alpine continental climate characteristics together with data scarcity. The Kaidu River, a typical watershed located in the south slope of the Tienshan Mountains, is one of the headwaters of the Tarim River, the largest endorheic basin in China. Understanding future hydrological processes and their related uncertainty help the sustainable development of countries along the ‘Silk Road’ (Li et al. 2015). In the literature, most studies in this area have focused on historical hydrological events (Shi et al. 2007; Liu et al. 2010; Piao et al. 2010; Chen 2014; Rumbaur et al. 2015). Although there have been some investigations for future scenarios (Liu et al. 2010, 2011; Sorg et al. 2012; Fang et al. 2015a; Xu et al. 2016), few of them (Duethmann et al. 2016) quantified the uncertainty sources. Characterizing the uncertainty sources is of high importance for a valid interpretation of the results.

This paper aims to investigate the impact of climate change on the hydrological system and assess the impact of GCM structural uncertainty on hydrological processes. To this end, a cascade of a GCM ensemble, downscaling methods and a HM were used to simulate future hydrological processes. Three main questions are addressed: (1) How will the future climate and hydrological processes change in this arid mountainous region? (2) Which one of the following issues contributes most to future hydrological processes: GCM structure uncertainty, representative concentration pathways (RCPs) or downscaling methods? (3) Do different uncertainty decomposition methods produce different results? Understanding these issues will enable us to better assess future hydrological changes and related uncertainties. The paper is organized as follows: the section below introduces the study area; the next section describes the cascade of the GCM ensemble, bias correction methods, HM and the methodology on how to decompose the uncertainty sources; a results and discussion section follows and the final section drawing conclusions.

The Kaidu River Basin (Figure 1), with a drainage area of 18,634 km² above the Dashankou hydrological station, is one of the four headwaters of the Tarim River. It originates in the Tienshan Mountains. Recharged mainly by rainfall and snowmelt (SM), the Kaidu River provides 42 × 10⁸ m³ amount of water for agricultural irrigation and ecological water conveyance for the lower reaches of the Tarim River, which is crucial to the local eco-environmental and economic development. The altitude ranges from 1,340 m to 4,796 m above sea level (a.s.l.) with an average elevation of 2,995 m and average slope of 23%. This watershed has a temperate continental climate with alpine characteristics. The average annual temperature at the Bayanbulak meteorological station is −4.1 °C and annual precipitation is 278 mm; precipitation generally falls as rain from May to September and as snow from October to April of the next year. The average daily flow at the Dashankou hydrological station is around 120 m³/s (equivalent to 201 mm runoff), ranging from 15 m³/s to 973 m³/s.

Figure 2 presents the framework for the hydrological modelling under future climate change. First, daily climate predictions from 21 GCM models from CMIP5 (Coupled Model Intercomparison Project Phase 5) under RCP4.5 and RCP8.5 were downloaded (http://cmip-pcmdi.llnl.gov/cmip5/; IPCC 2013), and then these grid-based climate predictions were downscaled/bias-corrected to the station scale using four precipitation (BC_pcp) and three temperature (BC_tmp) bias correction methods. These bias-corrected climatic variables were used to force the well-calibrated HM of the Soil and Water Assessment Tool (SWAT). Compared to other sources (e.g., model parameter, model structure), meteorological input contributes to the largest part of uncertainties in hydrological modelling (Wilby & Harris 2006; Graham et al. 2007; Chen et al. 2011, 2016; Bosshard et al. 2013), therefore, we only used one HM in this study. In total, there were 504 hydrological simulations from 504 different combined meteorological inputs (i.e., different combinations of 21 GCMs, 2 RCPs, 4 BC_pcps and 3 BC_tmps), as shown in Figure 2.

The state-of-the-art climate change projections of 21 GCMs under two emission scenarios (RCP4.5 and RCP8.5) were used as climatic data in this study (Table 1). These models are from over 20 institutes or universities. Of all GCM simulated climate variables, daily precipitation, and maximum and minimum temperatures from 1975 to 2099 were used.

RCP4.5 (lower emission scenario) is a stabilization scenario with the total radiative forcing rising until 2070, which will remain stable at 4.5 W/m². In contrast, RCP8.5 (higher emission scenario) is a continuously rising radiative forcing pathway (at a target of 8.5 W/m² in 2,100) with a further enhanced residual circulation and significant CH₄ increase (Van Vuuren et al. 2011). RCP4.5 and RCP8.5 are equivalent to B1 and A2 of the Special Report on Emission Scenarios (SRES).

To account for the low spatial resolution in GCM outputs, four BC_pcp and three BC_tmp methods were used to downscale the grid-based GCM outputs to the station scale (where there is a meteorological station). These correction methods are local intensity scaling (LOCI), power transformation (PT), distribution mapping (DM) and quantile mapping (QM) for precipitation as well as linear scaling (LS), variance scaling (VARI) and DM for temperature. These methods can be classified into mean-based (LS and LOCI), variance-based (PT and VARI) and distribution-based approaches (DM and QM). Table 2 briefly describes the characteristics of each method. The methods have been widely used in downscaling and bias correcting the climate model outputs (e.g., Schmidli et al. 2006; Fang et al. 2015b).

SWAT (Arnold et al. 1998), developed at the Agriculture Research Service of the United States Department of Agriculture, has been widely used for comprehensive modelling of the impacts of management practices and climate change on hydrological processes at a watershed scale (e.g., Jayakrishnan et al. 2005; Singh et al. 2015; Awan et al. 2016; Tamm et al. 2016). To represent the spatial variability, a watershed is first disaggregated into subbasins and each subbasin is further divided into hydrological response units based on soil and land use data. For more details, refer to SWAT manuals (http://www.brc.tamus.edu/).

The SWAT model was successfully applied in the Kaidu River Basin (Fang et al. 2015c). The SWAT model was first set up with digital elevation model (DEM) (www2.jpl.nasa.gov/srtm/), land use (from the Environmental and Ecological Science Data Centre for West China), soil map (from Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences), and observed meteorological data (at two meteorological stations Bayanbulak and Baluntai, Figure 1; China Meteorological Data Sharing Service System), which form the base for estimating meteorological input for each subbasin with the elevation band method. Then, model parameters were calibrated with the observed streamflow data at Dashankou station (hydrological station in Figure 1) and good model performances were achieved for both calibration and validation periods, as shown in Table 3. More details of the SWAT model setup and calibration can be found in Fang et al. (2015c).

When studying the impact of climate change on flow extremes, we used average annual 3-day maximum high flow (3DMHF) and average annual 7-day minimum low flow (7DMLF), as used in Sanborn & Bledsoe (2006).

To analyse the impact of climate change on hydrological processes, five periods are defined: 1986–2005 (control period), 2020–2039, 2040–2059, 2060–2079 and 2080–2099. For future changes of the precipitation, temperature and streamflow, we calculated the ensemble median and 25th and 75th percentiles instead of the mean value because these statistics have a lower sensitivity to outliers than the ensemble mean (Benestad 2004).

There have been several uncertainty quantification methods applied in climate change impact studies, as proposed in the ‘Introduction’. Standard deviation and ANOVA approach are two quantitative methods commonly used in assessing the contributions of GCMs, downscaling methods, RCPs and HMs to the total variance of hydrological variables within an ensemble of climate models (Yip et al. 2011; Déqué et al. 2012; Gampe et al. 2016; Vidal et al. 2016). Here, we used these two methods to decompose the contributions of GCMs, RCPs, BC_pcp and BC_tmp to prediction uncertainty.

The larger S/N is, the higher the credibility and more significant are the changes in the projected streamflow. Normally, S/N > 1 indicates that the projections are credible to a certain extent while S/N < 1 indicates low credibility or with unsignificant changes (Zhou & Yu 2006).

Further, the consistency of the simulations is estimated as the ratio of the number of simulations with negative projected relative streamflow changes to the number of all simulations (N = 504 in this case).

Evaluation statistics in Table 3 indicate good model performances forced by the observed climate variables: daily Nash–Sutcliffe coefficients (NS) (Nash & Sutcliffe 1970) larger than 0.8 and percent biases (PBIAS) within ±10% (Table 3) for both calibration and validation periods. For the high flow and low flow, ‘PBIAS’s of the 3DMHF and 7DMLF are −19.6% and 7.5%, respectively, during 1990–2002. The validation period is three times longer than the calibration period, indicating the calibrated HM is robust, and therefore could be used to study the impact of climate change on local hydrological processes.

Figure 3 compares observed streamflow series, and simulated streamflow series driven by observed meteorological inputs and bias-corrected GCM outputs (combinations of 21 GCMs, 4 ‘BC_pcp’s and 3 ‘BC_tmp’s) in the control period. The 90% percentile (shaded) of simulated streamflow driven with bias-corrected GCM outputs bracketed both the monthly average observations (left plot) and the observed streamflows at each exceedance (right plot). In addition, ‘PBIAS’s for 3DMHF and 7DMLF ranged from −39.4% to 77.6% and −55.5% to 150.4% with 67% and 34% of these simulations having absolute PBIAS values within 20%. The model performances show generally satisfactory results and could be used to study the impact of climate change.

The projected precipitation and temperature changes are presented in Figure 4. The medians of annual precipitation change are 8% and 16%, while their 25% and 75% quantiles are 3.1%–18% and 7.0%–22.5% under RCP4.5 and RCP8.5, respectively, for 2080–2099. There is a significant seasonal variation in the precipitation change with a substantial increase in the cold season (November to April of the next year) and small increase during summer (June, July and August) with large differences among GCMs. The uncertainty bands of projected precipitation gradually increases throughout the 21st century.

For temperature, all GCMs projected a continuous increasing trend with the median increasements being 3.6 °C and 6.5 °C, and their 25% and 75% quantiles ranging from 2.0 °C to 3.3 °C and 4.2 °C to 5.5 °C, respectively, under RCP4.5 and RCP8.5 for 2080–2099 (Figure 4). Compared to precipitation, uncertainty in temperature projections are considerably smaller.

Changes in the precipitation and temperature led to changes in the streamflow. The following results are based on 504 simulations forced by all combinations of RCPs, GCMs, BC_tmp and BC_pcp. Figure 4 shows the streamflow changes with their 25% and 75% quantiles. This indicates there is no general conclusion that these changes are definitely positive or negative or have the same magnitude. The largest decrease will be likely to occur during 2080–2099 under RCP8.5. The medians of annual streamflow change are −12.5% and −18%, while their 25% and 75% quantiles are −26% to 3.4% and −38% to −7% under RCP4.5 and RCP8.5, respectively, for 2080–2099. Note that most models project decreasing streamflow after the 2060s due to the continuously rising temperature. Seasonally, monthly average streamflow decreases by 15% and 27% for the summer season (June, July and August) while it increases by 3.0% and 3.7% for spring (March, April and May) under RCP4.5 and RCP8.5 in 2080–2099.

In a previous study conducted in the Kaidu River Basin, Liu et al. (2011) reached the conclusion that the qualitative impact results were highly consistent, while they are not in our case. As we used 21 GCMs and different bias correction methods, the climate change scenarios were expected to have larger ranges. This is further discussed based on credibility and consistency indices in the text below. Differences can also be found in the streamflow projections during the snow melt period. Liu et al. (2011) concluded that the flow changes are strongly positive during April–May (17.7%–29.7%) for 2046–2065 using a lumped conceptual model (VHM). Our conclusion is comparable to Liu et al. (2011) and Fang et al. (2015a), but with larger width (−8.3% to 30.6% under RCP4.5 and −18.4% to 23.6% under RCP8.5 during the counterpart period). The reason may be related to the fact that multiple bias correction methods were used in our study while only one (a perturbation approach, equivalent to QM) was used in Liu et al. (2011). The decrease in summer runoff has been supported in many other regions, e.g., the Aksu River in south Tienshan Mountains (Duethmann et al. 2016), a forested Canadian watershed (Chen et al. 2011), the midlatitude alpine regions of the Swiss Alps (Addor et al. 2014) and the Colorado River Basin (Christensen & Lettenmaier 2007).

Future changes in high flows and low flows represented by 3DMHF and 7DMLF are shown in Figure 5. The median changes in the high flows for different exceedances are projected to range from −8.5% to 17.4%, while those in low flow will decrease by 12.4% to 46.7%, which may result in potential drought and hinder agriculture irrigation for the oasis in the lower reaches, especially under RCP8.5. We should be careful when interpreting changes in the low flows as only 32% of the simulations have absolute PBIAS values within 20% for the control period. For the high flow, the extremely high flood, e.g., exceedance <0.1, will not have a significantly increasing trend, while the relatively small peaks with exceedance between 0.8 and 1.0 will increase, which may help ecological recovery in the lower reaches of the Kaidu River. Many previous studies (Ragettli et al. 2016; Zhang et al. 2016) concluded that the extreme flow has been increasing or will increase in many mountainous regions, e.g., the Aksu River Basin amd the Langtang River. This may be related to these rivers having a considerable part of the runoff fed by glacier melt water, while the contribution of glacier melt to runoff in the Kaidu River basin is approximately 10%, which cannot generate a severe flood under a warmer climate.

Figure 6 shows the projected changes in SM, surface streamflow (R_s), subsurface streamflow (R_g) and evapotranspiration (ET) for 2080–2099 under RCP8.5 (changes in the hydrological components under RCP4.5 (also shown) are similar but smaller and not discussed here). The changes exhibit an obvious seasonality, i.e., insignificant from October to March and significant from April to September during which SM and rainfall occur. SM increases by 16% and 18% in March to May and 49% and 79% in June to August under RCP4.5 and RCP8.5, respectively, for 2080–2099. The contribution of SM to streamflow will decrease from 0.22 for the control period to 0.19 and 0.17 for 2080–2099 under RCP4.5 and RCP8.5, respectively, which means that SM is decreasing in importance. The snow melting time will shift forward approximately 1–2 months. This shifting may be partially attributed to the increased temperature, which governs snow melt, as demonstrated by other studies (Barnett et al. 2005; Moore et al. 2007). R_s will shift forward with more water generated during March to May and less water during June to August. Changes in the annual R_g are insignificant (4%–5%), which indicates the groundwater flow is the most stable component. ET will increase throughout the 21st century with a median increment of 24%–42%.

Table 4 lists uncertainty contributions of streamflow from GCMs, RCPs, BC_pcp and BC_tmp using the standard deviation method and the ANOVA approach. For both methods, GCMs is the most important uncertainty source in streamflow projection, which coincides with previous studies (Buytaert et al. 2010; Chen et al. 2011; Bosshard et al. 2013). Based on the standard deviation method, all contributions increase over these four periods slightly, e.g., contributions related to GCMs and RCPs increased from 0.215 and 0.093 during 2020–2049 to 0.345 and 0.124 during 2080–2099, respectively, indicating that the uncertainties from each source have increased (Wilby & Harris 2006; Exbrayat et al. 2014). Based on ANOVA, GCMs dominates the uncertainty with its contributions ranging from 0.907 to 0.967, while other sources can be ignored. The uncertainty proportion of GCM or RCP did not show an obvious increase with time based on ANOVA, which is different from Wilby & Harris (2006).

The uncertainty from GCMs based on the standard deviation method is the most important, which accounts for 55.3% of the total uncertainty, which is much lower than that based on the ANOVA approach (over 90% uncertainty caused by the GCMs). The reason may be that the ANOVA uses the square index to quantify the uncertainty contribution, which tends to favour high uncertainty sources.

The uncertainty result demonstrates the high contribution of climate models in uncertainty estimation of streamflow and suggests that the most effective way to reduce projection uncertainty is to reduce uncertainties in climatic predictions, as shown in Zhang et al. (2015).

As uncertainty in hydrological modelling is inevitable, it is important to determine the credibility and robustness of the projected streamflow change. Here, we used signal-to-noise ratio (S/N) (Zhou & Yu 2006; Addor et al. 2014) to represent this credibility. The median S/N for 2080–2099 is only 0.214 and 0.246 (less than 1) for RCP4.5 and RCP8.5 (Figure 7), respectively. Higher credibility was found in winter months (January, February, November and December) and SM season (April) under RCP8.5, indicating that the projection has higher credibility in these months compared to other months.

Furthermore, the consistency is estimated based on 504 simulations. Approximately 63.4% and 66.9% of the simulations demonstrate a negative streamflow change in 2080–2099 under RCP4.5 and RCP8.5 compared to 1986–2005. Most simulations (70.0% and 74.4%) show decreasing trends during June to February in the next year while only 30.9% and 29.4% of the simulations show a decreasing trend in April for these two scenarios. The S/N and consistency results both suggest that the streamflows are likely to increase in April and decrease in winter months, and for other months the changes are with high uncertainty, thus caution needs to be taken when used for decision-making.

In this study, we analysed the climate change impacts on hydrological processes for an important headwater of the Tarim River Basin with uncertainty analysis. A well-calibrated SWAT model forced by the downscaled and bias-corrected outputs of 21 GCMs was applied to investigate the effects of climate change on hydrological processes and the impact of GCM structural uncertainty on the hydrological processes.

While all the state-of-the-art GCMs predicted an increased temperature, the predicted precipitation has both decreasing and increasing trends of different magnitudes. Precipitation will increase by 3.1%–18% and 7.0%–22.5% while temperature will increase by 2.0 °C–3.3 °C and 4.2 °C–5.5 °C (represented by their 25% and 75% quantiles), respectively, for 2080–2099 under RCP4.5 and RCP8.5.

For the 21st century, streamflow is likely to increase until the 2060s and then decrease thereafter. Streamflow will change by −26% to 3.4% under RCP4.5 and by −38% to −7% under RCP8.5, respectively, for 2080–2099. Seasonally, streamflow will decrease by −27% and −15% for the summer months (June, July and August), while it will increase by 3.0% and 3.7% in spring (March, April and May) under RCP4.5 and RCP8.5, which may result in a potential water shortage during the critical water-demand summer. The seasonal shift of streamflow may be related to the spring freshet because SM will shift forward for approximately 1–2 months.

GCMs-related uncertainty was the most important based on the standard deviation method and ANOVA approach, while uncertainties linked to RCPs and bias corrections for precipitation and temperature are less important. The standard deviation method generated more mediocre results compared to the ANOVA approach.

Although the impacts of climate change on hydrological processes have been investigated in many previous studies, this study presents a complete study on future hydrological changes, highlighting the uncertainties caused by climate models. This study provides useful information on predicting uncertainty and credibility for water resource management and agricultural planning.

The research was supported by the ‘Thousand Youth Talents’ Plan (Xinjiang Project Y371051), the National Natural Science Foundation of China (41630859) and the CAS ‘Light of West China’ Program (2016-QNXZ-B-12).