A comparative assessment of CMIP5 and CMIP6 in hydrological responses of the Yellow River Basin, China

Global warming has been recognized as the primary determining factor to augment climate change-related risks worldwide (Su et al. 2021). The Intergovernmental Panel on Climate Change (IPCC) in 2021 has reported that the global average surface temperature during the most recent decade (2011–2020) has exceeded the warmest centennial-scale range reconstructed for the present interglacial period (Masson-Delmotte et al. 2021). Increasing global mean surface temperature has caused higher evapotranspiration rates and leads to changes in precipitation. These changes are then expected to change water resources in the timing and amount (Cherkauer et al. 2021). Some investigations indicate that the renewable surface and groundwater resources in many regions are expected to decrease significantly in the 21st century (Chenoweth et al. 2011; Portmann et al. 2013; Llopart et al. 2020; Andrade et al. 2021). Policies should be made to deal with these changes. In order to design effective strategies for planning and management of water resources, future streamflow under a changing climate should be projected and understood, especially at the basin scale (Zhang et al. 2015), which is the focus of the present study.

Numerous studies have investigated the effects of climate change on different hydrological systems in upcoming time periods using general circulation models (GCMs) (Kim et al. 2013; Guo et al. 2018; Nilawar & Waikar 2019; Guo et al. 2020a, 2020b). However, raw GCM outputs are too coarse to be compatible with hydrological models. It is necessary to convert GCM outputs into local meteorological variables required for reliable hydrological modeling, referred to as downscaling (Musau et al. 2013; Bozkurt et al. 2018; Mishra et al. 2020). The Coupled Model Intercomparison Project (CMIP), which is organized by the World Climate Research Program (WCRP), has contributed to producing immense GCM outputs (Meehl et al. 1997, 2000). The GCM outputs from different phases of the CMIP have been at the heart of climate change studies. However, previous studies on hydrological projection were mainly based on the outputs from CMIP3 and CMIP5 under different designed emission scenarios, such as the Special Report on Emission Scenarios (SRESs) and Representative Concentration Pathways (RCPs) (Schnorbus & Cannon 2014; Ayers et al. 2016; Ficklin et al. 2016; Meaurio et al. 2017; Joshi et al. 2020). Although the CMIP5 generation is regarded to be superior to the CMIP3, some problematic features remained; for example, climate models commonly underestimated the precipitation intensity. The WCRP has announced the latest phase, referred to as CMIP6, and the CMIP6 models were reported to have a new and better representation of physical, chemical, and biological processes, as well as higher spatial resolution than CMIP5 (Stouffer et al. 2017; Zhu & Yang 2020; Li et al. 2021).

CMIP6 advocates emission scenarios for future climate projection based on the combination of RCPs and Shared Socioeconomic Pathways (SSPs). The scenarios in CMIP6 cover a broader range of emission features than considered in CMIP5, with high and very high greenhouse gas (GHG) emissions (SSP370 and SSP585), intermediate GHG emissions (SSP245), and very low and low GHG emissions and CO₂ emissions (SSP119 and SSP126) (O'Neill et al. 2016). In this phase, the combination of RCPs and SSPs has a clearer description of future society's socioeconomic evolution, leading to more reasonable future scenarios (‘The CMIP6 landscape’ 2019). Currently, a few studies have been reported in the literature with the latest simulations and outputs of CMIP6. Su et al. (2021) evaluated and compared the simulation of drought characteristics by four CMIP6 models and their corresponding CMIP5 predecessors in China. Akinsanola et al. (2021) assessed the CMIP6 models in simulating extreme precipitation statistics over Eastern Africa. Zhu et al. (2021) presented projections of climate extremes over China based on CMIP6 and reported that CMIP6 models outperformed CMIP5 over China, especially in simulating precipitation extremes. These studies generally used an ensemble of various GCMs from different groups around the world, allowing us to analyze discrepancies in the change of climate variables to evaluate how models diverge in projecting long-term impacts of climate change (Bağçaci et al. 2021). However, the above studies mainly focused on evaluating the performance of CMIP6 models in simulating the climate. Little efforts have been reported for future projections of hydrological regimes under different warming targets using updated CMIP6 models. Moreover, whether or not GCMs have improved from CMIP5 to CMIP6 remains unclear. Therefore, it is of great interest to systematically assess the performance of multi-model CMIP6 GCMs for the hydrological responses against observations and compare it with the previous generation of GCMs.

Zhang et al. (2015) indicated that the extent to which streamflow responds to human activities and climate changes on streamflow was different for different river basins. Modeling and predicting streamflow in a changing environment is critical for basin-scale water resource management. The Yellow River Basin (YRB) is the primary water supply and the hydropower generation source in Northern China. Investigating the role of multiple GCM ensembles in obtaining comprehensive knowledge of future streamflow at a regional scale across the YRB is still of substantial importance. Policymakers need to formulate a more reasonable water management policy for water resource distribution and reservoir regulation. On the other hand, no studies have investigated the hydrological performance difference between CMIP5 and CMIP6 models over the YRB in detail.

Therefore, this study aims to reveal the capability of available CMIP6 GCMs in simulating the hydrology over the YRB and compare their performance with that of CMIP5 GCMs. The specific questions to be addressed are: (1) How is the performance of the CMIP6 models in simulating hydrological variables compared to CMIP5? (2) What causes the differences in simulation skills among different models? In this study, outputs from CMIP6 and CMIP5 models will be downscaled first to ensure a higher spatial resolution, followed by evaluation against observations. Downscaled CMIP6 and CMIP5 outputs are then employed to drive a Soil and Water Assessment Tool (SWAT) hydrological model and project future streamflow under diverse scenarios. The hydrological responses are then evaluated and compared between CMIP6 and CMIP5.

The Yellow River, with a length of 5,464 km and a basin area of 0.75 million km², is the second-largest river in China. It is also the fifth-longest river in the world. Taking Hekou town and Taohuayu as cut-off points, respectively, the mainstream of the Yellow River is divided into three reaches: namely the upper, middle, and lower reaches (Figure 1). The YRB is the major source of water supply in Northern China, and water supply includes municipal use, industrial use, irrigation, and ecological use (Chang et al. 2014). Meanwhile, especially for the Upper Yellow River, one of the China's 13 major hydropower bases, more than 25 reservoirs have been built or planned in this area (Guo et al. 2021). Thus, water supply and hydropower generation are the two main benefit objectives for water utilization in the YRB. Policymakers still need to formulate a more reasonable water management policy for future water resource distribution and reservoir regulation in the YRB. However, during the past several decades, the hydrological regime in the YRB has significantly changed due to climate change and anthropogenic effects. Changes in the hydrological regime include an increase in temperature and a decrease in streamflow and precipitation. Multiple GCM ensembles can give comprehensive knowledge of future hydrological variations at a regional scale across the YRB. Attempts have been made to understand the causes of the hydrological changes in this study.

This study uses observed daily meteorological data from 1961 to 2010 at 68 meteorological stations in the YRB, including precipitation, temperature, wind speed, and relative humidity. The data sets are obtained from the National Meteorological Information Center of China Meteorological Administration. We analyze climate simulations from models participating in CMIP6 and CMIP5. The CMIP6 and CMIP5 model outputs, including the monthly precipitation and average, maximum, and minimum temperature, are available from the Earth System Grid data distribution portal (http://www.earthsystemgrid.org). The projection experiment in CMIP6 contains a new set of emission and land-use scenarios that combine SSPs and RCPs. We analyze 21 CMIP6 models (see Table 1), for which monthly model outputs for three SSP scenarios (SSP126, SSP245, and 585) are available. SSP126 is a low ending range of future scenario with an achieving forcing a level of 2.6 W/m² by 2100, SSP245 represents an intermediate ‘middle of the road’ scenario with an achieving forcing level of 4.5 W/m² by 2100, and SSP585 denotes a high emission scenario with a relevant forcing outcome of 8.5 W/m² by 2100 (O'Neill et al. 2016). We also analyze 18 CMIP5 models (see Table 2), for which the three RCP emission scenarios (RCP26, RCP45, and RCP85) are considered here. The radiative forcing trajectories in the SSPs and RCPs can reflect various possible combinations of economic, technological, demographic, and policy developments (Masson-Delmotte et al. 2021). SSP126, SSP245, and SSP585 scenarios of the CMIP6 update the corresponding RCP26, RCP45, and RCP85 scenarios of the CMIP5, respectively. The GCM outputs are gridded on a 2.5° × 2.5° resolution using the inverse distance weighted (IDW) interpolation. Three 30-year time slices are used in this study, 1971–2000 for the historical climate, 2026–2055, and 2066–2095 for the future climate..

Four hydrological stations, including Tangnaihai (TNH), Lanzhou (LZ), Sanmenxia (SMX), and Huayuankou (HYK) stations (Figure 1), provide the original monthly streamflow data in this study. The original streamflow series are naturalized, which means that the effects of human activities have been removed. The monthly streamflow data from 1971 to 2010 at the TNH station and from 1961 to 2010 at the other three stations are available. The streamflow data are used to calibrate and validate the hydrological model. The historical records for the four stations are split into calibration and validation periods. For the TNH station, 1971–1998 is used for calibration and 1999–2010 for validation. For the other three stations, 1963–1994 is used for calibration and 1995–2010 for validation.

Geospatial data include digital elevation model (DEM), land use, and soil map information. The DEM data with a horizontal grid spacing of 30 m resolution is provided by the Geospatial Data Cloud of China (http://www.gscloud.cn). It is used for defining the stream, area, and boundary of sub-basins. Land-use map data in 1990 were developed by the Chinese Academy of Sciences (http://www.resdc.cn) with a spatial resolution of 1,000 m. The soil map information was derived from Harmonized World Soil Database 1.2 (https://iiasa.ac.at).

We first used the BCSD method to correct and downscale the GCM projections of precipitation and temperature from 1971 to 2000 (baseline period). RMSE and CC were used to evaluate the downscaled model performance against the observations. The evaluation indexes using the Taylor diagram are shown in Figure 2. Among the models in raw CMIP6 compared to raw CMIP5, inter-model differences still exist. For raw CMIP5 and CMIP6, the RMSE values of the multi-model ensemble mean precipitation are 32.14 and 23.77 mm, respectively. The raw CMIP6 outputs slightly outperform raw CMIP5 in simulating precipitation in terms of accuracy statistics. This is a great agreement with Bağçaci et al. (2021) that CMIP6 can improve the amplitude of precipitation and alleviate errors departing from observations. However, CMIP6 in simulating temperature with an RMSE value of 4.91 °C displays worse performance in comparison to CMIP5 with an RMSE value of 2.24 °C.

Figure 2 shows that there is no significant difference in RMSE and CC values between CMIP5 and CMIP6 after correction. After correction and downscaling, the RMSE and CC of monthly precipitation are around 20.85 mm and 0.82, 20.02 mm, and 0.83 for CMIP5 and CMIP6, respectively. It is clear that BCSD downscaling can significantly correct the GCM precipitation. The monthly mean precipitation of BCSD projections has higher CC and lower RMSE than that of raw predictions. For CMIP5 and CMIP6, the RMSE of precipitation is reduced approximately by 40.27 and 33.29%, while the CC value is increased by about 10.48 and 16.86%. The improvement in temperature is smaller than that of precipitation. This is because the monthly mean temperature has a higher correlation coefficient (around 0.99) than precipitation. The RMSE value of temperature is reduced approximately by 0.61 and 0.56 °C, while the CC value is increased by about 0.519 and 0.515% for CMIP5 and CMIP6, respectively. The results show that both CMIP5 and CMIP6 can simulate temperature better than precipitation. Overall, the statistical downscaling method of BCSD is good at simulating the historical period of precipitation and temperature in the YRB.

We used the SWAT model to project monthly streamflow under GCM-derived and downscaled climate scenarios. The four hydrological stations were first selected to calibrate the SWAT model, i.e., TNH, LZ, SMX, and HYK stations. The streamflow data we used is the human-reconstructed natural streamflow. Model calibration and validation were conducted by comparing the SWAT simulated data with the original streamflow on a monthly basis. Figure 3 compares the simulated monthly streamflow with the observed streamflow values. The model performance values of each hydrological station are shown in Table 3. The NSE and R² values are above 0.75, and the PBIAS value is less than 12% of the four stations during the calibration and validation periods. For TNH and LZ stations, except for several years during which simulated peaks are underestimated, most periods have a good agreement between the simulated and observed streamflow, and the low flow is simulated very well. However, the SWAT model does not simulate the low flow very well for SMX and HYK stations. Compared with the streamflow collected at the upper hydrological stations (TNH and LZ), the streamflow collected at the lower stations (SMX and HYK) is more likely affected by human activities (e.g., reservoir operations and water supply). We assume that the human-reconstructed streamflow at SMX and HYK stations is lower than the natural streamflow since water with drawls could not be completely considered during reconstruction. This may be the reason why the worse performance is obtained in the lower basin. Every effort is made to achieve the best possible calibration and validation results for each hydrological station. The simulation results we obtained in our study outperform those reported in similar studies. For example, Yang et al. (2018) indicated a simulation performance of the HYK station with the NSE values of 0.236 and 0.394 during the calibration and validation periods, respectively. Overall, the above results indicate that the SWAT model is good at simulating the monthly hydrological procession of the YRB.

We then used the modified BCSD method to correct and downscale the GCM projections of precipitation and temperature during the period of 2026–2055 (near future) and 2066–2095 (far future). The period of 1971–2000 is referred to as the baseline period. The mean annual temperature and precipitation in 1971–2000 are 7.32 °C and 436.59 mm. Figures 4 and 5 demonstrate the spatial distribution of multi-model ensemble average changes in annual temperature and precipitation. It is noticeable that the yearly temperature and precipitation changes show an uneven spatial changing trend projected by both CMIP5 and CMIP6 models. Temperature changes in the middle and upper YRB are more extensive than those in the lower YRB, while precipitation changes in the lower and upper YRB are more significant than in the middle YRB. Similar spatial patterns of temperature changes under RCP85 (SSP585) are observed when compared to those under RCP26 (SSP126) and RCP45 (SSP245), but with a larger increasing magnitude of temperature under RCP85 (SSP585) than that under RCP26 (SSP126) and RCP45 (SSP245). An increase of 2–3 and 4–6 °C in temperature can be identified under RCP85 and SSP585 in the near and far future, respectively. Besides, a larger increasing magnitude of precipitation can be found under RCP85 (SSP585) when compared to that under RCP26 (SSP126) and RCP45 (SSP245), and most regions of the YRB under RCP85 and SSP585 are dominated by the increasing magnitude of precipitation of 5–10 and 10–30% in the near and far futures, respectively. In general, the comparison of CMIP6 SSP-based simulations with CMIP5 using the RCPs shows that the increase in simulated warming in CMIP6 versus CMIP5 arises because higher climate sensitivity is more prevalent in the CMIP6 model version.

Figure 6 shows the mean annual temperature and precipitation in both the near and far futures. In the future, the YRB will experience a warmer and wetter climate in most cases. It can be seen from Figure 6 that increasing temperature is apparent under CMIP5 (RCP26, RCP45, RCP85) and CMIP6 (SSP126, SSP245, SSP585), while CMIP6 projects a warmer climate than CMIP5. For the near future, CMIP5 and CMIP6 project an increased mean temperature by 1.64–2.20 and 1.92–2.39 °C, respectively. In the far future, these increments will reach 1.68–4.76 and 2.31–5.29 °C, respectively. The above results imply that higher emissions of GHG can cause larger temperature impacts. In addition, for both CMIP5 and CMIP6, moderate warming (<2.5 °C) is projected in the near future, and a much warmer climate is projected in the far future. Unlike the consistent increasing trend in temperature projections, CMIP5 and CMIP6 display considerable variability in precipitation projections. CMIP5 and CMIP6 changes in precipitation relative to 1971–2000 lie in the range of −4.77 to 13.23 and −4.37 to 23.14% for the near future, while that ranges from −6.90 to 21.73 and −2.70 to 37.33%, respectively, for the far future. The emblematic annual mean precipitation, when averaged over the whole of the YRB in CMIP6, increases by 3.45–4.70 and 6.77–15.40% for the future period of 2026–2055 and 2066–2095, while the counterpart in CMIP5 is 2.58–2.96 and 3.83–9.95%, respectively.

Figure 7 shows the annual mean temperature and precipitation trends in both the near and far futures. The different color shadows represent all GCM projections, while the solid line denotes the model ensemble average. It indicates that the multi-model ensemble range in annual precipitation shows more significant uncertainty than that in annual temperature. During the near future period of 2026–2055, the average basin temperature shows an increasing trend, and the increasing rate varies from 0.183 to 0.511 °C /10a and 0.2 to 0.526 °C/10a under CMIP5 and CMIP6, respectively. However, there is an interesting finding that the average basin temperature shows a downward trend under RCP26 and SSP126 during the far future period of 2066–2095. This is because both RCP26 and SSP126 represent scenarios with lower GHG emissions and are designed to limit climate warming (Masson-Delmotte et al. 2021). At the opposite end of the range, RCP85 and SSP585 represent the very high warming end of future emission pathways and thus obtain a much higher temperature among all scenarios. The annual precipitation shows an apparent decreasing trend in the baseline period. However, the annual precipitation starts to increase in most cases in future periods. The increasing rate is 4.365–12.813 mm/10a and 6.522–12.77 mm/10a under CMIP5 and CMIP6, respectively, during the near future, and the changing trend rate is −1.107 to 11.97 mm/10a and 1.022 to 19.786 mm/10a during the far future. There is only one exception: the annual precipitation shows a slight downward trend under RCP26 from 2066 to 2095.

Figure 8 shows the monthly precipitation and temperature projected by the multi-model under CMIP5 and CMIP6. It can be noticed that the YRB has four distinct seasons, both in the baseline and future periods. The range of multi-model temperature changes is similar in all 12 months. An increase in monthly temperature under CMIP5 ranges from 0.49 to 2.92, 0.96 to 2.87, and 1.17 to 3.50 °C in the near future, while that ranges from 0 to 3.24, 1.03 to 4.44, and 3.00 to 6.69 °C in the far future. The differences in monthly temperature among RCPs or SSPs are substantial, specifically in the far future. Similar patterns but a larger range of monthly temperature changes in the far future under CMIP6 compared to that under CMIP5 can be found. In contrast, there are quite different characteristics for the changes in monthly precipitation in the future. For example, in the near future, October and November show a decreasing trend, while May and July show the most remarkable tendency of increase. The monthly precipitation projected by the multi-model from May to October has a larger uncertainty than in other months under CMIP5 and CMIP6.

For the future period, the calibrated SWAT model simulated the monthly streamflow corresponding to different climate scenarios. According to the analysis made in previous sections, 234 scenarios (18 × 3 × 2 + 21 × 3 × 2) are identified in this study. Figure 9 shows the projected average monthly streamflow for different scenarios at the four hydrological stations, i.e., TNH, LZ, SMX, and HYK stations. Different scenarios produce a wide range of changes in the hydrological regime. The model uncertainty that becomes apparent by comparing different models is larger than the simulated streamflow of individual models. Of the 18 CMIP5 models used in this study, the GFDL-CM3 model provides the greatest estimate of the impacts of climate change on streamflow in the headwater catchment of the YRB, while MIROC5 generates the largest effects in the lower YRB. The scenario-based climatic changes have remarkable impacts on streamflow changes in the future. Compared with the simulated streamflow in the upper YRB (TNH and LZ stations) in the baseline period (Figure 9), the projected streamflow changes of the multi-model ensemble mean for the near future with the values of −0.94 and 0.40% are found, while that for the far future increases by 4.22 and 5.99%. As for the other two hydrological stations in the lower YRB (SMX and HYK stations), the multi-model ensemble mean streamflow for the near future projects a recession of 14.93 and 13.42%, and that for far future decreases with the declination of 9.67 and 7.87%. The decreased streamflow is well corroborated by results obtained by Zhu et al. (2016) and Yang et al. (2018).

Of the 21 CMIP6 models used in this study, the CanESM5–CanOE impacts streamflow significantly in the upper and lower YRB. In the upper YRB, the projected multi-model ensemble mean streamflow of TNH and LZ stations decreases for the near future with the declination of 3.43 and 4.75%. Several researchers also pointed out similar results when they evaluated streamflow changes under climate changes in the lower YRB and indicated that no evident increase in precipitation but a significant increase in temperature was the major cause behind decreased streamflow (Xu et al. 2009; Zhang et al. 2017). However, the projected multi-model ensemble mean streamflow increases with the increment of 5.51 and 3.28% for the far future. This may be caused by the substantial increase in temperature and the larger increasing magnitude of precipitation. In terms of lower YRB, the multi-model ensemble mean streamflow of SMX and HYK stations for two benchmark periods projects a declination of 13.07 and 6.52%, and 11.26 and 3.87%. The estimated streamflow reduction in the far future is not as extreme as in the near future, and the streamflow of CMIP6 is generally greater than that of CMIP5.

Figures 10 and 11 show the annual and monthly streamflow during different future periods compared with those corresponding to the historical conditions. There is no significant difference between CMIP5 and CMIP6. The three RCPs and three SSPs project a similar streamflow that the annual streamflow shows an increasing trend under various scenarios. The relationship between the streamflow and the magnitude of green gas emissions is nonlinear. That is to say, the streamflow response to precipitation and temperature changes is nonlinear when precipitation and temperature either increase or decrease. The seasonal shape of the hydrograph is similar to the baseline, but both tendencies of increasing and decreasing will occur under different scenarios. The streamflow shows a remarkable decreasing trend in spring while slightly decreasing in other seasons. The estimated streamflow reduction at a monthly scale in the far future is not as extreme as in the near future, especially at SMX and HYK stations. For example, in the near future, SMX station experiences a decreasing tendency with percentages of 16.97 and 14.90% under CMIP5 and CMIP6 scenarios, whereas, in the far future, it shows a decreasing trend with the percentages of 11.24 and 8.03%. The major contributors that led to lower variance included increases in streamflow from August to December. This is because the increased precipitation driven by climate change will generally increase streamflow during the winter and spring seasons.

In this study, the uncertainty we considered originates from two sources, i.e., GCMs and RCPs, or GCMs and SSPs. The uncertainty caused by GCMs is dominant for different hydrological stations, as shown in Figure 12. Compared to the near future, the effect of RCPs and SSPs uncertainty in the far future generally shows a slight decrease. GCMs uncertainty in the near future contributes more than 70% to the total uncertainty in annual mean streamflow, whereas in the far future it contributes more than 80%. This finding that GCMs uncertainty is the dominant source is in agreement with the conclusions in some previous studies (Gao et al. 2020; Lee et al. 2021). These studies indicate that an ensemble of various GCMs from different groups around the world can generally provide better water resource assessments than a single GCM since, in some cases, the uncertainty of climate models is greater than that of hydrological simulations. The lower contribution of RCPs and SSPs individual complies with the results that three RCPs or three SSPs project a similar streamflow. This may be due to the ensemble of multiple simulations, which can eliminate the significant differences among the three RCPs or SSPs. SSPs of CMIP6 are projected to have a lower contribution in contrast with that of CMIP5. The reason may be the larger number of GCM models used in CMIP6. More combinations of GCMs and emission scenarios may give better results. However, it can be concluded that inter-model uncertainties have a comparable magnitude between CMIP6 and CMIP5.

Scenarios show that the streamflow responses to precipitation and temperature changes are nonlinear when precipitation and temperature increase or decrease. To further investigate the relationship between precipitation, temperature, and streamflow, we have quantified the contribution of precipitation and temperature changes impacting streamflow at the mean annual scale. A typical result of many studies (Sillmann et al. 2013; Sonkoué et al. 2019; Akinsanola et al. 2021) agreed that the multi-model ensemble means tend to represent precipitation characteristics better than individual models. To eliminate the inter-model uncertainties under different RCPs or SSPs, we use the multi-model ensemble means to drive the hydrological modeling. In general, the fractional contribution of precipitation to streamflow changes is larger than that of temperature to streamflow changes (Zhang et al. 2017). As observed from Figure 13(a), the relationships between annual streamflow, annual precipitation, and annual mean temperature in the near future show that streamflow has closer relationships with precipitation in comparison with temperature. The fractional contribution rates of precipitation to streamflow changes are more than 60% under most CMIP5 and CMIP6 scenarios. This means that streamflow is more sensitive to precipitation change. It can be depicted from Figure 13(a) that the larger contribution of temperature change in the upper YRB than that in the lower YRB is most likely due to snowmelt.

However, we find that the change of temperature contributes more than that of precipitation in the far future (Figure 13(b)). That may be because the relative change in temperature between the far future and the baseline period is much more significant than the relative change in precipitation, particularly for scenarios with high green gas emissions, i.e., RCP85 and SSP585. The relative change in temperature is 65.26 and 72.43%, while that in precipitation was 9.95 and 15.40% under RCP85 and SSP585, respectively. This suggests that an increase in temperature would likely accelerate evaporation, which will exceed the impacts of precipitation and subsequently reduce streamflow (Li & Fang 2021). This agrees with Zhang et al. (2017), who reported that the decreased streamflow can be found in the headwater regions of YRB during the 2070–2099 period, which could be attributed to a more significant increase in temperature than precipitation in these regions.

Beyond GCMs and RCPs uncertainty, several other uncertainty sources were reported, like those stemming from hydrological models, hydrological parameters, and downscaling methods. For example, Lee et al. (2021) found that hydrological parameters might lead to high uncertainty in assessing climate change impacts, while Meaurio et al. (2017) reported that different downscaling methods might have a considerable contribution to different extreme high flows. Recent years have witnessed changes in streamflow processes due to increasing human activities, such as agricultural activities and the construction of dams and water reservoirs. These could introduce more uncertainty. Guo et al. (2020a) indicated that changes in underlying surface properties would impact the fluvial hydrological cycle and, hence, impact streamflow changes. Räsänen et al. (2016) pointed out that the streamflow of the Lancang-Mekong River in the dry season was more significant in the pre-dam period than in the post-dam period. As for the YRB, 25 reservoirs have been built or planned in the upper YRB with 16.3483 GW of total installed capacity, which made the YRB one of the 13 major hydropower bases in China (Si et al. 2019). The irrigation area of the YRB increased from 0.8 million hm² in 1950 to 7.33 million hm² at the end of 2000 (Wang et al. 2020). The construction of cascade dams and agricultural irrigation would result in a significant change in streamflow in the YRB. Many attempts have been made to investigate factors potentially driving changes in different streamflow components (Li et al. 2017).

Unlike climate change projected using scenario-based GCMs, human impacts and changes of underlying surface properties are hard to be evaluate and predict (Zhang et al. 2017). In addition to precipitation and temperature, agricultural irrigation and reservoir operations should be considered when analyzing streamflow changes at different hydrological stations. In this study, instead of assuming climate change and human activities scenarios, we further used generalized additive models for location, scale, and shape (GAMLSS) models to quantify how streamflow changes respond to these four factors according to the historical records from 1978 to 2010. A detailed description of GAMLSS models can be found in the Supplementary Material. In our study, the predicted streamflow series was the output variable, and the annual precipitation, temperature, food output, and hydropower generation were considered as the input variables. Streamflow defined by different quantiles was modeled using the GAMLSS with Gamma distribution function, the same as Li et al. (2017). Figure 14 shows the historical standardized precipitation, temperature, food output, and hydropower generation indices. As shown in Figure 1, there were no large reservoirs in the drainage areas controlled by the TNH station, and hence, the effects of reservoir-induced impoundment were neglected but were considered for LZ, SMX, and HYK stations. Specifically, for the LZ station, we evaluated the effects of two upstream reservoirs with high regulating capacities, i.e., Longyangxia (LYX) and Liujiaxia (LJX) reservoirs, on driving changes in streamflow components. Annual mean temperature, food production, and hydropower generation series showed an upward trend at all stations. In contrast, the annual mean precipitation series exhibited a large degree of variability, much more than the other three indices.

The coefficients of α_pr, α_tem, α_Food, and α_Power, respectively, reflected the impacts of precipitation, temperature, agricultural irrigation, and reservoir operations on streamflow changes, as shown in Figure 15. The coefficients of α_LYX and α_LJX reflected the effects of LYX and LJX reservoirs on streamflow of the LZ station, respectively. Positive and negative values indicated the potential to increase and decrease streamflow, respectively. Light (dark) color-filled markers denote significant parameters at a 5% (10%) confidence level, and hollow markers denote not significant parameters at a 10% confidence level. In general, precipitation tended to increase streamflow, and temperature tended to decrease streamflow. In most cases, statistical analysis of α_pr indicated that precipitation changes had significant positive influences on streamflow variations at the 10% significance level. However, it was in line with Li et al. (2017) that α_pr at the lower percentiles of the discharge distribution at the LZ station was negative but not significant. The α_tem coefficient values in most cases were smaller than 0, showing that temperature contributed to the decrease of streamflow changes. In the aspect of human activity factors, the α_food coefficient could reflect the prominent effects of water withdrawal on streamflow variations in terms of magnitude and variability. The α_food coefficient values at the LZ, SMX, and HYK stations were smaller than 0 at the 5/10% significance level, implying that agricultural irrigation had the potential to decrease streamflow. These results agreed with those of Singh et al. (2016), Traylor & Zlotnik (2016), and Li et al. (2017), showing similar effects of water withdrawal for irrigation on streamflow changes in other basins.

To obtain power generation benefits, attempts should be made to maintain high reservoir water levels and to reduce the water released from a reservoir. That is to say, it is generally believed that hydropower generation did not consume water. Correspondingly, the α_power coefficient values were higher than 0 at SMX and HYK stations. Specifically, for the SMX station, the reservoir release was reduced during the operation of the SMX reservoir, and thus, the α_power coefficient values were significant at the 5/10% significance level. Reservoirs like LYX and LJX could keep water for downstream areas to cope with possible droughts in the future, by which the interests of hydropower generation and downstream water use seemed synergic in the long run (Si et al. 2019). The α_LYX coefficient values were almost negative at the 5/10% significance level related to different streamflow components, while the α_LJX coefficient values were positive at the 10% significance level. This implied that the LYX reservoir had the potential to decrease streamflow, but the LJX reservoir could increase streamflow. Compared with the evident effects of precipitation at the TNH station, that at the other three stations would be reduced to the effects of reservoir-induced impoundment and irrigation-induced withdrawal. Overall, precipitation played a critical role in changes in streamflow components, and human activities, i.e., agricultural irrigation and reservoir operations, tend to have more substantial impacts on hydrological processes across the YRB.

In this study, we analyzed the performance of CMIP5 and CMIP6 models in projecting hydrological responses of the YRB in the near future period of 2026–2055 and the far future period of 2066–2095. The calibrated SWAT model was used to simulate monthly streamflow under BCSD downscaled climate scenarios. GCMs uncertainty and RCPs or SSPs uncertainty were quantified, while the sensibility of streamflow changes to temperature and precipitation changes was also identified. The results of this study were of both theoretical and practical merits in terms of the management of water resources under the influences of changing climate in the YRB. The main results are summarized as follows:

• (1)

  The raw CMIP6 products slightly outperformed raw CMIP5 in terms of accuracy statistics for simulating precipitation but displayed worse performance in temperature in the study area. The statistical downscaling method of BCSD was good at simulating precipitation and temperature in the YRB. There was no significant difference in RMSE and CC values between CMIP5 and CMIP6 after BCSD correction.

• (2)

  Even though both corrected CMIP5 and CMIP6 indicated an increasing trend in the temperature in the YRB, CMIP6 resulted in a relatively higher increase than CMIP5. For the near future, CMIP5 and CMIP6 projected an increased mean temperature of 1.64–2.20 and 1.92–2.39 °C, respectively. In the far future, these increments would reach 1.68–4.76 and 2.31–5.29 °C, respectively. An interesting finding was that the average basin temperature showed a downward trend in the far future under the scenarios with low GHG emissions (RCP26 and SSP126). CMIP5 and CMIP6 displayed considerable variability in precipitation projections due to the well-known fact that GCMs were not very reliable in simulating precipitation. Changes in the precipitation of CMIP5 and CMIP6 relative to 1971–2000 lay in the range of −4.77 to 13.23% and −4.37 to 23.14% for the near future, while they ranged from −6.90 to 21.73% and −2.70 to 37.33%, respectively, for the far future.

• (3)

  The three RCPs and three SSPs projected a similar streamflow, although the projected streamflow was different from the historical values. In most cases, future streamflow would decrease, and the estimated streamflow reduction in the far future was not as extreme as that in the near future, and the streamflow of CMIP6 was generally greater than that of CMIP5. Uncertainty in the projected streamflow was dominated by GCMs uncertainty for both CMIP6 and CMIP5. However, it was concluded that inter-model uncertainties had a comparable magnitude between CMIP6 and CMIP5. The streamflow was more sensitive to precipitation changes than temperature changes in the near future, while it was more sensitive to temperature changes in the far future.

Agricultural irrigation and reservoir operations tend to have more substantial impacts on hydrological processes than precipitation and temperature. The results of this study are of theoretical and practical merits in terms of management of water resources under influences of changing climate. However, we claim a number of limitations within this study that reduce our overall certainty in this projection. One limitation not considered is the use of the single downscaling method and the hydrological model. The predictive uncertainties of streamflow projection are further amplified when it comes to different hydrological model structure and parameterization, and downscaling techniques. Another limitation is that we treated the land use/cover as fixed in this study to focus on the climatic changing factors. However, conclusions based only on one land-use/cover style may not be optimal and may change with the inclusion of other land-use/cover predictors. Further investigations are needed to comprehensively evaluate the influence of climate change and land-use/cover change on future streamflow changes. We also assumed that GCMs of CMIP5 and CMIP6 were equal in our balanced streamflow prediction. However, in future research, additional information would be welcomed to evaluate the GCMs credibility, which potentially impacts the streamflow projection and improves the certainty of the forecasting results.

This research is funded by the National Key Research and Development Program of China (2021YFC3201100) and the National Natural Science Foundation of China (52009121, 52109037). The authors express their gratitude for the support from the naturalized streamflow data set provided by the Yellow River Conservancy Commission of The Ministry of Water Resources. We thank the anonymous reviewers and editors for their comments, which greatly helped improve the quality of this paper.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.