Evaluation Climate Change Impacts on Water Resources Over the Upper Reach of the Yellow River Basin

In this study, a climate-streamflow modeling framework (CSF) is advanced to generate future climate projections and assess climate change impacts on water. The proposed CSF incorporates global climate models (GCMs), meteorological factors downscaled by the providing regional climate impacts for studies (PRECIS), and stepwise-clustered hydrological model within a general framework. It has advantages in (1) transferring large scale climate variables from global climate models to high-resolution meteorological datasets by the PRECIS, and (2) quantifying the climate change impacts on streamflow simulation by employing the stepwise cluster analysis method to reflect nonlinear relationships between predictand and predictor. Correspondingly, a real case of streamflow simulation at the upper Yellow River basin is applied to demonstrating the efficiency of the CSF. Results disclose that: (i) an increasing trend of average temperature exists in future with the highest temperature increments happening in December under RCP4.5 and more increments occurring in Summer under RCP8.5; (ii) there would be no visible precipitation changes in future Winter when compared with historical precipitation, while remarkable rainfall reduction may occur in May and June; (iii) compared with historical streamflow rates, the future streamflow would mainly change during May to October in which remarkable streamflow reduction may be observed in May but explicit increases may occur in July.


Introduction
Climate change impacts on hydrology and water resources can imply shifts in both alteration (shifts in timing and averages) and intensification (increasing number and severity of extreme events) in the hydrological cycle (Greve and Seneviratne 2015;Tian et al. 2021;Hu et al. 2022). Streamflow, as the integrated component of hydrology in a basin, is one of the major factors remarkably affected by climate change. Moreover, extensive uncertainties exist in streamflow predictions under climate change, which are embedded in hydrologic and meteorological inputs from climate projections, river morphology and geographical features (Raseman et al. 2020;Koukoula et al. 2021;Fan and Huang, 2021). Consequently, projections of change in local water resources are wrought with uncertainties surrounding natural variability, and hydrological method (Nguyen et al. 2020).
Over the past decades, various meteorological factors, such as observation datasets and downscaled outputs from global climate models (GCM) or regional climate models (RCM), have been employed to assess the potential impacts of climate change and variability on the hydrological regime (Gao and Xie 2014;Lazin et al. 2020). For instance, Graham et al. (2007) indicated that, in the analysis of the climate change impacts on future runoff, the most important uncertainty source comes from GCM forcing, which has a larger impact on projected hydrological change than the selected emission scenario or RCM used for downscaling. Faramarzi et al. (2013) analyzed the impact of climate change on water resources, reflected the seasonal patterns of blue water. Wang and Liu (2019) discussed the effect of floods and river sedimentation on the Yellow River and the training strategy for preventing flood disasters in the past 2000 years. In general, despite significant advances in climate change science and modelling techniques, the uncertainty associated with such projections (rather than predictions) is likely to continue for the foreseeable future (Blásquez and Nuñez 2013;Tsakiris et al. 2007;Li and Xie 2022).
For evaluation of climate change impacts water resources, multiscale and complete meteorological datasets (such as rainfall, streamflow, and temperature) should be required and taken into consideration, however, records of hydrologic processes are usually short and often have missing observations (Nikolopoulos et al. 2011). In fact, hydrological systems are inherently spatial, nonlinear and time-variant (e.g. both the surface runoff processes and the rainfall-runoff relations are nonlinear) (IGR 1980). The nonlinearity and dynamic behavior of non-periodic hydrologic datasets have been indicated in water resources literature as issues that influence the performance of modeling tools that ignore nonlinearity and dynamics inherent in the data structure (Friedli et al. 2021;Ansell and Dalla-Valle 2022). Therefore, it is desirable to develop more flexible hydrological simulation approaches that can be used in modeling hydrological processes under the conditions of limited data availabilities. Correspondingly, techniques of stepwise cluster analysis (SCA) can be employed for reflecting nonlinear relationships between meteorological factors and hydrological process, as well as dealing with non-continue datasets through cluster trees (Fiseha et al. 2014). On the other hand, large discrepancies among global climate models (GCMs) exist with respect to regional projections (Llopart et al. 2020). Nevertheless, downscaling procedures are based on the empirical relationships between large-scale atmospheric predictor variables and local surface parameters such as precipitation and temperature (Genç 2021). Correspondingly, dynamic downscaling techniques (e.g. regional climate models) can be employed for generating high-resolution meteorological variables, and for displaying the underlying climate change signal (Meng et al. 2016). Besides, climate change projection can be employed for impacts evaluation of hydrological process, e.g., the frequency of extreme events such as droughts and floods (Ning et al. 2016;Wang et al. 2022).

Methodology
Currently, the data distribution center of the intergovernmental panel on climate change (IPCC) provides links for datasets from various global climate models (GCM) on related scenarios for impact assessments (Ma et al. 2020). The CMIP5 (coupled model intercomparison project phase 5) experiments include the historical climate simulation experiments for the 20th century and prediction experiments for the 21st century driven by "Representative Concentration Pathways" concentrations (Zhu et al. 2020). However, the resolutions of GCM models are too coarse to represent fine-scale physical processes in the climate system (Wilby 2005) especially for some complex terrestrial regions such as the upper Yellow River reach. The methods used to convert GCM outputs into local meteorological variables required for reliable hydrological modeling are usually referred to as downscaling techniques (Milly et al. 2005). Different approaches have been developed for downscaling/bias-correcting which can mainly be classified into two categories of statistical and dynamical downscaling (Ramteke et al. 2020). Specifically, the dynamical downscaling techniques mainly rely on regional climate models, which aim to downscale climate fields produced by coarse resolution GCMs, thereby provide information at fine, sub-GCM grid scales more suitable for studies of regional phenomena and for application to vulnerability, impacts, and adaptation (VIA) assessments (Giorgi 2019). The providing regional climate impacts for studies (PRECIS) model is developed by the Met Office Hadley Center, which has been one of the widely-used RCMs applied over any area of the globe to provide regional climate information for impacts studies (Wang et al. 2015). The PRECIS model can be driven by outputs from diverse GCMs (e.g. Hadley Centre Global Environment Model version 2 (HadGEM2)), and is able to transfer the resolution of GCMs (1° × 1°) to be fine enough (25 km × 5 km or 0.22° × 0.22°) to capture high resolution region-specific climate responses. More detailed introductions are described by Taylor et al. (2012).
Daily outputs from global climate models (GCMs) under the medium emission (RCP4.5) and high emission scenarios (RCP8.5) are downscaled by PRECIS model for generating high -resolution datasets. Meteorological data from 1960 to 2018 have been provided by the National Meteorological Information Centre of China, and can be accessed from the China Meteorological Forcing Data Sharing Service System (http:// data. cma. cn/). Meteorological datasets used in this study are precipitation (P, mm), minimum (T min ), maximum (T max ), and mean (T mean ) air temperature (°C). Correspondingly, the downscaled maximum and minimum temperature and precipitation time series are used as inputs to drive stepwise-clustered hydrological model (SCHM). Figure 1 presents the framework of 1 3 hydrometeorological variables integrated with the SCHM. The hydrological model based on stepwise cluster analysis (SCA) is developed to imitate the complex nonlinear relationships between climate input variables and targeted hydrological variables. The stepwise cluster analysis (SCA) is used in SCHM as it can capture discrete and nonlinear relationships and produce non-functional maps between explanatory and response variables. This method has been used for a number of studies especially in climate impact analysis and streamflow predictions (e.g., Wang et al. 2013, Fan et al. 2016, Wen et al. 2022, which have demonstrated the capabilities of SCA for reflecting complex processes in hydrological cycle. In detail, a cluster tree can be generated to reflect predictands (streamflows) and predictors (such as temperature, precipitation). Since stepwise cluster analysis method has the ability to screen the most relevant and important predictors by identifying the predictor that leads to a minimum Wilks' statistics during the tests of optimal cutting points, all the predictor candidates that might affect streamflow could be considered as predictors.
The criteria for cut and merge operation of SCA is based on Wilk's statistic. According to Wilks' likelihood-ratio criterion, if the cutting point is optimal, the value of Wilks' Λ ( Λ = |W|∕|W + H| ) should be minimum (Wilks 1962), where W and H are within and between-group sums of squares and cross products (SSCP) matrices. The symbol of |W| indicates the determinant of a matrix. Consider two sets of dependent variables e and f denoted as the following vectors: e i = e i1 , e i2 , … , e ir , i = 1, 2, … , n e ; f j = f j1 , f j2 , … , f jr , j = 1, 2, … , n f , where n e and n f are the sample size of e and f, respectively, and r is the dimension of the dependent variables, the H and W can be given by:    where e and f is the sample mean of set e and f, respectively: According to Rao's F-approximation (R-Statistic) (Rao 1965), we can have: The nonlinear relationship between the predictors and streamflow is identified and defined using a cluster tree, where the values of predictors (input variables) determine which cluster (i.e., end node) to enter, and mean value of the samples in the end-node determines the predicted flow value (the output variable). The flow chart of SCA is shown as Fig. 1.

Case Study
The upper Yellow River has a mainstream length of 3,472 km, from the origination to Hekou district as presented in Table 1. The upper Yellow River, with a drainage area of 385,996 km 2 , has an annual runoff of 2.04 × 10 10 m 3 , accounting for 34.5% of total annual runoff of the Yellow River basin (Zhang et al. 2009). Datasets of monthly and daily streamflow for the upper Yellow River are collected at the Tangnaihai hydrological station. The Tangnaihai hydrological station, located in the northeast of the Tibetan Plateau, is considered as a key hydrological monitoring station in the upper Yellow River basin (Fig. 2). Long term climatic records suggest a noticeable warming trend of 0.31-0.35 °C/10-year over the upper Yellow River in the past 50 years (Xin et al. 2014).
Based on the SCHM, the complex interrelationship between meteorological inputs and streamflow will be established for the upper Yellow River, which is to be further employed to reveal potential impacts of climate change on water availability in the upper Yellow River, and provide scientific support for developing resilience strategies to mitigate the adverse impacts of climate change. The simulated results generated from SCHM during calibration and validation will be evaluated by using the Nash-Sutcliffe efficiency (NSE) (Nash and Sutcliffe 1970), and the normalized root mean square errors (NRMSE). The criteria are defined as follows: Fig. 2 The upper Yellow River Basin and its main gauging stations where Q i is the observed flow in month i; Q max and Q min is respectively the maximum and minimum monthly flows; C i is the simulated flow on month i; n is the number of simulated months; Q is the average measured flow.

Characteristics of Meteorological Variables Under Climate Change
The upper Yellow River is located in a complex terrestrial terrain, flowing across both plateau (Tibetan Plateau and Losses Plateau) and plain (Hetao) regions, and having a total drop of nearly 3500 m and an average slope of 0.1%. These terrestrial complexities lead to noticeable spatial variations for both meteorological and hydrological variables, especially under climate change. In this study, the meteorological variables for streamflow forecasting are downscaled PRECIS, which can generate high-resolution projections over the upper Yellow River region. The performance of the PRECIS model on the projections of climate change over China has been demonstrated by some studies (e.g., Zhu et al. 2018;Guo et al. 2019). Datasets of daily precipitation, maximum, minimum and average temperature are chosen as predictand variables for further streamflow forecasting in the upper Yellow River region. Forty years  of historical records of predictands are available at Tangnaihai station and are used for calibrating and validating the developed SCHM. Figure 3 presents the comparison of monthly minimum temperature (i.e., Tmin) during the baseline period (1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004) and the two future periods (i.e. 2020s (2020-2030) and 2030s (2031-2039)) under RCP4.5 and RCP8.5. Explicit increases will be expected for both RCPs in the two projection periods. However, the increasing pattern under  Fig. 3b and d that the Summer time would have more temperature increase in 2030s under RCP8.5 than that under RCP4.5, with the temperature increment around 2 °C. In general, the annual mean temperature increment would be around 1.34 and 1.96 °C in the two projection periods under RCP4.5, and around 1.37 and 2.13 °C in the two projection periods under RCP4.5. The comparison for monthly maximum temperature (i.e., Tmax) among the baseline and two projection periods under two RCPs are presented in Fig. S1 in the Supplementary Materials, whilst Fig. S2 shows the comparison for monthly mean temperature (i.e., Tmean). The changing patterns for Tmax and Tmean in the two future projection periods under each RCP show similar features with the temperature changes of Tmin. In detail, the largest temperature increment may happen in December for Tmin, Tmax and Tmean under RCP4.5, while under RCP8.5, a hotter summer may occur than that under RCP4.5. However, there are also some differences for temperature increments for each temperature index. For instance, the temperature increment in the 2030s would not necessarily be larger than the increment in 2020s, such as the changes of Tmax in December under RCP4.5. Moreover, the annual mean increment for Tmax is respectively 1.36 and 1.66 °C for 2020s and 2030s under RCP4.5, while such a Tmax increase under RCP8.5 is 1.00 °C in 2020s and 1.82 °C in 2030s. Also, the Tmean increase under RCP4.5 is 1.32 and 1.80 °C respectively for 2020 and 2030s, whilst the temperature increment for Tmean under RCP8.5 would be 1.20 °C during 2020s and 1.97 °C during 2030s. These results imply that even though the scenario of RCP8.5 is associated with higher GHG emissions than RCP4.5, more temperature increase resulting from higher GHG emissions may occur after 2030s in the upper Yellow River region.
Precipitation is a key component of the hydrological cycle and the changes in hydrological processes are mainly reflected by the temporal and spatial variations in precipitation. Figure 4 shows average monthly precipitation in the baseline period and two projections periods under RCP4.5 and RCP8.5. It can be observed from Fig. 4a and c that the future Fig. 4 Comparison of mean values for monthly precipitation between historical simulation and future projections under two RCPs precipitation variations under the two RCPs would show similar features with the precipitation in the baseline period with most precipitation occurring from April to October. However, there are still some explicit differences for the precipitation between the baseline and two projection periods under two RCPs. As presented in Fig. 4b and d, the precipitation changes (between future and historical rainfall) show different patterns in two projection periods (i.e., 2020s and 2030s) under two RCPs. For instance, the upper Yellow River region may have minor precipitation decrease in July in 2020s but the largest increase (around 15 mm/month) in 2030s under RCP4.5. In comparison, explicit increment may be observed in July in 2020s under RCP8.5 but obvious rainfall reduction may happen in 2030s. In general, more precipitation may occur in both 2020s (around 3 mm/year) and 2030s (16 mm/year) under RCP4.5. Under RCP8.5, precipitation may have an increment around 31 mm/year in 2020s but some decrease (around 12 mm/year) in 2030s.

Climate Change Impacts on Hydrological Process
To quantify the climate change impact on basin hydrology, the prediction of potential runoff variation using projected climate variables as inputs would be generated from the stepwise-clustered hydrological model (SCHM). In detail, the relationships between climatic variables and the streamflow in the upper Yellow River are quantified through the SCHM based on historical observations at the Tangnaihai station. In detail, the monthly observations from 1961 to 2000 are employed to establish the stepwise-clustered hydrological model (SCHM) for the upper Yellow River basin whilst the data from 2005 to 2012 are adopted to verify the applicability of the developed hydrological models. The established SCHM is then applied for revealing the impacts of climate change on the streamflow in the upper Yellow River basin. Figure 5 shows the cluster tree of monthly streamflow generated by SCHM under α value of 0.05. The cluster tree consists of 424 codes, with 172 cutting actions and 79 merging actions. Figure 6a presents the monthly time series of simulated and observed streamflow at the Tangnaihai station during the calibration period . The monthly-average simulated and observed streamflow would respectively be 687.9 and 701.9 m 3 /s in 1960s (1961)(1962)(1963)(1964)(1965)(1966)(1967)(1968)(1969) (as shown in Table 2). In contrast, the monthly-average simulated and observed streamflow would be 592.8 m 3 and 549.4 m 3 /s in 1990s (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000), respectively. Consequently, the stepwise-clustered hydrological model (SCHM) can well reflect the complex relationships between climatic variables and river flows. The simulated and observed streamflow during the validation period (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012) would be shown in Fig. 6b, which indicates that the predictions from the SCHM would fit well with the corresponding observations in most time periods.
The calculated Nash-Sutcliffe efficiency (NSE) and normalized root mean square errors (NRMSE) values are 0.89, and 0.054 respectively in the calibration period , whilst these two evaluation indices would respectively be 0.72 and 0.12 in the validation period (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012). Both the NSE and NRMSE values imply a good and acceptable relationship mapping between the mean values of observed and simulated streamflow rates. Indeed, validation results indicate that the calibrated factor values (temperature and precipitation) are acceptable in predicting streamflow. Consequently, the developed SCHM can be used for characterizing the impacts of climate change on the water resources in the upper Yellow River basin.

Conclusions
In this study, a climate-streamflow framework (CSF) has been advanced for assessing climate change impacts on the hydrology of the upper Yellow River basin, which incorporated global climate model, dynamic downing technique and stepwise cluster analysis into a hydrological framework. CSF has the following advantages: (1) it can operate different temporal resolutions of variables, through dealing with continuous and discrete meteorological variables; (2) it can transfer large-scale factors to generate high-resolution meteorological variables by using PRECIS model; (3) it can project characteristics of climate change (temperature and precipitation patterns) and their impacts on streamflow by applying stepwise-clustered hydrological method with the downscaled meteorological factors. The CSF has been applied to upper Yellow River basin to reveal the impacts of climate change on the monthly streamflow. Results disclose that monthly average temperature would increase in the two projection periods (i.e., 2020s, 2030s) under RCP4.5 and RCP8.5. However such increases in Tmin, Tmean and Tmax may show some discrepancies among different periods under different RCPs. Compared with the baseline period (i.e. , 1986-2004), the largest temperature increments may happen in January for all the three temperature indices under RCP4.5. Moreover, a hotter summer may be expected under RCP8.5 than that under RCP4.5 especially in 2030s. Compared with the temperature variations, the precipitation changes show discrepant patterns in which there are no visible variations in Winter in future under both RCPs whilst rainfall decrease may happen in May and June in future. Moreover, under different RCPs, the precipitation may change dramatically and even inversely.
By applying CSF in upper Yellow River, the variability of observed streamflow can be reflected in simulation characteristics. The calculated NSE and NRMSE between observed and simulated streamflow would be 0.89 and 0.054 during calibration , and 0.72 and 0.12 in the validation (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012) periods. The variability of streamflow during future period (2020 and 2030s) has been assessed through the stepwise-clustered hydrological model driven by the projections of average temperature, maximum and minimum temperatures, and precipitation. The results indicate that the monthly streamflow rates in Winter under both RCPs would be similar to those in the baseline period. The significant streamflow changes in two future periods would be more likely to occur between May and October, in which remarkable streamflow reduction may be observed in May but obvious streamflow increments may be expected in July.