Evaluating the impact of climate change on surface water resources in the upper Ganjiang River Basin, China

Based on the IPCC AR5, global annual mean temperature exceeded 0.85 °C since 1880, and the annual mean temperature in China increased >1.0 °C in the past three decades, which was dramatically higher than the world's average (Warszawski et al. 2014; Fang et al. 2018). In the next 30 years, surface temperature is expected to rise ranging from 2.4 to 6.1 °C in China, while annual runoff may increase roughly by 3–10% (Liu et al. 2014). Influenced by the monsoon climate, China has become particularly sensitive to the global warming, which may bring about a descent in streamflow in some basins, such as the lower Yellow River Basin (Piao et al. 2010). Research demonstrated that climate change was detected to be a dominant driving factor yielding runoff reduction in major basins across China, and it was estimated to account for 53.5% of streamflow changes during the period 2001–2014 (Liu et al. 2017). Furthermore, a change in climate owing to ascending temperature and intensified evaporation has aggravated the situation of hydrological cycles, altered spatio-temporal distribution of water resources and further enlarged the evolution of climatic variables (e.g., precipitation and evaporation), and these, in turn, will inevitably produce an effect on the ecological system (John et al. 2020). Therefore, projecting climate change and its impacts on surface runoff are essential for allocation and scheduling of water resources, and researchers have conducted much significant work in recent years, basically following the pattern of future climate scenario design-hydrological model-impact assessment.

General circulation models (GCMs) have been used extensively to characterize future climate change, which can reflect the dynamic change of climate system through numerical simulation (Zinyengere et al. 2017). Employing the outputs provided by GCMs to drive a hydrological model is a widely adopted method in view of the fact that the hydrological model is a useful tool for simulating runoff processes (Chen et al. 2012). To compare the impact of two delta-change methods on different precipitation scenarios, Yuan et al. (2016) applied the Xinanjiang (XAJ) model to predict surface streamflow, and more accurate runoff estimates can be obtained in the watershed with stronger precipitation change signals in the Pearl River Basin. Zhu et al. (2019) used the Soil and Water Assessment Tool (SWAT) model to discuss the uncertainty of future runoff and concluded that monthly runoff increases are concentrated between May and October (July and August are not included) in the Biliu River Basin. Jin et al. (2020) employed the Variable Infiltration Capacity (VIC) model to depict the water cycle process, suggesting that annual precipitation may ascend while runoff may descend in the Yellow River source region. Yan et al. (2020) applied the Water and Energy Transfer processes in Large basins (WEP-L) model integrated with the global Bayesian model averaging (BMA) method to make the runoff projections become convincing and revealed that runoff in the Yellow River Basin during the period 2050–2070 would be extraordinarily lower than that in the baseline period. Although there are structural differences in different GCMs, the temperature changes they reflect are quite consistent, whereas precipitation features appear as large fluctuations across various models (Wen et al. 2016). Sun & Peng (2020) took the RCP4.5 scenario as an example to estimate the range of runoff variation and reported that future annual runoff varies between 268.9 and 544.2 mm from 2011 to 2100. In summary, the hydrological model makes it possible to obtain more potential changes and detailed information on various climatic characteristics of the catchment by predicting surface runoff and providing continuous simulations of daily and monthly runoffs. Moreover, this approach can enhance the simulation accuracy in surface runoff by making use of satellite-based data (Crow & Ryu 2009). Nevertheless, considering that the hydrological model is a data-intensive method, an estimation of future runoff is complicated and fraught with uncertainties. Also, this approach is always constrained by multiple data sets (e.g., vegetation, topography, soil) (Teng et al. 2012). Thus, it is evident that a variety of uncertainties are introduced through the hydrological model in the form of model parameters and structures, and the impact generated by the hydrological model may be unneglectable especially for the complex model (Thompson et al. 2013). As a consequence, the hydrological model may not be the most appropriate choice to estimate future runoff in some cases, and it may be inapplicable for the large-scale catchment with insufficient data or limited hydrological information.

In contrast with the hydrological model, the water-balance method is identified as an alternative technique owing to its fewer model parameters and simpler calculation principles, and this approach has received tremendous attention in speculating the hydrological process under changing environments. Chen et al. (2017) analyzed that >60% watersheds in China would face water deficiency and revealed the increasingly unbalanced distribution of water resources since the middle of the last century. Xing et al. (2018) established an elasticity method to estimate future runoff in 35 catchments across China, and the Wang–Tang equation was the most reasonable water-balance equation to strengthen the prediction reliability. Yuan et al. (2018b) constructed a basin-scale water-balance approach and investigated that natural runoff would increase by 14.4–16.8% under the RCP4.5 scenario over the Yellow River Basin. Zhang et al. (2018b) quantified the impact of leaf area index (LAI) changes on the streamflow, and it is estimated that the downtrend in runoff is likely to sustain. Li et al. (2020) applied the elasticity method coupled with the Blaney–Criddle equation into the upper Heihe River Basin and reported that the predicted runoff may decline by 5.6 and 6.7% during two periods 2021–2050 and 2051–2080, respectively. Guan et al. (2021) established the elasticity method to predict runoff and investigated that increasing runoff during the projection period may alleviate water shortages in northern China. Previous studies have illustrated the feasibility of the water-balance model coupled with GCMs in ascribing future streamflow patterns, and the estimated runoff calculated from the Budyko's framework is quite similar to that calculated from the hydrological model (Li et al. 2013). It is also worth pointing out that investigative assessments are always earlier than the hydrological modeling, where the Budyko hypothesis is widely accepted to address the trouble of flexibly distinguishing the influences of climate change on surface streamflow (Kazemi et al. 2019). Furthermore, the water-balance model based on the single-parameter Budyko-type equation is particularly suitable for ungauged catchments, because only estimates of long-term water availability are needed in these catchments (Chiew 2010).

The Ganjiang River Basin (GRB), located in southern China with abundant water resources, is the largest and longest river in Jiangxi province. Over the past decades, the GRB has undergone severe water shortages. The contradiction between water supply and demand is particularly remarkable, and the phenomenon of upstream and downstream competition for water resources is increasingly highlighted in the dry season (Zhang et al. 2018a). Multi-year mean runoff in the upper Ganjiang River basin (UGRB) accounts for about 22.5% of that in the Poyang Lake Basin (PLB), and it is the largest among all three-grade water resource partitions over the PLB (Yu et al. 2014). Projecting future surface runoff in the UGRB is crucial for minimizing these adverse impacts, as it would be beneficial to develop more effective adaptation measures and achieve sustainable water resource management. Aimed at probing into climate change and its effects on surface water resources in the UGRB, this study is organized as follows: (1) to calibrate a single parameter within the Budyko's framework and compare the feasibility of various water-balance models, and to seek for the optimal Budyko-type equation for capturing the runoff variation; (2) to build the statistical relation between temperature and potential evapotranspiration and to investigate the sensitivity of surface runoff to climatic variables; and (3) to further project the future changes of hydrometeorological factors during the period 2021–2050 under various GCMs, with taking multiple representative concentration pathways (RCPs) into account. The research framework of this paper is shown in Figure 1.

The GRB is situated in 113°34′–116°37′E and 24°31′–28°45′N, covering an area of >83,500 km². As the longest river of the PLB, the Ganjiang reaches a length of 766 km and lies mainly in the Jiangxi Province. The GRB has four distinctive seasons, that is, summers are hot and rainy while winters are dry and cold, with annual precipitation ranging from 1,400 to 1,800 mm and mean temperature of 17.8 °C (Song et al. 2020).

The Wan'an Reservoir is located 2 km upstream of Wan'an County, and it plays an important role in alleviating the water crisis and maintaining the availability of water resources. The Dongbei hydrological station is the outlet control station over the UGRB, which can well reflect the runoff changes of the study area. Annual rainfall amount, surface water resource amount, and groundwater resource amount in the UGRB are 633.2 × 10⁸, 344.7 × 10⁸ and 95.9 × 10⁸m³, equivalent to 1,574, 857 and 238 mm, respectively (Yu et al. 2014). The location of the UGRB is illustrated in Figure 2.

Monthly runoff data from 1959 to 2017 were supplied by the hydrological yearbook issued by the local hydrological bureau. Observed daily meteorological data (1959–2017) of 13 meteorological stations (i.e., Ganxian, Changting, Guangchang, Ji'anxian, Xunwu, Longnan, Nanxiong, Ninghua, Ningdu, Nanfeng, Yongfeng, Suichuan and Jingangshan, labeled in Figure 2) were collected from the National Meteorological Information Centre (NMIC). The Penman–Monteith method is employed to calculate potential evapotranspiration for a single meteorological station. Daily precipitation, temperature and evapotranspiration of the single station are reorganized into monthly precipitation, temperature and evapotranspiration, and then precipitation, temperature and evapotranspiration in the study area are calculated according to the Thiessen polygons method, respectively. It should be noticed that runoff depth is used to represent surface water resources. Precipitation, evapotranspiration and surface runoff during the reference period (1961–2010) are 1,560, 1,466 and 821 mm, respectively.

RCPs were defined in terms of a possible range of global radiation forcing value in 2100, and future mitigation emission scenarios are ranked as RCP2.6, RCP4.5, RCP6.0 and RCP8.5. Supposing that the forcing level will reach the highest point at ∼3 W/m² before 2100, if this value drops to 2.6 W/m² by 2100, it stands for the RCP2.6 scenario; if this value stabilizes without overshoot pathways to 4.5 W/m² after 2100, it represents the RCP4.5 scenario; if this value stabilizes without overshoot pathways to 6.0 W/m² after 2100, it means the RCP6.0 scenario; if this value climbs rapidly to 8.5 W/m² by 2100, it symbolizes the RCP8.5 scenario (Yuan et al. 2018a).

Continuous daily-series meteorological data were derived from multiple GCMs. These climate models were chosen since they were widely accepted with persuasive feasibility in climate impact studies in China and can provide possible explanations concerning future climate which might make differences to some significant areas (Zhou et al. 2018). In addition, it should be noticed that Yuan et al. (2019) have demonstrated that these five climate models chosen in this paper are applicable in the similar catchment, the Fu River basin. Meteorological data contained the period from January 1951 to December 2050, with the geographical scope covering 70.25°–140.25°E and 15.25°–55.25°N (Piani et al. 2010). Precipitation and temperature data used in this study were bias-corrected and resampled, so as to ensure consistency between the observed data and the predicted data. In this paper, the statistical downscaling method and bilinear interpolation were used to uniformly interpolate the output data to a 0.5° × 0.5° grid, and these data were corrected based on the statistical deviation of probability distribution. A detailed description of five GCMs is listed in Table 1 (Taylor et al. 2012).

Three RCPs including RCP2.6 (low), RCP4.5 (medium) and RCP8.5 (high) scenarios were chosen to project the future precipitation, temperature and discharge variation, and 21 grid cells representing the future climatic variables in and around the UGRB are marked in Figure 2.

Observed data during the period 1959–2017 were adopted to calibrate and verify the water-balance model, and the study period was divided into two parts: the calibration period (1959–1998) and the validation period (1999–2017). Then, a calibrated water-balance model was driven by climatic variables during the projection period (2021–2050) provided by the GCMs, and the relative variation characteristics of simulated runoff are used to characterize the trend of runoff.

For reasons of ensuring the water-balance model performs well in the study period, it is most important to calibrate the parameters in various Budyko-type equations during the period 1959–1998, followed by verifying these parameters during the period 1999–2017. After calibration, dimensionless parameters ⁠, ⁠, and k are 1.70, 0.03, 0.02 and 0.99, respectively. The performance for different Budyko-type equations is shown in Table 2.

It can be concluded that all RE are <11%, all NSE are >0.84 and all DC are >0.84 in the calibration period, while all RE are <15%, all NSE are >0.64 and all DC are >0.72 in the validation period. Therefore, it is evident that four Budyko-type equations are highly consistent in runoff simulation for meeting the requirements of accuracy, and the water-balance model is applicable in the study area.

In contrast with other equations, the Budyko-CY equation has a better performance with respect to the NSE during two periods. As a consequence, the Budyko-CY function was screened as the optimal equation to project the runoff variation in the near future (2021–2050). Comparison between the simulated runoff and the observed runoff is shown in Figure 3.

The statistical relation between monthly mean temperature and monthly potential evapotranspiration was established via the MATLAB toolbox, taking the observed data during the period 1961–2010 into account. Figure 4 shows that monthly temperature and evapotranspiration conform to the non-linear relation composed of an exponential function and a linear function, and it was clear that this equation can fit the statistical relation well with the square of correlation coefficient of 0.894.

Based on daily-series meteorological data of 21 grid cells in and around the UGRB generated by various GCMs, monthly precipitation and average temperature of 21 grid cells can be counted at first, and monthly precipitation and average temperature in the study area can be summarized. Then, monthly potential evapotranspiration in the study area can be calculated by means of the statistical relation proposed in Figure 4. Subsequently, surface runoff under multiple climate scenarios can be simulated with reference to the Budyko-CY function expressed in Equation (7). Ultimately, projected climate change and its possible influences on water resources can be estimated according to Equation (8), and the response of surface runoff to the variations in precipitation and temperature is shown in Figure 5.

In terms of the response of surface runoff to precipitation, on the condition of given precipitation, when annual temperature rises by 1 °C, the annual surface runoff is expected to decline by 2.5–4.7% on average, or 22.6–42.5 mm. In the case of a 30% ascent in precipitation, every 1 °C change in temperature may lead to the largest variation range in runoff, and corresponding runoff is expected to decrease by 7.8% when average temperature rises from 2 to 3 °C. Relatively speaking, in the case of a 30% reduction in precipitation, variation amplitude in runoff is the smallest for 1 °C change in temperature, and corresponding runoff is expected to decrease by 1.6% when average temperature changes from −3 to −2 °C. The foremost reason is that actual regional evaporation is relatively large under warm and humid conditions. Therefore, runoff is more sensitive to temperature rise under given precipitation.

In terms of the response of surface runoff to temperature, under the circumstance of given temperature, for every 10% increase in precipitation, annual surface runoff is expected to rise from 12.1 to 14.3% on average, that is, 110.7–130.6 mm. In the case of a 3 °C ascent in temperature, every 10% change in precipitation may cause the smallest variation range in runoff, and corresponding runoff is expected to increase by 11.0% when annual precipitation varies from −30 to −20%. Instead, in the case of a 3 °C reduction in temperature, variation amplitude in runoff is the largest for 10% variation in precipitation, and corresponding runoff is expected to increase by 15.0% when annual precipitation changes from 20 to 30%. Chiew (2010) revealed that the greater the runoff coefficient is in the watershed, the less sensitive it is to climate change. As far as the UGRB is concerned, its multi-year mean precipitation and streamflow during the period 1961–2010 are 1,560 and 821 mm, so the runoff coefficient is 0.53 over the UGRB. Therefore, the UGRB is not sensitive to climate change, and the analysis based on the water-balance model is consistent with it.

A comparison of three climatic factors is shown in Tables 3, where base represents the observed monthly average during the baseline period, while RCP2.6, RCP4.5 and RCP8.5 represent the projected ensemble mean value of five GCMs during the projection period (2021–2050).

In the next 30 years, the projected multi-year mean precipitation will vary as 10, −9 and −8 mm under three RCP scenarios, while the projected temperature will increase by 0.69, 0.68 and 0.97 °C, the projected potential evapotranspiration will increase by 162, 148 and 232 mm, respectively. Taking the temperature as an example, relative to the reference period, an average temperature will rise every month, especially in September and October, with warming by 1.1/1.1/1.6 °C and 1.0/0.9/1.2 °C under three RCPs. Similarly, the increased potential evaporation is mainly concentrated in July and August with monthly variation of 40/46/56 mm and 61/43/84 mm, respectively. The variation in monthly precipitation is specifically manifested as follows: rising in June, August and September, and falling in May, July and October.

The projected annual precipitation in the UGRB is shown in Figure 6. The variation range and the ensemble mean for each RCP are shown concurrently, with the assumption that the ensemble mean is able to characterize the variation trend of annual precipitation, temperature and runoff, and five global climate models are equally plausible. It is worth noting that the solid lines in red, blue and green represent the multi-mode ensemble mean of annual runoff under three climate scenarios (i.e., RCP2.6, RCP4.5 and RCP8.5), respectively, whereas the dotted lines of corresponding colors represent the maximum and minimum annual runoffs in various GCMs during the projection period of 2021–2050. It is obvious that the annual precipitation under the RCP2.6 scenario presents the greater fluctuation of 558 mm due to its variation range, while that under both the RCP4.5 and RCP8.5 scenarios is 494 mm. The maximum and the minimum annual precipitation during the period of 2021–2050 are 2,465 mm (GFDL-ESM2M) and 900 mm (GFDL-ESM2M) under the RCP2.6 scenario, while they are 2,809 mm (GFDL-ESM2M) and 667 mm (MIROC-ESM-CHEM) under the RCP4.5 scenario and 2,560 mm (GFDL-ESM2M) and 693 mm (GFDL-ESM2M) under the RCP8.5 scenario, respectively. It is not difficult to find that the GFDL-ESM2M model is closely associated with annual rainfall extremes.

Figure 7 shows multi-year mean precipitation of various GCMs under three climate scenarios. It can be concluded that multi-year mean precipitation during the projection period changes by −8.5 to 8.7% (RCP2.6), −4.5 to 3.9% (RCP4.5) and −8.1 to 6.3% (RCP8.5), respectively. The average precipitation in the GFDL-ESM2M, HadGEM2-ES and NorESM1-M models will increase, with variation ranges of 1.6–7.2, 0.3–2.2 and 1.4–8.7%, while that in IPSL-CM5A-LR and MIROC-ESM-CHEM models will decrease, with variation ranges of −8.5 to −4.5% and −8.1 to −4.2%, respectively. In aggregate, the IPSL-CM5A-LR model has the minimum variation in future projected precipitation, with the average change of −6.9%, while the NorESM1-M model has the maximum variation with the average change of 5.4%.

The projected annual temperature in the UGRB is shown in Figure 8. On a multi-year average scale, it is clear that temperature shows an upward trend in the projection period with the gradient of 0.29 °C/10a, 0.25 °C/10a and 0.61 °C/10a under three RCP scenarios, respectively. Especially in the last 10 years of the projection period, temperature is expected to increase by 1.03, 0.94 and 1.68 °C relative to that during the period of 1961–2010, suggesting that future temperature is expected to increase in the UGRB. The annual mean temperature under the RCP8.5 scenario presents the greater variation amplitude, with the maximum value of 21.74 °C in 2050 and the minimum value of 19.40 °C in 2040, while the variation amplitudes under RCP2.6 and RCP4.5 scenarios are 1.37 and 1.45 °C, revealing that the higher the temperature rises, the greater the variation range in temperature will be.

Figure 9 shows a multi-year mean temperature of various GCMs under three climate scenarios. From the perspective of RCP, the multi-year average temperature is expected to increase by 0.17–1.43 °C (RCP2.6), 0.22–1.18 °C (RCP4.5) and 0.53–1.72 °C (RCP8.5), respectively. When it comes to the GCM, all five models show a rise in temperature, with variation ranges of 0.22–0.53 °C (GFDL-ESM2M), 0.92–1.23 °C (HadGEM2-ES), 0.64–1.03 °C (IPSL-CM5A-LR), 1.18–1.72 °C (MIROC-ESM-CHEM) and 0.17–0.64 °C (NorESM1-M), respectively. The GFDL-ESM2M model has the minimum warming mean temperature of 0.33 °C, while the MIROC-ESM-CHEM model has the maximum warming mean temperature, reaching up to 1.44 °C under three RCPs. Different from the MIROC-ESM-CHEM and HadGEM2-ES models, the GFDL-ESM2M and NorESM1-M models both show a small increase in temperature under all RCP scenarios.

The projected annual runoff in the UGRB is shown in Figure 10. The annual mean runoff is expected to increase by 8.4, 6.5 and 5.0% from 2021 to 2050 in contrast with that from 1961 to 2010, suggesting that surface water resources will increase in the next 30 years. The maximum runoff depth values are 1,190, 1,116 and 1,092 mm, while the minimum runoff depth values are 725, 692 and 662 mm under three RCPs, respectively. Studies demonstrated that a higher temperature may bring about a decrease in annual runoff in the UGRB. Nevertheless, different from RCP2.6 and RCP4.5 scenarios, the annual runoff under the RCP8.5 scenario shows a downward trend during the projection period, especially after 2040.

Figure 11 shows multi-year mean runoff depth of various GCMs under three climate scenarios. Surface water resources will increase in the future under various scenarios, except for the MIROC-ESM-CHEM model under three RCPs, as well as the IPSL-CM5A-LR model under RCP2.6 and RCP8.5 scenarios. The variation in projected runoff is similar between the GFDL-ESM2M model and the NorESM1-M model, because the former has the 21.2% ascent under the RCP2.6 scenario and 16.4% ascent under three RCPs, while those in the latter are 23.3% rise and 17.3% rise, respectively. On the contrary, the MIROC-ESM-CHEM model seems to be on a downward trajectory, with the variation amplitude of −12.5 to −2.1% and the mean variation value of −6.4%. The variation range under the RCP4.5 scenario is the smallest, varying from −2.05 to 16.02%, while variation ranges under RCP2.6 and RCP8.5 scenarios are −4.55 to 23.25% and −12.46 to 17.53%, respectively.

The variations of the projected multi-year mean precipitation and temperature over the UGRB during the period 2021–2050 are relatively close to the conclusions proposed by Deng & Wang (2019), and both are precipitation declines and potential evaporation increases; however, in terms of annual runoff, the opposite conclusions are drawn. In this paper, the annual mean runoff is expected to increase by 8.4, 6.5 and 5.0% under three RCPs in the next 30 years, while the latter reported a decreasing trend about −9.9, −19.5 and −11.4%. The reason for the opposite conclusions of projected runoff lies in the different climate models selected. Taking climate models as an example, multiple global GCMs from different countries were selected in this paper, while the latter just chose a climate model, namely BCC CSM1.1. Zhao et al. (2020) applied the HBV model and the same five climate models into the GRB to reveal that annual runoff showed an upward trend under the RCP4.5 scenario, while that may decrease after the middle of the 21st century under RCP2.6 and RCP8.5 scenarios. It is evident that the selected hydrological model, study area and study period all have a great influence on the projected runoff.

Since daily-series data of climatic factors are generated by multiple GCMs, the rationality of the results obtained is directly related to the selected climate models, which will create uncertainties inevitably. If different climate models are adopted, the different analyses and conclusions may be achieved in practice. Thus, it is still a challenge to quantitatively describe the variation trend and range for climatic variables.

In this study, precipitation, temperature and runoff in the UGRB were projected, and the impact of climate change on water resources was evaluated by employing the Budyko-type functions driven by multiple representative GCMs. The conclusions are summarized as follows:

• The constructed water-balance model performed well on simulating annual runoff, because NSE and DC are both >0.84 and >0.64 during two periods. The Budyko-CY function was the optimal water-balance equation screened from the Budyko-Fu, Budyko-Zhang, Budyko-WT and Budyko-CY functions, for its having the maximum NSE. Compared with the hydrological model, Budyko-type functions have the advantage of few parameters and simple calculations. Therefore, this approach is appropriate in predicting future streamflow in large-scale regions or ungauged catchments with limited hydrological information.

• When temperature rises by 1 °C, the surface runoff depth is expected to decline by 2.5–4.7% on average, that is 22.6–42.5 mm; when precipitation rises by 10%, surface runoff is expected to rise by 12.1–14.3% on average, that is, 110.7–130.6 mm. The rainfall elasticity of streamflow is strongly correlated to the runoff coefficient, annual rainfall and streamflow, where the runoff coefficient in the UGRB is 0.53, suggesting that this catchment is not sensitive to a climate change.

• The projected temperature over the UGRB presents an obvious escalating trend during the period 2021–2050 with a slope of 0.29 °C/10a, 0.25 °C/10a and 0.61 °C/10a under three RCP scenarios, respectively. The annual mean temperature is estimated to rise by 0.17–1.43, 0.22–1.18 and 0.53–1.72 °C, while the annual runoff is expected to increase slightly with the variability of 8.4, 6.5 and 5.0%, respectively. Five GCMs all show a rise in temperature, with variation ranges of 0.22–0.53 °C (GFDL-ESM2M), 0.92–1.23 °C (HadGEM2-ES), 0.64–1.03 °C (IPSL-CM5A-LR), 1.18–1.72 °C (MIROC-ESM-CHEM) and 0.17–0.64 °C (NorESM1-M), respectively.

As a whole, this research provides possibilities for climate scenarios that may occur in the future and results presented herein are capable of giving a reference to rational allocation and optimal scheduling of water resources in the UGRB. In addition, leveraging of the characteristics of a variety of climate models to prospect surface runoff will remain a tough task over the coming decades.

This research was financially supported by the Young Elite Scientists Sponsorship Program by CAST (2019QNRC001) and the National Key Research and Development Program of China (2018YFC0407405). The anonymous reviewers and the editor are thanked for providing insightful and detailed reviews that greatly improved the manuscript.

P.S., C.W. and G.D. conducted the research and wrote the paper. J.S., L.K. and M.L. helped to revise the paper. X.L. and H.W. gave the comments. All authors have read and agreed to the published version of the manuscript.

The authors declare no conflict of interest.

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