The impacts of climate changes on watershed streamflow and total dissolved nitrogen in Danjiang Watershed, China

Assessing the variability of watershed hydrochemical processes under possible climate change conditions is crucial for effectively managing water resources and the environment (Shrestha & Wang 2019; Ross & Randhir 2022). Previous studies have shown that the global climate will change significantly in the future under the influence of human activities (Sha et al. 2021a; Orke & Li 2022). The changes reflecting on the watershed scale will result in changes in local precipitation and temperature, which in turn would modify the watershed hydrochemical process and lead to changes in streamflow and non-point source pollution features (Arab Amiri & Gocić 2021). A higher drought or flood risk can be observed in different regions owing to heavier evapotranspiration, earlier and more snowmelt, and extreme precipitation (Aksoy et al. 2021; Amiri & Gocić 2021). In addition, more non-point source pollution loads can be observed in most farmland-dominated watersheds, leading to higher risks of water quality degradation and eutrophication. Quantitative estimation of the impacts of future climate change on watershed hydrochemical processes is of great significance to watershed management as supporting information for effective decision-making applications (Akomeah et al. 2021).

In the past decades, various mechanistic, empirical and semi-empirical, and statistical models have been used as effective estimation tools to assess the impacts of climate change on watershed hydrochemical processes for decision-making support (Burgan & Aksoy 2020; Jakimavičius et al. 2020; Daneshi et al. 2021; Gocić & Arab Amiri 2021; Rashid et al. 2022). By using a suitable watershed model with a particular framework and complexity, watershed hydrochemical processes can be simulated to estimate the yields and source apportionment of streamflow and non-point source pollution. Based on a validated watershed model, the responses of watershed hydrochemical processes under different climate conditions can be modeled by updating the related model weather data for scenario analysis. The sets of climate change scenarios can be subjective assumptions or reanalyzed data based on the results of related sophisticated global climate models (GCMs). Among these, the GCM outputs from the Coupled Model Intercomparison Project (CMIP) are the most widely used (Miralha et al. 2021; Modi et al. 2021; Chen et al. 2022).

CMIP is a standard experimental protocol for climate change scenarios established by the Working Group on Coupled Modelling (WGCM; Eyring et al. 2016a, 2016b). It can be regarded as an international standard of climate change scenarios for scientists to drive different GCMs to assess future climate characteristics in an equivalent way to facilitate further applications. The ensemble of GCM outputs based on CMIP scenarios has been widely used in Intergovernmental Panel on Climate Change (IPCC) assessment reports. Many studies worldwide have used different watershed models to estimate watershed hydrological and/or hydrochemical responses to future climate changes based on GCM outputs under different representative concentration pathway (RCP) scenarios proposed in CMIP5 (Brouziyne et al. 2021; Panondi & Izumi 2021; dos Santos et al. 2022). Recently, the WGCM proposed a new set of emission scenarios driven by different socioeconomic assumptions called shared socioeconomic pathways (SSPs) as the newest CMIP6 scenarios. Many studies have compared the GCM outputs of CMIP5 and CMIP6 and have found better accuracy in new scenarios, particularly for estimating instability and extreme trends (Zhu et al. 2020; Ayugi et al. 2021; Chhetri et al. 2021; Kim et al. 2021). There is a great demand to estimate watershed hydrochemical responses based on state-of-the-art CMIP6 scenarios (Yao et al. 2021). However, owing to the limited update of tools and approaches, related reports are pending.

The main innovation of this study is to propose a new technical approach that enables watershed response estimations for CMIP6 climate change scenarios based on an ensemble of results from multi-GCMs. The generalized watershed loading function (GWLF) model was used to simulate the watershed hydrochemical process. By using historical and synthetic future daily weather data, the GWLF models their respective streamflow and total dissolved nitrogen (TDN) fluxes to assess the impact of climate change by comparing their differences. However, as using a single GCM for response estimation would lead to significant uncertainty, the main challenge in achieving GWLF application is obtaining daily weather data that reflect the ensemble of multi-GCM outputs. Although the original GCM results were on a daily scale, the ensemble of multi-GCMs could only be calculated on a monthly/annual scale as daily weather is random and daily results from different GCMs cannot be directly averaged. Thus, only monthly climate change estimations over a 20-year period were obtained. To address this issue, weather generator models are effective for time-downscaling analysis to bridge the gap between the monthly climate change estimation of multi-GCMs and the daily weather data demand of watershed model inputs. In this study, the Long Ashton Research Station Weather Generator (LARS-WG) model was used as the downscaling tool to generate synthetic weather series for GWLF scenario analysis. However, as the LARS-WG does not have the capability for the latest CMIP6 scenarios, a Python batch procedure was developed for pre-treating GCM outputs to expand the capability of the LARS-WG application in CMIP6 scenarios.

The Danjiang River Watershed (SDRW) in China was used as the study area. We estimated the impact of climate change on watershed streamflow and TDN pollution, and assumed that the behaviors and intensities of water transformation and pollution loading remain constant. The changes in streamflow and TDN fluxes in the watershed outlet were estimated by updating only future temperature and precipitation as inputs to the GWLF. The results showed that there are generally wet and hot trends in the future, and more non-point source pollution of TDN was expected. The variability of watershed hydrochemical processes under emission-controlled climate scenarios would be relatively moderate, indicating the justifiability of positive climate policies for realizing emission peaks and carbon neutrality to obtain environmental benefits and achieve sustainable development.

The watershed and sub-watershed boundaries are delimited with ArcHydroTools based on Digital Elevation Model (DEM) maps, as well as the river lines that are further calibrated based on remote sensing images. There is one Water Quality Monitoring Station and one Hydrological Station in the study watershed, which provided the historical data for watershed model calibration and verification. The historical weather data from the Meteorological Station closest to the study area is used to operate the watershed model and build the weather generator model. The sources of the original data used in this study are summarized in Table 1.

The GWLF model was used to model the watershed hydrochemical process for streamflow and TDN estimation (Haith & Shoemaker 1987). GWLF is a semi-distributed watershed model that can provide useful results with moderate resources for data gathering and calibration (Santos et al. 2020). It has a distributed hydrological framework for runoff estimation based on the Soil Conservation Service Curve Number (SCS-CN). A series of lumped parameter algorithms were used for groundwater estimations based on linear reservoirs and TDN estimations based on the pollution concentration. The regional nutrient management (ReNuMa) model, a derivative of GWLF with a more robust nutrient algorithm, was used in this study. ReNuMa has the same hydrological framework as GWLF, but a NANI framework to calculate the nitrogen concentration of runoff based on human activity to better account for the nitrogen contribution from different land use areas (Hong et al. 2005, 2011). In addition, two additional algorithms based on a previous study were used to refine the modeling accuracy during low-flow periods, which included the segment function approach for the saturated zone and the leakage transport approach for the unsaturated zone (Sha et al. 2014). The GWLF/ReNuMa model performed well in estimating the yields and source apportionments of watershed streamflows and pollution fluxes (Hu et al. 2018; Liu et al. 2018). Based on various scenario analyses, the valid GWLF/ReNuMa model can be used to support best management practices (Qi et al. 2020) and estimate the impacts of possible climate change on watershed hydrochemical processes (Li et al. 2019; Sha et al. 2021b). In this study, we used GWLF to model the hydrochemical processes of SDRW to estimate the monthly/annual streamflow and TDN loads in the present and future.

The observed monthly streamflows at the Ma'jie Hydrological Station from 2009 to 2013 were used to calibrate the transport parameters of the GWLF for streamflow estimation. Observed historical data from 2014, 2015, 2017, and 2018 were used to validate the calibrated parameters. The observed monthly TDNs at Gu'yu'kou Water Quality Station from January 2018 to December 2019 were used to calibrate the nutrient parameters of GWLF, and the reserved water quality data from January 2020 to April 2021 were used to validate the effectiveness of the calibrated nutrient parameters. The nonlinear least squares method was used to calibrate the model parameters using the solver macro add-in procedure. To minimize the sum of the squared errors between the observed and estimated values was set as the calibration target. The Nash–Sutcliffe coefficient (R²_NS), coefficient of determination (r²), and mean relative error (MRE) were used to judge model accuracy (Burgan & Aksoy 2022). The yields and source apportionments of the watershed streamflow and TDN in the current state were estimated based on the validated GWLF model. By updating the GWLF model input weather data, the possible yields and source apportionments of watershed streamflow and TDN under various projected climate change scenarios were estimated. The impacts of climate change on watershed hydrochemical processes were estimated by comparing the current and possible future watershed streamflow and TDN features. The methods for obtaining synthetic future daily weather data for GWLF scenario analyses are described in Sections 2.3 and 2.4.

The weather characteristics of the study area were described based on a series of semi-empirical distributions using the LARS-WG model. The weather conditions were divided into dry and wet days, which indicated precipitation and no precipitation in a single day, respectively. Dry and wet days alternate but have variable lengths. The lengths of the dry and wet days are described in the LARS-WG with a set of semi-empirical distributions for different months. On wet days, precipitation was sampled based on a set of semi-empirical distributions for different months. The daily maximum and minimum temperatures were also estimated separately for dry and wet days. For each month, there were two semi-empirical distributions each for dry and wet days to sample the daily maximum and minimum temperatures. These semi-empirical distributions were determined by a series of model parameters that are normally calibrated based on historical weather records in the study area. In this study, 67 years (1954–2020) of observed daily weather data from the Shang-Zhou Meteorological Station (closest to the study area) were used to calibrate the LARS-WG model parameters to determine the semi-empirical distributions of precipitation and temperature. Subsequently, based on the calibrated model parameters, 67 years of synthetic daily weather data were generated and compared with the observed data to validate the parameters. Consistent with existing research (Bayatvarkeshi et al. 2020; Kavwenje et al. 2021), three statistical tests for eight indicators were used to test whether the observed and synthetic values were derived from the same population (Table 2).

The LARS-WG model was used for downscaling analysis to obtain future synthetic weather time series, which were further used in GWLF for scenario analyses to estimate the changes in streamflow and TDN loads. The LARS-WG model can be considered as a stochastic weather generator. Based on the validated model parameters, it can be used to generate synthetic daily weather data of any length statistically ‘identical’ to the observations to extend the available data scale. Furthermore, by updating the model parameters to alter the related semi-empirical distributions, various scenario analyses concerning future climate change can be implemented. There are many possible strategies to design future climate scenarios, among which the CMIP proposed by the WGCM was employed in this study. The four newest SSPs (SSP1-26, SSP 2-45, SSP 3-70, and SSP 5-85) in CMIP6 were used as climate change scenarios. These are the pathways with different emissions and shared socioeconomic assumptions. SSP1-26 and SSP 2-45 represent a controlled world with positive climate policies, while SSP5-8.5 and SSP3-7.0 represent an uncontrolled world with failed climate policies and severe greenhouse gas emissions. Here, we assess future climate change under each scenario and reflect these changes in the LARS-WG model parameters.

The future increases in precipitation are relatively moderate compared with the changes in temperature. Over time, the precipitation increased slightly. In most scenarios and periods, the increase in precipitation was mainly concentrated in the spring and summer months. In all scenarios and time periods, contrary to the increase in temperature, the precipitation in September decreased significantly. There was no significant difference in the changes in precipitation between emission-controlled and uncontrolled scenarios. In the most recent period (2030s), the decreases in precipitation in autumn and winter (September to December) were relatively significant under uncontrolled scenarios. At the end of this century (2090s), the increase in precipitation in spring was relatively significant under uncontrolled scenarios. In the middle of this century (2050s and 2070s), the changes in precipitation under emission-controlled scenarios were relatively moderate.

The downscaled GCMs’ outputs mentioned above were used to build user-defined scenario files for LARS-WG to generate corresponding synthetic weather data series, which were further used in GWLF to estimate the impacts of climate change on watershed streamflow and TDN, as described in Sections 3.2 and 3.3.

Under different scenarios in different periods in the future, streamflow changes in each month were found to vary significantly. The monthly streamflows increase from May to July and decrease from August to November under most scenarios and future periods. This is because under the general increasing trends of precipitation in the future, the temperature increases from May to July are small, and the increased evapotranspiration is less than the additional increases in precipitation, which manifests as an increase in streamflow. However, the warming caused by future climate change is mainly concentrated from August to November, resulting in significant increases in evapotranspiration which are greater than the increases in precipitation during this period, and is manifested as a decrease in streamflow. The changes in the monthly streamflow from January to March were not significant. On the one hand, this is due to the relatively low variation in temperature and precipitation during this period. On the other hand, as the temperatures during this period are low, the increases in evapotranspiration caused by the increases in temperature were not significant. In summary, a stable trend was reflected in the change in monthly streamflow from January to March.

Changes in future TDN fluxes are influenced by changes in hydrochemical processes owing to climate change. On the one hand, active climate policies need to be implemented to reduce the disturbance of hydrochemical processes by climate change; however, under the assumption that various possible changes in climate are unavoidable, the changes in source apportionments of TDN flux due to changes in hydrochemical processes need to be estimated and targeted control measures should be implemented.

The decreasing trends of the annual streamflow in the study area were generally similar to those in other cases. However, under the best emission-controlled scenario of SSP1-26, other cases showed an increase in annual streamflow (Sunde et al. 2018; Sha et al. 2021a), but the current study area still showed a decreasing trend. This indicates that the study area is more sensitive to climate change. This was probably due to the active evapotranspiration in this area, which is more sensitive to increases in temperature. The greater evapotranspiration caused by the higher temperature in the future would offset the increase in precipitation, leading to a decrease in streamflow, even in the mildest SSP1-26 scenario. The changes in monthly streamflow in the summer months were similar to those in other cases, showing a decreasing trend (Panondi & Izumi 2021). This was mainly due to the stronger evapotranspiration caused by the higher temperatures in summer. However, the decrease in monthly streamflow in the study area in the fall and winter months was still significant, which was quite different from those in other areas (Wang & Kalin 2018). This means that the streamflow in the study area is more sensitive to future climate change in autumn and winter. The increasing trends of the annual TDN flux in this study area were lower than those in other similar areas (Sha et al. 2021a). This was mainly due to the decrease of non-point source pollution loading resulting from the decrease in streamflow. However, the increases in TDN flux in the spring and early summer months in this study area were still remarkable, similar to other studies.

The main limitations of this study are as follows. First, there was potential uncertainty in the synthetic weather series generation of LARS-WG. Historical weather data for model calibration may not be sufficiently long, and potential extreme weather conditions may be ignored. Second, the GCMs used for downscaling and ensembles may be insufficient, resulting in uncertainties. Third, the time step of the GWLF results was limited and the monthly estimations may not be sufficiently detailed. It was difficult to judge the extreme days of high streamflow and/or TDN flux, which is important for risk management. Future research should focus on more detailed weather simulations and watershed hydrochemical estimation. More GCMs outputs were expected for the ensemble and new downscaling technologies, such as image super-resolution of machine learning for more detailed weather data at the watershed scale, were expected. In addition, more detailed watershed hydrochemical models are expected with shorter time steps for extreme status estimations and more scenario analysis for decision-making support.

Based on the agreement between observations and modeling results, the proposed technical method of combined application of the LARS-WG and GWLF models was found to be effective for estimating the impacts of climate change on watershed streamflow and TDN in the CMIP6-SSP scenarios. This can be used as an alternative approach to other similar areas. More GCM outputs and detailed watershed hydrochemical models are required for future applications.

There are generally decreasing trends in annual streamflow under future climate change scenarios. The monthly distributions of annual streamflow will generally increase from May to July and decrease from August to November. Positive climate policies to control greenhouse gas emissions will mitigate changes in streamflow by the end of the 21st century. However, there are generally increasing trends in annual TDN fluxes under future climate change scenarios. The monthly TDN fluxes mainly increase from May to July. The additional TDN flux was mainly contributed by agricultural sources through runoff. To mitigate the increase in TDN fluxes, positive climate policies and targeted management strategies concerning agricultural sources should be implemented.

The main limitation of this study is the potential uncertainty of the results. The limited weather data used in the LARS-WG may lead to the neglect of extreme weather conditions. The limited GCMs outputs used in the downscaling may lead to uncertainty in the ensemble. Limited monthly streamflow and TDN fluxes may prevent detailed short-term extreme estimations for risk management.

Future research is expected to provide more detailed estimations and practical scenario analyses to support decision-making. New GCMs outputs based on CMIP6 should be used in the downscaling of the ensemble to avoid uncertainty. A more detailed modeling of watershed hydrochemical processes is expected to provide short-term extreme status estimations. More detailed scenario analyses are expected to support local management.

The authors would like to acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modelling, coordinated and promoted CMIP6. We thank the climate modeling groups for producing and making available their model output, the Earth System Grid Federation (ESGF) for archiving the data and providing access, and the multiple funding agencies that support CMIP6 and ESGF. We also acknowledge the WorldClim data website (http://www.worldclim.org) for the high spatial resolution global weather and climate data. In addition, the authors would like to acknowledge the China Meteorological Data Service Centre (http://data.cma.cn) for the long-term historical weather data of the study area.

This research was supported by the Open Research Fund Program of State Key Laboratory of Hydro-science and Engineering, Tsinghua University (sklhse-2021-B-07).

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.