Hydrological simulations are a main method of quantifying
the contribution rate (CR) of climate change (CC) and human activities (HAs)
to watershed streamflow changes. However, the uncertainty of hydrological
simulations is rarely considered in current research. To fill this research
gap, based on the Soil and Water Assessment Tool (SWAT) model, in this
study, we propose a new framework to quantify the CR of CC and HAs based on
the posterior histogram distribution of hydrological simulations. In our new
quantitative framework, the uncertainty of hydrological simulations is first
considered to quantify the impact of “equifinality for different
parameters”, which is common in hydrological simulations. The Lancang River
(LR) basin in China, which has been greatly affected by HAs in the past 2 decades, is then selected as the study area. The global gridded monthly
sectoral water use data set (GMSWU), coupled with the dead capacity data of
the large reservoirs within the LR basin and the Budyko hypothesis
framework, is used to compare the calculation result of the novel framework. The results show that (1) the annual streamflow at Yunjinghong
station in the Lancang River basin changed abruptly in 2005, which was mainly due to the construction of the Xiaowan hydropower station that
started in October 2004. The annual streamflow and annual mean temperature
time series from 1961 to 2015 in the LR basin showed significant decreasing and increasing trends at the
The hydrological cycle of the watershed and water resource systems is deeply influenced by climate change (CC) and human activities (HAs) (Bao et al., 2012; Chandesris et al., 2019; Han et al., 2019; Teuling et al., 2019). CC mainly refers to changes in precipitation and evapotranspiration that are caused by rising temperatures and water vapor (Hegerl et al., 2015), while the impact of HAs is mainly reflected in the following aspects: reservoir construction changes the spatial and temporal distribution of streamflow processes (Hennig et al., 2013; Chandesris et al., 2019), land use changes change the characteristics of the underlying surface of the watershed, in turn affecting the streamflow of the watershed (Yang et al., 2017), and population increase leads to an increase in the amount of water used for domestic consumption (Teuling et al., 2019). However, identifying which CC and HAs are the main factors driving the changes in the water cycle of river basins is of great significance for water resource managers to formulate policies for sustainable water resource utilization (Dey and Mishra, 2017; Liu et al., 2017). If CC is the dominant driving factor, then hydrometeorologists need to assess the future trends of meteorological factors, such as precipitation and temperature, to change their water resource management strategies in a timely manner. Conversely, if HAs are the dominant factor, water resource managers should evaluate whether the impact of these HAs exceeds the local water resource carrying capacity and then adjust their related policies (Fu et al., 2004).
Numerous published articles have focused on how to quantify the CR of CC and HAs to the streamflow change in river basins (Liu et al., 2019; Bao et al., 2012; Chandesris et al., 2019; Han et al., 2019; Kong et al., 2016; Xie et al., 2019). In general, the commonly used methods of attribution analysis can be divided into the following three categories: (1) conceptual methods, such as the Budyko framework (Li et al., 2007; Liu et al., 2017), (2) hydrological simulation methods (Liu et al., 2019), and (3) analytical methods, such as the climate elasticity method (Liang et al., 2013). What these three methods have in common is that they all need to first test the annual streamflow sequence through nonstationary testing methods (such as the Mann–Kendall test) and then divide the study period into the natural period (before the break point) and the impacted period (after the break point). The first type of method needs to first calculate the sensitivity of the basin's precipitation and potential evapotranspiration to hydrological variables, and then the hydrological changes caused by CC can be calculated combined with the hydrological sensitivity parameters through the changes in precipitation and potential evapotranspiration in the impacted period and natural period so that the CR of HAs is obtained based on the water balance equation (Li et al., 2007). The second type of method simulates multiple scenarios by changing one impact factor with other fixed factors to evaluate the CR of the changed factor using lumped or distributed hydrological models (Liu et al., 2019). The core of these methods is the modeling of two situations where only one impact factor state has been changed, and the difference between the two simulation results is regarded as the influence of the changed factor. The third type of method is mostly based on numerical calculation, taking the climate elasticity method as an example (Liang et al., 2013): this method introduces the concept of climate elasticity to define the quantitative relationship between changes in streamflow and climatic variables (precipitation, evapotranspiration, etc.), and the CR of HAs to streamflow changes can be obtained by subtracting the CR of climate variables. Among the three types of methods, the second types of methods are the most widely used because it has the following advantages: (1) relatively small data requirements (one only needs to input the meteorological and hydrological data to the hydrological model), (2) relatively simple theoretical assumptions, and (3) quantifying the CR of CC and HAs to streamflow changes at the monthly scale.
Various related published articles are briefly reviewed as follows. Bao et al. (2012) used the variable infiltration capacity (VIC) model to investigate the impacts of CC and HAs on streamflow changes in the Haihe River basin, China, and they concluded that HA accounted for more than 70 % of the decrease in streamflow at Guantai station. Wang et al. (2013) used a two-parameter hydrological model to quantify the contribution of CC and HAs to streamflow changes in three river basins (i.e., Zhanghe, Chaohe, and Hutuo River), and they found that HAs were the dominant factor in streamflow changes. The above literature review shows that these studies all used hydrological simulations with fixed parameter sets to quantify the impact of CC and HAs. As pointed out by Abbaspour et al. (2004) and Zhao et al. (2018a), there is a phenomenon of “equifinality for different parameters” (Beven, 2006) in hydrological calibration and simulation, which also means that we cannot ignore the uncertainty of model parameters in the process of quantifying the CR of CC and HAs to streamflow changes because two sets of parameters with the same performance (with the same Nash–Sutcliffe efficiency coefficient) may lead to very different results; this will further influence the decision-making of water resource managers in making effective and sustainable water resource utilization policies. In the last few decades, great progress has been made in evaluating the uncertainty of hydrological simulations (Abbaspour et al., 2004; Beven and Binley, 1992; Yang et al., 2008; Zhao et al., 2018a; Farsi and Mahjouri, 2019); however, in studies related to quantifying the CR of CC and HAs for streamflow changes, few studies have considered the uncertainty of hydrological simulations (Farsi and Mahjouri, 2019). According to our literature search, Farsi and Mahjouri (2019) first considered the uncertainty of hydrological simulations in the process of quantifying the CR of CC and HAs to streamflow changes, but they only constructed the posterior distribution of the CRs of CC and HAs; in their research, they did not specify how to calculate the CRs of CC and HAs while considering the uncertainty of hydrological simulations. Therefore, to fill this research gap, in this study we propose a new method to quantify the contribution of CC and HAs to streamflow changes considering the uncertainty of hydrological simulations, which in summary is developed using the posterior histogram distribution of hydrological simulations.
The Lancang River (LR) is located in Southwest China and is the largest transboundary river in the Indo-China Peninsula; it is usually called the
Mekong River (MR) after flowing out of China (Grumbine and Xu, 2011).
The abundant water and ecological diversity of the Lancang–Mekong River basin nurtures tens of millions of people in many countries along the Lancang–Mekong River. The upstream flow of the river provides guarantees for irrigation and fishery water use in the countries along the MR during the
dry season, and the water conservancy facilities of the LR during the peak
of the flood period also provide important engineering guarantees for
downstream flood control (Piman et al., 2012, 2016). In the
past 3 decades, a series of hydropower stations have been constructed in the LR basin to meet the flood control and drought relief requirements of downstream countries and the power needs of Southwest China. Therefore, it
is particularly important to quantify the CR of CC and HAs to streamflow
changes in the LR basin. However, so far, there are still few corresponding studies. Han et al. (2019) chose the LR basin as the study area
and then divided the research period into three periods, the natural period,
transition period, and impacted period, and combined them with the
construction time of six large hydropower stations in the LR area. Finally,
they found that the CR of HAs during the impact period exceeded 95 %,
using the coupled routing and excess storage (CREST) model, which was
probably due to the construction of the Nuozhadu hydropower station.
However, there are still areas for improvement in their research: (1) the
results of the hydrological simulation were relatively poor (with monthly
NSE
In this paper, the break point of the change in flow regimes was identified using the Mann–Kendall break point test. Then, the study period was divided into a natural period (before the break point) and an impacted period (after the break point). The Soil and Water Assessment Tool (SWAT) model was used for monthly streamflow simulation at Yunjinghong station. Next, the monthly SWAT model was calibrated and validated using the sequential uncertainty fitting procedure version 2 (SUFI-2) (Abbaspour et al., 2004). Uncertainty analysis was also conducted with the SUFI-2 method, and then the posterior histogram frequency distribution (HFD) of the CR of CC and HAs was obtained. Finally, the proposed quantification framework was compared with two other methods: one was the Budyko framework, and the other was to use the LR basin's gridded monthly sectoral water withdrawals in the period from 1971 to 2010 (Huang et al., 2018) together with the dead reservoir storage capacity data of the six constructed hydropower stations along the mainstream of LR to separate the CR of HAs.
The Lancang River (LR) originates in the northeastern Tanggula Mountains,
Qinghai Province, China, and flows through China's Qinghai Province, Tibet
Autonomous Region and Yunnan Province. It is the largest international river
in Southeast Asia, and it is called the Mekong River after it flows out of
China. Its mainstream has a total length of
Locations of the Lancang River (LR) basin, Yunjinghong hydrological station, constructed dams on the mainstream of the LRB, and main rivers and elevations (m).
Basic information for the six large constructed dams on the mainstream of the LR basin.
(Notation: dead storage capacity refers to the storage capacity below the dead water level of the reservoir, which does not participate into runoff regulation during the normal operation of the reservoir.)
The LR features an arid climate in the upper mountainous areas, while the
lower reaches are dominated by humid climates. The average annual
precipitation of the whole basin is
The China Gauge-based Daily Precipitation Analysis (CGDPA) product was
developed by the China Meteorological Administration (CMA) using data from
The digital elevation model (DEM) used in this study was downloaded from
NASA's Shuttle Radar Topography Mission (SRTM) database at a spatial resolution of
The global gridded monthly sectoral water use (GMSWU) data set for 1971–2010
was obtained from
Hydrological simulation is one of the main methodologies to quantify the CR of CC and HAs to streamflow variations; however, in the past, related studies have rarely considered the uncertainty involved in hydrological simulations (Farsi and Mahjouri, 2019). In this section, we will introduce a new quantitative framework to quantify the influence of the common phenomenon of “equifinality for different parameters” in hydrological simulation on the quantitative results by constructing the posterior distribution of streamflow simulations during the implementation process. The specific implementation flowchart is shown in Fig. 2. The detailed execution steps are shown as follows.
Step 1. Inspection of break points in the annual streamflow sequence; based on the result of the breakpoint test, the entire time series is divided into a
natural period (before the break point) and an impacted period (after the
break point). Step 2. Sensitivity analysis of the parameters in the hydrological model. Step 3. According to the results of the parameter sensitivity analysis, selection of the more sensitive parameters and input of the
hydrometeorological data of the natural period (before the break point) to
calibrate the hydrological model with 1000 runs. Step 4. Selection of the parameter sets with Nash–Sutcliffe efficiency (NSE) coefficients greater than 0.75 in 1000 simulations, input of the hydrometeorological data of the impacted period, and further calculation of
the CR of CC and HAs to the streamflow change corresponding to each
simulation result. Step 5. Construction of the posterior histogram distribution (PHD) of the CR of CC and HAs (with a 5 % step), and then the histogram with the highest
frequency is treated as the uncertainty CR interval of CC and HAs to the
streamflow change. The arithmetic mean of the results in the interval is treated as its
true CR.
In step 4, to ensure the number of streamflow simulation samples, we set the simulation results with NSE to greater than 0.75 to at least 500 times. If the setting is not met, then step 3 is repeated until the cumulative simulation times are greater than 500 times.
Flowchart of the newly proposed quantitative framework.
In this step, the trends and break points of the hydrometeorological data
are detected using the nonparametric Mann–Kendall monotonic trend test (Gilbert, 1987; Kendall, 1975; Mann, 1945) and the Mann–Kendall break
point test (Sneyers, 1991), respectively. The main consideration of using the Mann–Kendall test is that this method assumes no particular distribution
for the tested time series (Song et al., 2019; Xu et al.,
2018). Significance levels of
The Mann–Kendall (MK) monotonic trend test was developed by Mann (1945), Kendall (1975), and Gilbert (1987), which has been widely used to detect the presence of an upward or downward trend of the
hydrometeorological time series, and the advantage of this test is that the
time series does not need to follow a certain distribution (Hamed
and Ramachandra Rao, 1998). This method first tests whether to reject the
null hypothesis (
The break point of the hydrometeorological time series denotes a change from
one stable state to another stable state (Xu et al., 2018). It
occurs when the climate system breaks through a certain threshold. The
Mann–Kendall break point test has been widely used to test break points for hydrometeorological time series, signaling when abrupt changes start
(Sneyers, 1991). This test method is used to determine the break point of
the observed annual streamflow in this study. The defined statistic UF
After identification of the break points in the annual streamflow series, the study period is divided into a “natural period” (before the break point) and an “impacted period” (after the break point) (Wang et al., 2015; Bao et al., 2012). The “natural period” means that there is no significant increase or decrease in streamflow during this period, and it also means that relatively slow CC is the dominant factor and that the impact of HAs is very small during this period. Consequently, the impacted period indicates a significant change in streamflow during this period, mostly due to factors such as the construction of water conservancy engineering facilities, increased water consumption for irrigation, changes in land use, and increased water consumption in cities and towns.
The SWAT model is a semi-distributed, physical process-based hydrological model developed by the Agricultural Research Service of the United States Department of Agriculture (USDA-ARS) (Arnold et al., 1998). The SWAT model first divides the study area into several subbasins based on DEM data, and then each subbasin is further divided into several HRUs (hydrologic response units) based on land use and soil data sets. Then, streamflow generation at the subbasin scale is calculated following the principles of water balance and energy balance after inputting the meteorological data sets. Finally, the total flow of river basin exports is calculated according to the Muskingum method (Tang et al., 2019; Arnold et al., 2012b). We chose to use the SWAT model in this study because numerous published studies have proven that this model has excellent performance in hydrological simulations across the world (Tang et al., 2019; Zhao et al., 2018a, b; Lee et al., 2018).
The calibration of model parameters is executed using the independent
software SWAT-CUP, which was developed by Abbaspour et al. (2007).
This software is freely available and provides five parameter calibration
and uncertainty analysis methods. In this study, the sequential uncertainty
domain parameter fitting version 2 (SUFI-2) algorithm (Abbaspour et al.,
1997, 2004) was used to perform parameter calibration and
uncertainty analysis, because this method has proven to have the advantages
of shorter calculation time, ease of implementation, and ability to set arbitrary objective functions (Zhao et al., 2018a; Tuo et al., 2016; Wu and Chen, 2015). The performance of the SWAT model was evaluated by the
Nash–Sutcliffe efficiency coefficient (NSE) (Nash and Sutcliffe, 1970) and relative error (RE):
In this section, we introduce how to calculate the CR of CC and HAs to streamflow variations and how to construct the posterior histogram distribution (PHD) of the CR to consider the uncertainty of hydrological simulations.
A schematic diagram of the attribution evaluation of streamflow changes is
shown in Fig. 3.
Schematic diagram of the contribution rate (CR) of climate
change (CC) and human activities (HAs) to streamflow change using SWAT
modeling. (Notation:
The hydrological and meteorological data in the natural period are input
into the SWAT model, and using the SUFI-2 method to calibrate the model, a
set of parameters represents the characteristics of the catchment under natural conditions with less impact from HAs. Then, this set of parameters is
brought back into the SWAT model using the meteorological data of the
impacted period. Based on the above simulation results, the CC induced in
streamflow can be calculated as follows:
Before the construction of the PHD of the CR of CC and HAs, the sensitivity
of the parameters of the SWAT model is first conducted. Based on the related
published literature (Zhao et al., 2018a; Yang et al., 2008; Malagò
et al., 2015) and the authors' experience, the Latin hypercube and global sensitivity methods were used to perform the uncertainty analysis
(Abbaspour et al., 2007). The global sensitivity analysis method is
the estimation of the average change in the objective function caused by the
change in each parameter, and all parameters change during the whole
process. A
Twenty-two selected SWAT model parameters in the sensitivity analysis at Yunjinghong station.
(Notation: R_, V_, and A_represent multiplying, replacing, and adding the corresponding parameter values, respectively, in the process of calibrating the parameters.)
Based on the sensitivity analysis results, nine parameters with the highest sensitivity were selected to recalibrate the model with 1000 simulations. According to the recommendations in Tuo et al. (2016) and
Moriasi et al. (2007), the performance of the hydrological simulation
can be divided into four grades based on the NSE values: very good
performance (0.75
In order to evaluate the calculation accuracy of the novel framework
proposed in this study to quantify the CR of CC and HAs to streamflow
changes, the Budyko framework was used first. This framework was developed
by Budyko (1961) and links climate variability to streamflow (
The changes in the catchment streamflow due to CC, which are mainly
characterized by precipitation (
Through the above formulas, we can separate the CR of CC to streamflow variations and further compare it with the calculation results of the new method proposed in this paper.
In addition to the Budyko framework, we also used the GMSWU data introduced
in Sect. 2.2 and the reservoir dead storage capacity data to roughly
separate the CR of HAs from the streamflow changes in the LR basin. The GMSWU data set provides five types of water withdrawals (i.e., irrigation,
livestock, domestic use, mining, and manufacturing) within the period of 1970 to 2010 in the LR basin, and it was generated by downscaling
country-scale estimates of different sectoral water withdrawals from the
Food and Agriculture Organization (FAO) of the United Nations AQUASTAT,
which ensured its good accuracy (Huang et al.,
2018). Here, AQUASTAT refers to the FAO's Global Information System on Water and Agriculture (
The results of the Mann–Kendall break point test for the annual streamflow at Yunjinghong station within the period from 1961 to 2015 are shown in Fig. 4. Since the intersection of the UF and UB curves in Fig. 4 is within the
confidence intervals (of 0.05 and 0.01), the break point of the annual
streamflow in the LR basin occurred in 2005. Combined with the construction of reservoirs in the LR basin, the construction of the Xiaowan hydropower
station started in October 2004 (with total storage capacity
The Mann–Kendall break point testing statistics of the annual streamflow for the LR basin from 1961 to 2015.
Figure 5 shows the MK monotonic trend testing statistics of the annual and
monthly streamflow for the LR basin from 1961 to 2015. A positive
The Mann–Kendall monotonic trend test statistics of the annual and monthly streamflow for the LR basin from 1961 to 2015.
The time series and MK break point test results of the annual areal precipitation and mean temperature for the LR basin from 1961 to 2015 are presented in Fig. 6. In general, changes in the annual precipitation were more complicated than changes in the mean temperature in the LR basin. The precipitation showed a fluctuating trend, while the mean temperature almost showed a continuous rising trend throughout the study period.
Time series and the Mann–Kendall break point test statistics of the annual precipitation
As shown by the time series of the annual precipitation in the LR basin in Fig. 6a, there was a slightly decreasing trend in the LRB during the last
55 years, especially in the past 10 years, but this trend was not
significant according to the result of the MK test at the
The time series of the annual mean temperature in the LR basin presented in Fig. 6b shows that the annual mean temperature in the study area changed
relatively smoothly before 1998. After 1998, the temperature began to rise
significantly and exceeded the significance level of 0.01. The annual mean
temperature in 1963 reached 5.2
The MK monotonic trend test statistics of the annual and monthly
precipitation and mean temperature for the LR basin from 1961 to 2015 are presented in Fig. 7. The annual precipitation in the study area showed an
insignificant decreasing trend (
The Mann–Kendall monotonic trend test statistics of the annual and monthly precipitation and mean temperature for the LR basin from 1961 to 2015.
As described in Sect. 3.4.2, the sensitivity of 22 selected parameters was evaluated using the SWAT-CUP software (Abbaspour et al., 2007;
Abbaspour et al., 1997), and this software integrates the global sensitivity
analysis method and the parameter optimization methods (such as SUFI-2). SWAT-CUP can perform a combined optimization and uncertainty analysis using
a global search procedure and deal with a large number of parameters through
Latin hypercube sampling. The sensitivity evaluation indexes, the
Basin-wide sensitivity ranking calculated from 22 selected parameters using SUFI-2.
(Notation: V_ represents replacing the default value with the given value, R_ represents the relative change (%), and A_ represents adding the given value to the original parameter value.)
As mentioned above, the nine parameters with the highest sensitivity rankings that controlled different stages of the basin's streamflow production and
flow concentration were selected to recalibrate the model using the SUFI-2 method, and the number of simulations was set to 2000. To reduce the
influence of the initial value of the model parameters on the simulation
results, during the model parameter calibration process, 1961 and 1962 were
set as the warming-up period. Table 4 shows the evaluation metrics of the
simulation using the SWAT model at a monthly scale with the largest NS
value. For the calibration period from 1963 to 1990, the NSE and RE were
found to be equal to 0.94 % and
Evaluation metrics, Nash–Sutcliffe efficiency, and relative error of the SWAT model on a monthly scale.
Normalized monthly observed and simulated streamflow at Yunjinghong station for the calibration (from 1963 to 1990) and validation (from 1991 to 2004) periods. The blue histogram shows the monthly precipitation in the LR basin. The normalized streamflow was calculated from the observed and simulated streamflow divided by their average values.
According to the method described in Sect. 3.4.2, simulations with NSEs
greater than 0.75 among the 1000 simulations were selected. Figure 9 shows the
number of simulations with 0.75
Number of simulations with NSEs greater than 0.75 during the calibration (1963–1990), validation (1991–2004), and whole periods (1963–2004).
The 575 simulations with NSEs greater than 0.75 were selected to construct
the posterior histogram frequency distribution (PHD) of the CR of CC and HAs
to streamflow changes in the LR basin. Figure 10 shows the number of simulations of the CC CR in 5 % intervals and their corresponding NS box
plots. In total, 167 out of 575 simulations calculated that the CR of CC in
the LR basin to runoff reduction was 40 %–45 %, and the average NSE was 0.84. Then, 131 and 92 of the simulation results had calculated climate
CRs of 35 %–40 % and 45 %–50 %, respectively. The CR of CC in other
intervals had relatively few simulations. The NSE value of the CR between
70 % and 75 % was largest (NSE
Histogram of the number of simulations of the CR (with 5 % steps) of climate change to streamflow reduction in the LR basin at the annual scale and corresponding Nash–Sutcliffe efficiency box plots.
Table 5 shows the average values of the main hydrological and meteorological
elements and their changes during the natural period and the impacted
period. During the impacted period, compared with the natural period, the
multiyear average streamflow decreased by 396 m
Hydrological and meteorological elements in the natural (1963–2004) and impacted periods (2005–2015) of the LR basin and their changes during the two periods.
The monthly CR of CC and HAs to the changing streamflow at Yunjinghong station was also analyzed using the new framework proposed in this study, and the results are shown in Fig. 11. In general, only June and November had a large CR of CC, which reached 95 %–99.9 % and 70 %–75 %, respectively, while the CR of CC in the other 10 months was relatively small. The trends of the streamflow and the precipitation and mean temperature in the study area shown in Figs. 5 and 7 indicate that the streamflow in June and November showed a decreasing trend (Fig. 5), while the precipitation in June decreased significantly (passing the significance level of 0.05), and the temperature increased significantly (passing the significance level of 0.05) (Fig. 7). This significant decrease in precipitation and the significant increase in temperature were the main reasons for the decrease in the streamflow in June; that is, the decrease in the streamflow in June was mainly caused by CC. The main factors that led to the decrease in the streamflow in November were also the decrease in precipitation and the significant increase in temperature (Fig. 7). From the results of each month, the CR of CC in March and April was the smallest, reaching 10 %–15 %, followed by July (15 %–20 %), May, August, and September (20 %–25 %), October (25 %–30 %), January and February (30 %–35 %), and December (45 %–50 %).
Histogram of the number of simulations of the CR (with a 5 % step) of climate change to streamflow reduction in the LR basin on a monthly scale.
The mean CR of CC and HAs at the monthly scale, which was calculated by averaging the CRs of all simulation results within the highest frequency, is displayed in Fig. 12a, and the monthly precipitation, potential evapotranspiration, and runoff depth during the natural period and the impacted period are shown in Fig. 12b. Overall, the monthly CR was consistent with the annual results, and the CR in a total of 10 months was mainly due to HAs that led to a decrease in the streamflow in the LR basin. It is worth noting that the CR of CC in June reached 96 %. Figure 12b shows that the precipitation in June during the impacted period was significantly reduced compared with the natural period (with a 20.2 mm decrease). At the same time, the increase in potential evapotranspiration in June was also relatively obvious (with a 9.2 mm increase). Figure 12b clearly shows that the streamflow in the LR basin during the impacted period was significantly reduced compared with the natural period in June to October, and the precipitation had little change, except in June. Therefore, we can conclude that the main reason for the decrease in the streamflow in the LR basin was HAs, as shown in Fig. 12a. In this study area, the main cause of the streamflow changes was mainly due to the construction of reservoirs (such as Manwan and Xiaowan), and at the same time, the water storage of these water conservancy facilities during the flood period also provided engineering support for protecting the safety of downstream life and property. Conversely, during the dry season (from January to May), the streamflow in the impacted period showed an increasing trend compared with the natural period, and the increase in runoff during these 5 months was mainly due to HAs (Fig. 12a), which might have been caused by the release of water from the reservoirs during the dry season. For example, in 2016, due to the influence of El Niño, the countries along the lower Mekong River all suffered severe drought. The Chinese government immediately asked the Jinghong Reservoir to release water urgently, which effectively helped downstream countries mitigate a series of possible effects caused by drought and water shortages (D. Li et al., 2017).
CR of CC and human activities to the changing monthly
streamflow at Yunjinghong station
In this subsection, the new proposed framework that considers the uncertainty of hydrological simulations was compared with the Budyko framework, five sections of water withdrawal data from the LR basin, and the equivalent streamflow depth converted from the dead storage capacity of six large hydropower stations.
Table 6 shows the CR of CC and HAs to annual streamflow changes at Yunjinghong station, which was calculated from the Budyko framework. The actual evapotranspiration was calculated from the annual precipitation minus the annual streamflow depth. As shown in Table 6, compared with the natural period, the precipitation and streamflow depth in the impacted period showed a decreasing trend.
CR of climate change (CC) and human activities (HAs) calculated by the Budyko framework.
The precipitation decreased by 25 mm, and the streamflow depth decreased by 86.5 mm. In contrast, the actual evapotranspiration showed an increasing trend, which may be related to the continuous increase in temperature in recent decades. The CR of CC and HAs to streamflow changes accounted for 37.2 % and 62.8 %, respectively, which was basically consistent with the results calculated by the new framework proposed in this study (the difference was 5.4 %).
Figure 13 shows the annual water withdrawals (i.e., domestic, irrigation,
livestock, manufacturing, and mining) in the LR basin during the period from 1970 to 2010 and changes in the installed capacity and dead reservoir storage from 1992 to 2015. In addition to the amount of water use for
irrigation, the other four types of water use withdrawals all showed an
increasing trend from 1970 to 2010, with domestic water consumption
increasing the most (linear slope
Annual water withdrawals of the LR basin during the period from 1970 to 2010. The linear trend lines are indicated by blue (1970–2004), green (2005–2010), and red (1970–2010), and in the last panel, the total dead storage capacity and installed capacity of the LR from 1992 to 2012 are shown.
According to the method introduced in Sect. 3.5, the changes in the
streamflow caused by HAs in the LR basin were separated, which mainly included the five sections of water consumption changes and the same amount
of water depth as the total dead storage capacity of the reservoir. Figure 14
shows the CR of the five types of water withdrawals by HAs and the
construction of the reservoirs to the streamflow changes in the LR basin during the impacted period (from 2005 to 2015) compared with the natural period (from 1961 to 2004). Overall, the CR of HAs to streamflow changes was
59.91 %, while that of CC was 40.09 %. This result was also consistent
with the results calculated in Sect. 4.3.1. Among them, the streamflow
depth caused by the construction of the reservoir was reduced by
CR of domestic, irrigation, livestock, manufacturing, and mining water withdrawals and reservoir construction and CC to the streamflow changes at Yunjinghong station from 1961 to 2015.
In this paper, we proposed a novel framework to quantify the CR of CC and HAs to streamflow changes considering the uncertainty of hydrological simulations. This is because the phenomenon of “equifinality for different parameters” in hydrological simulations greatly affects the quantification results. To preliminarily investigate the impact of model simulation uncertainty of the quantitative results, we selected the two simulation results with the largest NSEs in this study for analysis. The evaluation metrics and CR of CC and HAs are shown in Table 7, which shows that both simulations can simulate the monthly streamflow at Yunjinghong station in the LR basin accurately, and the two simulations have almost the same evaluation performance. However, the attribution analysis obtained from the two hydrological simulations showed completely different results. In the first simulation result, according to the method introduced in Sect. 3.4.1, the streamflow changes in the LR basin were mainly caused by CC, but in the second hydrological simulation, the opposite conclusion was drawn; that is, HAs dominated. These were almost the same hydrological simulation results but with opposite conclusions from the attribution analysis; this was one of the reasons why we must consider the uncertainty of the model parameters in the attribution analysis of CC and HAs using hydrological simulations. The results of Sect. 4.3.1 and related published studies (Han et al., 2019) in the LR basin show that the streamflow changes in the LR basin were mainly caused by HAs.
Results of the CR of climate change and human activities for runoff changes with almost equal model performance (monthly) using the SWAT model.
(Notation: CC and HA represent the climate change and human activities, respectively; NSE and RE represent the Nash–Sutcliffe efficiency coefficient and the relative error, respectively.)
Table 8 shows the values of nine highly sensitive parameters of the two simulation results and the streamflow values simulated by the two
simulations in the natural period and the impacted period. Table 8 and the
calculation methods introduced in Sect. 3.4.1 show that the watershed
streamflow reduction caused by CC calculated by the first and second simulation results was
Values of nine sensitivity parameters with similar simulation results and their simulated streamflow in the natural and impacted periods.
(Notation: NP: “natural period”, IP
In this study, 575 parameter combinations with good simulation results (NSE
greater than 0.75) were selected, with a step size of 5 %: it is proposed to reduce the influence of hydrological modeling uncertainty in the
quantitative results by constructing the posterior histogram distribution of
the CR of CC and HAs to watershed streamflow change. However, it is
undeniable that there are still unreasonable parameter combinations in the
simulation results with high probability (167 times). For the LR basin, it
is almost impossible to obtain the measured values of all nine parameters with high sensitivity (Table 3). Therefore, in order to further explore the
possible influence of unreasonable parameter values on the quantitative
results, we selected two parameters related to snowmelt streamflow (SMTMP and SFTMP) to exclude unreasonable parameter combinations. According to the parameter value ranges recommended by Abbaspour et al. (2007) and
other related references (Arnold et al., 2012a; Yang et al., 2017), in
this study, the reasonable value range of these two parameters is set to
Histogram of the number of simulations of the CR (with 5 % steps) of climate change to streamflow reduction in the LR basin at the annual scale and corresponding Nash–Sutcliffe efficiency box plots after excluding the parameter combinations.
In Sect. 4.4, the water withdrawals of domestic, irrigation, mining,
livestock, and manufacturing, and in addition, dead storage capacity of
constructed reservoirs as well as the impact of HAs, were separated. Then, the impacts of HAs on streamflow changes were separated. However, HAs also
influenced the land use change on rainfall-runoff characteristics. Figure 16
shows the land use in the LR basin in 2015. Grassland was the largest land use in the upper LR basin, while the lower reaches were dominated by forest.
Due to the high-altitude terrain in the upper reaches, unused land and
glaciers were mainly distributed in this area. Table 9 shows the areas of
land use types in the LR basin in 1980, 1990, 2000, 2010, and 2015. In general, the water area of the LR basin showed a significant reduction from 1980 to 1990, which was possibly due to the decrease in the area of glaciers
due to the increase in temperature from 1980 to 1990 (Fig. 6). In contrast,
the water area increased by nearly 38 % from 2010 to 2015, which was
mainly due to the construction of Nuozhadu hydropower station (with a total
storage capacity of 22.7 km
Areas (km
(Nation: permanent glacier in Table 5 is a second-level type which belongs to “Water”.)
The area of farmland in the LR basin showed a decreasing trend during 2000–2010 and 2010–2015, which is also the main reason for the reduction in the irrigation water consumption in the basin, which is consistent with the results shown in Fig. 13. The areas of the cities all showed an increasing trend in the three periods of 1980–2000, 2000–2010, and 2010–2015 (by 33.3 %, 17.1 %, and 33.1 %, respectively), while the other three types of land use/land cover (i.e., forest, grassland, and unused land) did not change significantly in the three periods. In summary, no significant changes were found from 1980 to 2015 in the forest and grassland of the LR basin (accounting for 38.4 % and 47.2 % of the total area, respectively). Although the city area has undergone significant changes, it accounts for a very small total area of the basin (0.17 %). The change in the water area was mainly due to the construction of the reservoirs, so the method used in Sect. 4.3 to separate the contribution of HAs to the reduction in the streamflow in the LR basin used is reasonable.
Land use classification in the LR basin in 2015.
As analyzed above, there was no particularly significant change in the precipitation and potential evapotranspiration from 1961 to 2015 in the LR basin. HAs mainly included the construction of reservoirs, resulting in changes in the streamflow. Attribution analysis results showed that the CR of HAs was 57.6 %, and the corresponding CC was 42.4 %. This result was basically consistent with Han et al. (2019), but the CR of HAs was smaller than that of calculation results. This may be due to the following reasons.
The streamflow data of different time spans were used to obtain different
break points. They used streamflow data from 1980 to 2014 to obtain the
break point in 2008, and this study used data from 1961 to 2015 to identify
the break point in 2005. Different hydrological models were used. They used the coupled routing
and excess storage (CREST) model with an NSE of 0.57, while the SWAT model
used in this study had an NSE of 0.94. Longer series of streamflow data and simulation data were used.
As indicated by D. Li et al. (2017) and Han et al. (2019), as the streamflow data series became longer in the impacted period, the impact of reservoir scheduling on the streamflow changes on an average scale for many years gradually decreased. D. Li et al. (2017) selected Chiang Saen station, which was the nearest station to Yunjinghong station downstream of the LR basin, for their research, and then they divided the streamflow series into three stages, the pre-impact period (1960–1991), the transition period (1992–2009), and the post-impact period (2010–2014). They concluded that the construction of the reservoirs in the LR basin led to a decrease in the streamflow process during the flood period and an increase in the dry period, which was consistent with the results of our study (Sect. 4.3.2). Their results also showed that HAs contributed 61.88 % to the streamflow reduction at Chiang Saen station, which was also close to the results of our study (57.4 %).
A new quantitative framework for calculating the CR of CC and HAs to watershed streamflow variations was proposed in this study, and it was successfully applied to the LR basin with relatively accurate results. From our perspective, this method can effectively quantify the influence of the “equifinality for different parameters” that may exist in the use of hydrological simulation methods to quantify the CR of CC and HAs. At the same time, we also believe that this framework can be applied to other watersheds based on the following aspects. First, in Sect. 4.4, the Budyko framework and sectional water withdrawal data within the basin were used to compare with the new framework. Second, the results of the comparison with published research on the LR basin (Han et al., 2019) also proved that the framework has good accuracy and applicability. Third, in the process of comparing with the new framework, we fully considered the impact of various HAs within the study area, including five types of water withdrawals (i.e., irrigation, livestock, living, mining, and manufacturing) and the impact of reservoir storage and the land use/land cover change. Of course, due to the highly nonlinear relationship between the parameters of the hydrological model, we suggest that readers ensure that the selected simulation results with NSEs greater than 0.75 are large enough when applying the novel framework in other research areas (this study had 500 simulations). It is undeniable that this method still has certain uncertainties and limitations when it is applied to other watersheds. First, if there are multiple break points in the annual streamflow sequence, then, when selecting the unique break point, it is necessary to consider the abrupt change points of the time series of other meteorological elements (precipitation, temperature, etc.) in the basin. At the same time, the impact of strong human activities (reservoir construction, large-scale water transfer project construction, etc.) on the abrupt change of streamflow in the basin should also be considered (Dey and Mishra, 2017). Finally, a unique break point is selected to divide the research time series into a natural period and an impacted period, and then the quantitative framework proposed in this study can be applied. Second, because the SWAT model has good applicability at the Yunjinghong station in the LR basin, it can meet the 500 best simulation requirements set by the framework proposed in this study, but the hydrological model may have different applicability in different research areas. Therefore, the application of this framework in other research areas may have limitations, which needs to be further verified. Third, because this study uses the parameter combinations obtained by the natural period to input the meteorological element data of the impacted period for calculation, this may also bring uncertainty to the calculation results, which is usually called “transferability” (Fu et al., 2018).
Although the new quantitative framework proposed in this study considers the uncertainties in hydrological simulations, the framework is still based on traditional hydrological simulation methods to separate the CR of CC into streamflow change and then to deduce the CR of HAs. Therefore, inevitably, there are still uncertainties in the calculation process. For example, the construction of large-scale reservoirs and changes in land use/land cover (urbanization, etc.) are important factors that alter the climatic state of a local region, specifically in that they change the temporal and spatial distribution characteristics of local regional hydrometeorological elements (Y. Li et al., 2017; Degu et al., 2011). This change in meteorological elements was regarded as part of the impact of CC in this study; however, it was also caused by both HAs (reservoir construction) and CC. On the other hand, there are uncertainties in the division of the natural period and the impacted period in this study, which assumed that the impact of HAs on streamflow changes in the natural period was negligible; however, there were almost no periods within a watershed that were completely unaffected by HAs, and the impact of HAs on streamflow variations in the natural period was ignored in these studies. In this study, there was also a strong disturbance of HAs during the natural period (i.e., reservoir construction: Manwan and Dachaoshan) (Table 1). In addition, our study selected the NSE as the objective function to calibrate the SWAT model, which may also bring uncertainties in the quantitative results. As indicated by Gupta et al. (2009) and Gupta and Kling (2011), using the NSE as an objective function to calibrate a hydrological model may tend to underestimate the peak streamflow. Although the CR in our study was calculated by the average streamflow over multiple years, it still brought a given amount of uncertainty to the quantitative results. Therefore, follow-up research should strengthen the optimization of the objective function and benefit from field investigation of the actual meaning of the parameters. Since the impacts of CC and HAs on the hydrological processes of the watershed are complicated and interconnected, it is still a challenge to completely separate the impacts of CC and HAs on streamflow variations (Xin et al., 2019). Further consideration should be given to quantifying the impact of specific HAs, such as land use change and water withdrawal, and then to separating the impact of CC and HAs on streamflow changes as completely as possible.
In this study, we proposed a new framework that considered the uncertainties of model simulations to quantify the CR of CC and HAs to streamflow changes. This framework was developed based on the posterior histogram frequency distribution (PHD) of the CR of CC and HAs. Then, we selected the LR basin for the case study. Over the past 3 decades, after the construction of the Manwan Reservoir in 1987, six large reservoirs were constructed within the basin before 2014. The streamflow process in the watershed also has significant changes on multiyear average and monthly scales. The Mann–Kendall monotonic trend test and the Mann–Kendall break point test were used to test the trend and identify the break point of the annual streamflow data at Yunjinghong station within the period of 1961 to 2015. Then, the available period was divided into the natural period (before the break point) and the impacted period (after the break point). Afterwards, the SWAT model and the SUFI-2 method were used to construct the posterior histogram distribution (PHD) of the CR of CC and HAs. Finally, the Budyko framework and the basin-wide gridded monthly sectoral water use (GMSWU) data set were used to compare with the newly proposed framework. The main conclusions of this study are as follows.
The new proposed framework can be used to quantify the CR of CC and HAs
in the LR basin which can fully solve the local optimal solution for hydrological simulation parameters in current related studies. The results
of comparison using the Budyko framework and Gridded Monthly Sectoral Water
Use (GMSWU) data set also showed that the new framework has high accuracy
(the error range is within 6 %). The break point of the streamflow sequence during 1961–2015 at
Yunjinghong station was identified in 2005. The streamflow significantly
decreased ( The quantification results calculated using the new proposed framework
showed that, on an annual scale, compared with the natural period of 1961–2004, the CR of CC and HAs (CR of CC and HAs) were 40 %–45 % (with an
average CR of 42.6 %) and 55 %–60 % (with an average CR of 57.4 %),
respectively. The CR of CC and HAs derived from the Budyko framework were
37.2 % and 62.8 %, respectively, and the error between the two
calculation results was 5.4 %. The CR of HAs calculated using the GMSWU
data and the reservoirs dead capacities was 58.0 %, which also proved that
the new proposed framework in this study can be used in the LR basin. Quantitative analysis results on a monthly scale in the LR basin showed that, except for June and November, streamflow changes in other months were
caused by HAs. Further analysis showed that the streamflow in June during
the impacted period decreased by 6.9 mm compared with that in the natural
period, while the precipitation and potential evapotranspiration decreased
and increased by 20.2 and 8.83 mm, respectively; the streamflow decreased
by 5.34 mm in November, while the corresponding precipitation and potential
evapotranspiration changed by
In summary, this study provides a new calculation framework that considers the uncertainty of hydrological simulations to quantify the CR of CC and HAs to streamflow changes. The results of this case study also provide a reference for understanding the dominant factors of streamflow changes in the LR basin and improving water resource management measures for the transboundary Lancang–Mekong River basin. Of course, this new proposed framework also needs to be applied and verified in more research areas. In addition, this framework only considers the dual impacts of CC and HAs. However, in practical applications, water resource decision-makers are more willing to understand the specific impacts of HAs such as irrigation water and land use changes. Therefore, in future research, efforts should be made to expand the framework to quantify the CRs of individual items of CC and HAs.
The observed precipitation, temperature, wind speed, and relative humidity data sets can be obtained by contacting the author via email (sheny@cma.gov.cn, Shen et al., 2014).
XT proposed the research framework and wrote the draft manuscript. GF and CG improved the language of the manuscript. GW provided the data for this study. CL, YL, ZB, SZ, and JJ jointly processed the data and helped build the hydrological model.
The contact author has declared that none of the authors has any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This research was financially sponsored by two research programs in China: (1) the National Key Research and Development Program (no. 2018YFC1508104), (2) the Guangdong-Hong Kong Joint Laboratory for Water Security (no. 2020B1212030005), (3) the National Natural Science Foundation of China (nos. 92047301, 52079079, and 51879163), and (4) the Second Tibetan Plateau Scientific Expedition and Research Program (no. 2019QZKK0203).
This research has been supported by the National Key Research and Development Program of China (grant no. 2018YFC1508104), the Guangdong-Hong Kong Joint Laboratory for Water Security (grant no. 2020B1212030005), and the National Natural Science Foundation of China (grant nos. 92047301, 52079079, and 51879163).
This paper was edited by Fabrizio Fenicia and reviewed by two anonymous referees.