ABSTRACT
Identifying runoff changes and quantifying the impacts of climate change and human activities are important for water resources planning and management in river basins. The impacts of climate change and human activities on runoff can be more accurately assessed through scientific hydrological modeling and data analysis methods. In this study, an integrated assessment framework was established to quantitatively separate the driving mechanisms of runoff at different time scales. The results show that the runoff of Wu River has shown a decreasing trend since 2004, with a change degree of 56%, and the monthly average flow indexes of August and September have changed significantly, both are over 90%. The NSE coefficient of the SWAT simulation effect is above 0.8, we validate the simulation results based on the LSTM model. It was found that climate change was the main factor affecting the runoff of Wu River, with the contribution rates reaching 60 and 57%, respectively. Pearson correlation analysis found that rainfall was the most important factor affecting the runoff. The results of this study are helpful to formulate effective water resources management policies and measures to ensure the sustainable utilization and management of water resources in the Wu River Basin.
HIGHLIGHTS
Twenty-three parameters were selected for rate determination in SWAT-CUP by referring to a large number of literatures.
The SWAT model is used to compensate for the inadequate simulation of physical processes in the LSTM model when reconstructing runoff.
The effects of climate change and human activities on runoff variability at annual, monthly, and seasonal scales were quantified.
INTRODUCTION
As a key link in the hydrological cycle, changes in river runoff not only affect the allocation and utilization of water resources, but also have far-reaching impacts on the ecological environment and human society (Patterson et al. 2013). With the rapid development of human society, the impact of human activities on the natural environment is becoming more and more significant, while global climate change also has a non-negligible impact on runoff (Ran et al. 2024). According to the IPCC report, global temperatures are rising (Hu et al. 2018), global warming is a serious problem that cannot be ignored, resulting in faster evaporation of surface water bodies, allowing more water to enter the atmosphere, directly affecting the formation and distribution of runoff. At least 3,700 large dams (each with a capacity of more than 1 MW) are planned or under construction globally, mainly in countries with emerging economies. These dams are expected to increase current global hydropower capacity by 73% (Zarfl et al. 2015). The construction of the Aswan Dam has caused a sharp reduction in runoff from the Nile River (Goharian et al. 2022). China's rivers are affected by climate change and several rivers are experiencing a decrease in runoff (Huang et al. 2015; Zhai & Tao 2017; Chen et al. 2022). Changes in river runoff affect surface water evaporation and heat transport, which in turn have an impact on climate, an impact that may further exacerbate climate change. These irregular climate changes and frequent human activities are significantly affecting the basin hydrological cycle processes. Therefore, how to accurately quantify and attribute the impacts of human activities and climate change on runoff is of great theoretical significance and practical value for understanding the pattern of runoff change, optimizing water resource allocation and coping with climate change (Huang & Qiu 2022; Li et al. 2022). Understanding the driving mechanisms of runoff can help to better manage water resources. By analyzing the impacts of different driving factors on runoff, future runoff trends can be predicted, providing a scientific basis for water resources planning and management. At the same time, understanding the different driving mechanisms can also help optimize the use and distribution of water resources and improve the sustainable use of water resources.
At this stage, the study of basin runoff response to climate change and human activities has become a hot issue in the study of water resources evolution. Human activities and climate change are considered to be the dominant factors affecting runoff (Yan et al. 2020). Among them, the impacts of human activities on runoff are complex and diverse, significantly altering the natural state of runoff through changes in land use, implementation of water conservancy projects, and development of agriculture (Lv et al. 2023). For example, changes in land use, such as land cover type conversion during urbanization, significantly alter the processes and characteristics of surface runoff. The increase in impervious area as a result of urbanization increases the surface runoff coefficient, which increases runoff volume and flood flow (Ramezani et al. 2023). In addition, the construction and operation of hydraulic measures such as reservoirs, dams, and irrigation canals, also have a significant impact on runoff by regulating the spatial and temporal distribution of runoff and altering the peaks and valleys of runoff (Su et al. 2023). At the same time, global warming has led to changes in precipitation patterns, resulting in increased floods and droughts in some areas, and higher temperatures have accelerated the rate of surface water evaporation. Changes in these climatic elements not only directly affect the generation and distribution of runoff, but also influence the availability of water resources and ecosystem stability through the process of hydrological cycle (Gray et al. 2023).
In order to quantify the contribution of human activities and climate change to runoff, researchers have used a variety of methods and models to analyze it. Previous studies have analyzed runoff driving mechanisms in three main ways (Shahid et al. 2021). One is to quantify the driving mechanisms using empirical formulae such as quantitatively distinguishing the response mechanisms of rainfall and climate change to runoff using models such as Budyko coupled hydrothermal model, cumulative volume slope method, double cumulative curve method, and regression analysis (Wang et al. 2021). When utilizing these empirical equation methods to quantify the driving mechanisms, attention needs to be paid to the applicability and limitations of the models, as well as the quality and completeness of the data. Second, the natural runoff is reconstructed based on the data-driven model, and the causes of runoff changes are analyzed based on the differences between the simulated runoff and the actual runoff. For example, long short-term memory (LSTM) neural network, BP algorithm, GRU and forest-on-the-fly models are used to reconstruct and simulate the intra-year process of watershed runoff (Liu et al. 2022; Zhang et al. 2023; He et al. 2024). Among them, LSTM models are widely used in hydrological forecasting and runoff simulation due to their superiority in handling serial data, as they have the advantages of non-linear forecasting ability, faster convergence, and capturing long-term correlation of time series (Yuan et al. 2018). However, these models often focus on the analysis of time-series data and ignore the influence of the actual subsurface parameters of the watershed, which leads to a certain deviation between the simulation results and the actual situation. Thirdly, physical hydrological models are used to reconstruct the natural runoff history by combining various climatic and subsurface parameters, and the commonly used models include SWAT (soil and water assessment tool), Variable Infiltration Capacity (VIC) Macroscale Hydrologic Model, and WEP hydrological models to quantify the runoff driving mechanism (Jia et al. 2006; Akhter et al. 2022). Among them, the SWAT model is highly regarded for its excellent simulation capability of basin-scale hydrological characteristics. It is widely used in the fields of river simulation, pollutant transport simulation, simulation of changes in sub-basin surface conditions, and simulation of climate change response, and has been proved to be a reliable tool for simulating changes in watershed runoff (Tian et al. 2017). In view of the fact that most of the previous scholars only used a single model for modeling and simulation, the results were not convincing. Therefore, the SWAT model was used to simulate and reconstruct runoff in the basin to quantify the driving mechanism of runoff, LSTM was used to verify the simulated values, and the SWAT model was used to make up for the shortcomings of the LSTM model in simulating the physical process of runoff reconstruction. The accuracy of simulation data and the applicability of the method are guaranteed.
The Wu River is an important tributary in the upper reaches of the Yangtze River (the third longest river in the world) and the largest river in Guizhou Province, China, with rich mineral resources in the basin, which is also a veritable ‘golden waterway’. Due to human activities and climate change, its runoff has declined significantly (Wei et al. 2021), but different scholars have used different models and related methods to quantify the driving mechanisms. However, in quantifying the driving mechanisms, the models and related methods used by different scholars differ in their principles and data collection methods for the recovery of the Wu River Basin (WRB) runoff, and therefore, the analysis of runoff attribution is subject to certain limitations and relativities (Zheng et al. 2021). In addition, the current study needs more in-depth analysis at more detailed time scales (e.g., seasons and months) to reveal the drivers of runoff changes in the basin more comprehensively.
In summary, this study aims to (1) analyze the degree of alteration of hydrological indicators of runoff using the indicators of hydrological alteration (IHA) and the method of variation amplitude (RVA) methods; (2) reconstruct the runoff using the SWAT model and the LSTM model to quantitatively isolate the driving mechanisms of the runoff at different time scales; and (3) analyze the correlation between rainfall and runoff using the Pearson's correlation analysis and the cross-wavelet.
STUDY AREA AND DATA
Study area
Geographic location (a) and land use (b and c) of the Wu River Basin.
Data sources
The day-by-day runoff data from Wulong Hydrological Station in the Wu River Basin used in this study were obtained from the Hydrological Bulletin of the Yangtze River Water Resources Commission (http://www.cjh.com.cn) from 1980 to 2019. Meanwhile, the long-term meteorological data were obtained from the National Meteorological Data Center website (http://data.cma.cn). Daily meteorological data from 13 national meteorological stations within the Wu River Basin between 1980 and 2019 were selected for the study, which included indicators such as maximum, minimum and average air temperature, rainfall, relative humidity, and sunshine hours. The geographic location information of the specific hydrological and meteorological stations is detailed in Supplementary material 1. 500-meter elevation data from the Geospatial Data Cloud (GDC) were used for elevation data, while land use data were obtained from the China Regional Land Cover Database (CRLCD), and soil data were taken from the World Soil Database (HWSD). See supplementary material 1 for details of each hydrographic station.
METHODS
Trend analysis and mutation tests
Using the results of the Mann-Kendall test can help us to determine the presence of statistically significant trends in the annual meteorological and hydrological data of the study area. Also, by performing a reverse sort on the raw time series data and performing the same statistical calculations, it can be used to detect the presence of mutation points. When two curves intersect at the 95% confidence level, it can be inferred that a mutation has occurred at that point in time (Guo et al. 2021).
The change of the curve can be utilized to verify the accuracy of the mutation point. Cumulative distance level is a commonly used method to determine the trend of the data by the curve intuitively (Pei et al. 2022), will be n moments of the cumulative distance level of all calculated, and can be plotted for trend analysis of the cumulative distance level curve. In order to detect the sudden change points in the time series, the sliding t-test is usually used and combined with other methods such as the cumulative distance level method for analysis.
The sliding t-test method determines whether there is a significant difference by comparing the means of two subsequence sample data to determine the location of possible mutation points. The specific calculation method of sliding t is shown in the literature (Du et al. 2019).
Analysis of hydrological conditions
IHA, a method used to assess the characteristics of river flow changes, is based on historical hydrologic data and calculates several indicators related to river flow changes, such as dry period duration, flood flow, and average flow. These indicators can be used to analyze the trend and magnitude of river flow changes, assess the ecological condition of the river, and predict future changes in river flow (Table 1).
IHA parameters and their grouping
Clusters . | IHA indicators . | IHA hydrological indicator parameters . |
---|---|---|
Group I | Monthly traffic sizes (1–12) | Average flow rate, January–December |
Group II | Magnitude and duration of annual extreme flows (13–23) | Average annual 1d, 3d, 7d, 30d, and 90d minimum (high) flow, baseflow index |
Group III | Time of occurrence of annual extreme flows (24–25) | Time of occurrence of annual minimum and maximum values |
Group IV | Frequency and duration of high and low pulses (26–29) | Number of high and low pulses; high and low pulse duration |
Group V | Rate and frequency of change (30–32) | Rate of increase; rate of decrease; number of reversals |
Clusters . | IHA indicators . | IHA hydrological indicator parameters . |
---|---|---|
Group I | Monthly traffic sizes (1–12) | Average flow rate, January–December |
Group II | Magnitude and duration of annual extreme flows (13–23) | Average annual 1d, 3d, 7d, 30d, and 90d minimum (high) flow, baseflow index |
Group III | Time of occurrence of annual extreme flows (24–25) | Time of occurrence of annual minimum and maximum values |
Group IV | Frequency and duration of high and low pulses (26–29) | Number of high and low pulses; high and low pulse duration |
Group V | Rate and frequency of change (30–32) | Rate of increase; rate of decrease; number of reversals |
Range of variability approach (RVA) was proposed by Richter (Richter et al. 1996) in 1996 to give the changing status of hydrological indicators by comparing the actual condition with the expectation.



SWAT model
The SWAT model can use mathematical–physical equations to describe the physical processes such as flow production, catchment, and sand transport in hydrological processes, providing a basis for the future direction of watershed management. The model is a typical semi-distributed hydrological model that can take into account natural factors such as topography, climate, soil, and land use and land cover, making it a strong physical basis (Liu et al. 2019).







Neural network modeling
LSTM is a temporal recurrent neural network (RNN) that evolved from the traditional RNN. LSTM is a special type of RNN with memory structures for learning long-term information (Xiang et al. 2020). When dealing with long-term dependencies, traditional artificial neural networks are unable to associate previous information with the current time step, while LSTM effectively solves this problem by introducing gating mechanisms and cell states, enabling the model to better capture long-term dependency information in sequences.
Correlation analysis methods
The cross-wavelet transform (XWT) was developed on the basis of traditional wavelet analysis and continuous wavelet transform (CWT) to examine the time–frequency spatial relationship between two time series (Wu et al. 2022).
Generally speaking, the greater the absolute value of the correlation coefficient between two variables, the higher the degree of correlation between the two variables (Niu et al. 2021).
Methods of quantitative runoff attribution analysis






RESULTS
Degree of change in hydrological indicators
Degree of change in hydrological indicators of the Wu River. H for high change; M for medium change; L for low change.
Degree of change in hydrological indicators of the Wu River. H for high change; M for medium change; L for low change.
Hydrologic evolution patterns at multiple time scales
Characteristics and differences in runoff changes at different time scales.
Calibration and validation of the SWAT model for the Wu River Basin
Sensitivity analysis helps in identifying the model parameters that have a critical impact on the model output. This in turn helps in calibrating the model by considering only the calibrated sensitive parameters, which can significantly reduce the model run time for good results (Himanshu et al. 2016). Referring to previous studies, we selected 23 parameters (Table 2) for rate calibration. Running the model again after rate-setting the parameters, the simulation effect is significantly better than the case without rate-setting, which further proves the necessity of parameter rate-setting. By running SWAT-CUP, we singled out five coefficients with the strongest correlation. v__ALPHA_BF.gw indicates the coefficient of groundwater influence on surface water, with a value of 0.245240, which implies that groundwater influence on surface runoff is moderate during the simulation; r__CN2.mgt is used for estimating the potential maximum retention, and 0.014232 is a relatively small value that may indicate low groundwater recharge potential for that soil type or management condition; and r__SOL_AWC().sol is used to indicate the effective water content of the soil. The value of 0.276019 is a direct reflection of the availability of water in the soil. A higher AWC value means that more water is available in the soil for plant uptake and utilization, which is beneficial to crop growth and development; v__TIMP.bsn is usually associated with the baseflow recession coefficient (BRC), which is used to characterize the process of groundwater flow recession, and this value of 0.867486 describes the groundwater flow. The value of 0.867486 describes the rate of reduction or recession of groundwater flow; v__CH_K2.rte is a parameter related to channel hydraulic conductance (CHC), and the specific value of the parameter, 88.974030, represents a certain specific channel hydraulic conductivity, which describes the ease with which water can flow in the channel.
SWAT model simulation parameter rate determination results
Coding . | Physical meaning . | Final parameter range . | Parameter value . |
---|---|---|---|
r__CN2.mgt | Runoff curve coefficient | [−2.418037, 2.768037] | 0.014232 |
v__ALPHA_BF.gw | Baseflow partition coefficient | [−0.129182, 0.624182] | 0.245240 |
v__GW_DELAY.gw | Groundwater delay parameter | [−143.93898, 252.238968] | 195.585526 |
v__GWQMN.gw | Shallow groundwater runoff coefficient | [177.922531, 534.577454] | 455.043396 |
v__GW_REVAP.gw | Groundwater evaporation coefficient | [0.102056, 0.266444] | 0.255101 |
v__ESCO.hru | Soil evaporation compensation coefficient | [0.493343, 1.481657] | 0.984535 |
v__CH_N2.rte | Main channel Manning's coefficient | [−0.128748, 0.157248] | −0.120454 |
v__CH_K2.rte | Effective hydraulic conductivity of the main channel | [56.838894, 160.836105] | 88.974030 |
r__SOL_AWC().sol | Effective water content of surface soil | [−3.324999, 0.894999] | 0.276019 |
r__SOL_K().sol | Soil saturated infiltration coefficient | [−0.804859, 3.594859] | −0.070106 |
v__SFTMP.bsn | Snowfall temperature threshold | [−1.632810, 15.132810] | 12.198827 |
v__SMTMP.bsn | Surface snowmelt depth temperature | [−4.633020, 6.133020] | −0.574223 |
v__SMFMX.bsn | Total water storage capacity | [−1.532886, 12.832887] | 12.186427 |
v__SMFMN.bsn | Minimum soil water storage capacity | [−2.582731, 12.482732] | 6.110041 |
v__TIMP.bsn | Rainfall infiltration snowmelt temperature | [0.307835, 0.905115] | 0.867486 |
v__SURLAG.bsn | Surface runoff retardation time | [−6.276535, 13.920784] | 6.104422 |
r__SOL_Z().sol | Soil depth | [2.158459, 7.484040] | 2.866761 |
v__HRU_SLP.Hru | Average slope | [0.465837, 1.399163] | 0.961433 |
v__REVAPMN.gw | Re-evaporation coefficient of shallow groundwater | [72.663216, 410.421753] | 125.691307 |
r__SOL_BD().sol | Soil bulk weight | [0.680043, 3.044957] | 2.536500 |
v__CANMX.hru | Maximum canopy retention | [−9.164934, 63.664932] | −3.557034 |
v__BIOMIX.mgt | Biomass mixing rate | [0.280860, 0.844140] | 0.770350 |
r__SOL_ALB().sol | Soil albedo | [−0.047284, 0.151034] | 0.043744 |
Coding . | Physical meaning . | Final parameter range . | Parameter value . |
---|---|---|---|
r__CN2.mgt | Runoff curve coefficient | [−2.418037, 2.768037] | 0.014232 |
v__ALPHA_BF.gw | Baseflow partition coefficient | [−0.129182, 0.624182] | 0.245240 |
v__GW_DELAY.gw | Groundwater delay parameter | [−143.93898, 252.238968] | 195.585526 |
v__GWQMN.gw | Shallow groundwater runoff coefficient | [177.922531, 534.577454] | 455.043396 |
v__GW_REVAP.gw | Groundwater evaporation coefficient | [0.102056, 0.266444] | 0.255101 |
v__ESCO.hru | Soil evaporation compensation coefficient | [0.493343, 1.481657] | 0.984535 |
v__CH_N2.rte | Main channel Manning's coefficient | [−0.128748, 0.157248] | −0.120454 |
v__CH_K2.rte | Effective hydraulic conductivity of the main channel | [56.838894, 160.836105] | 88.974030 |
r__SOL_AWC().sol | Effective water content of surface soil | [−3.324999, 0.894999] | 0.276019 |
r__SOL_K().sol | Soil saturated infiltration coefficient | [−0.804859, 3.594859] | −0.070106 |
v__SFTMP.bsn | Snowfall temperature threshold | [−1.632810, 15.132810] | 12.198827 |
v__SMTMP.bsn | Surface snowmelt depth temperature | [−4.633020, 6.133020] | −0.574223 |
v__SMFMX.bsn | Total water storage capacity | [−1.532886, 12.832887] | 12.186427 |
v__SMFMN.bsn | Minimum soil water storage capacity | [−2.582731, 12.482732] | 6.110041 |
v__TIMP.bsn | Rainfall infiltration snowmelt temperature | [0.307835, 0.905115] | 0.867486 |
v__SURLAG.bsn | Surface runoff retardation time | [−6.276535, 13.920784] | 6.104422 |
r__SOL_Z().sol | Soil depth | [2.158459, 7.484040] | 2.866761 |
v__HRU_SLP.Hru | Average slope | [0.465837, 1.399163] | 0.961433 |
v__REVAPMN.gw | Re-evaporation coefficient of shallow groundwater | [72.663216, 410.421753] | 125.691307 |
r__SOL_BD().sol | Soil bulk weight | [0.680043, 3.044957] | 2.536500 |
v__CANMX.hru | Maximum canopy retention | [−9.164934, 63.664932] | −3.557034 |
v__BIOMIX.mgt | Biomass mixing rate | [0.280860, 0.844140] | 0.770350 |
r__SOL_ALB().sol | Soil albedo | [−0.047284, 0.151034] | 0.043744 |
Analysis of model run results
Other data such as meteorological data remained unchanged. The simulation was based on land use data of the base period and the impact period, and the SWAT model was run for simulation. As shown in Table 3, the simulation result coefficients R2 and NS of the base period and the impact period were both above 0.8, indicating a good simulation effect. Especially, the R2 coefficients before the mutation were all around 0.90. The simulation effect after the mutation was lower than before, but also reached 0.69 and 0.74, respectively. The operation effect of the SWAT model in the Wu River Basin is ideal, and the monthly average simulated runoff can be used as the basis for the change of attributable discharge.
Statistics of coefficient indicators of simulation effect of the SWAT model under different scenarios
Data type . | R2 . | NS coefficient . | Mutation years ago R2 . | R2 after the mutation year . |
---|---|---|---|---|
Land use data for 1980 | 0.8121 | 0.81 | 0.8962 | 0.6882 |
Land use data for 2020 | 0.8334 | 0.83 | 0.9051 | 0.7441 |
Data type . | R2 . | NS coefficient . | Mutation years ago R2 . | R2 after the mutation year . |
---|---|---|---|---|
Land use data for 1980 | 0.8121 | 0.81 | 0.8962 | 0.6882 |
Land use data for 2020 | 0.8334 | 0.83 | 0.9051 | 0.7441 |
LSTM model validation
The SWAT model has a better performance in hydrological simulation, which can consider the influence of land use, soil type, and other factors on hydrological processes, while the LSTM model is good at processing time series data and capturing long-term dependencies in the time series data. The LSTM model is a kind of deep learning model, which is able to automatically learn long-term dependencies in time series data. Compared with the SWAT model, which is more cumbersome in data acquisition and processing, the LSTM model does not need to specify too many rules and parameters, and can automatically learn runoff patterns and regularities through a large amount of data. The LSTM model primarily relies on historical data to capture the combined impact of various factors, including climate change. However, without additional input and explanation, it may not be able to effectively differentiate between the effects of human activity and natural climate change. Therefore, this study predominantly utilizes the LSTM model to offer robust predictive power based on its observed data, emphasizing overall changes in traffic patterns and potential future trends. Additionally, the SWAT model is employed to provide a mechanistic understanding of how specific human activities and climate factors alter hydrological processes in order to facilitate analysis of the relative contribution of land use change.
Particle Swarm Optimization (PSO) is an optimization algorithm widely used in constrained optimization, dynamic optimization, multi-objective optimization, dynamic multi-objective optimization, and other problems, which has the advantages of less requirements on the optimized function and faster convergence (Yang et al. 2022). Therefore, we combine the PSO algorithm to optimize the LSTM parameters, and we set the initial range of each parameter as follows: calculate the global optimal fitness value, set the initial learning rate, regularization parameter, and the number of hidden layers in the ranges of [0.001–0.01], [0.0001–0.1], and [10–30], and set the dropout to 0.2, and the maximum number of training times to 500. The number of training is 500. Some other hyperparameters such as output dimension are optimized by a grid approach. After PSO optimization, the global optimal fitness value is 0.0517, the optimal learning rate is 0.0047, the optimal regularization parameter is 0.0001, and the optimal number of hidden layers is 15.
Driving mechanism analysis
Comparison of the contribution value and influence weight of each change quantity between SWAT (a–c) and LSTM (d–f) simulation results.
Comparison of the contribution value and influence weight of each change quantity between SWAT (a–c) and LSTM (d–f) simulation results.
The LSTM simulation results show that from the monthly scale (Figure 8(d)), the relative contribution of human activities is higher only in March, April, May, and November, in which March and November have the greatest impacts of human activities, reaching 0.61 and 0.72, respectively, while the rest of the months have a greater relative contribution of climate change, in which the relative contribution of climate change reaches the peak in January and July, reaching 0.81 and 0.71, respectively. From the seasonal scale (Figure 8(e)), except for spring, which is more affected by human activities, all the other three seasons are more affected by climate change, with summer and winter being the most significant, reaching 0.68 and 0.67, respectively. From the annual scale, climate change has a more significant effect, reaching 0.57. By comparing the monthly and seasonal scales, the results of the contribution rate on the monthly scale are similar to the results of the contribution rate on the seasonal scale, and the results of the contribution rate on the seasonal scale are similar to the results of the contribution rate on the monthly scale. In general, climate change has a more significant impact on the runoff change of the Wu River. By looking at Figure 8(f), it can be seen that climate change plays a dominant role in the flood season and is the main factor influencing the reduction of the runoff of the Wu River, while the difference between land use in the flood season and out of season is more significant.
DISCUSSION
Attribution analysis and model comparison
In recent years, under the influence of climate change and human activities, river runoff has decreased to different degrees in different regions (Zhang et al. 2018), and this effect is not caused by a single element in the basin, but by the interaction between multiple elements. In this study, based on 32 hydrological indicators, the hydrological conditions of the Wu River Basin were analyzed before and after 2004, and the overall change was 56%, which is a moderate change. The changes in hydrological conditions in the Wu River Basin provide a basis for the subsequent analysis of runoff changes. The SWAT model is based on the comparative analysis of reconstructed natural runoff and measured runoff, and calculates the contribution values of climate change and human activities on the monthly scale and the influence weights on the yearly, seasonal, and monthly scales. As shown in Figure 8(c) and 8(f), on the monthly scale, human activities and climate change increased synchronously and peaked in the flood season, the contribution of climate change to runoff change increased by 792.53 m3/s in May–June, and the contribution of human activities to runoff change increased by 460.96 m3/s in June–July. This is directly related to the construction of various water conservancy projects in the Wu River since 2004, and the reservoirs and the river water lifting. The construction and operation of the projects have a direct impact on the river runoff. This conclusion is consistent with the results of Wei et al. (2021).
Compared with the traditional RNN, the LSTM model is able to learn and preserve long-term dependencies in time series data more efficiently, and is suitable for dealing with non-linear relationships, long sequence data, and tasks with long-term memory dependencies. By comparing Figure 8(a)–8(c) and Figure 8(d)–8(f), the simulation results of the two models are basically the same. From the LSTM simulation results, in terms of annual scale, the contribution rate of climate change is 0.57, which is slightly lower than that of 0.60 in the SWAT model, and as a whole, the coefficient of determination R2 of the LSTM model is slightly lower than that of the SWAT model. According to the study of Hernandez et al. (2000), the SWAT model reflects well on the rainfall relationship. The results indicate that the SWAT model simulates a stronger correlation between runoff and rainfall and a higher sensitivity to rainfall changes, resulting in better SWAT model simulation results for monthly runoff than the LSTM model. Since the meteorological data in this study include rainfall, the SWAT model simulation results are more significant. On the other hand, the training samples of the LSTM model are small because the simulation scale is monthly. Deep learning is very dependent on the size of the training samples, and too small training samples may not be able to provide more time-series information, resulting in lower simulation accuracy, which may also be the reason why the coefficient of determination R2 of the LSTM model is smaller. The contribution of climate change to runoff reduction in the Wu River calculated based on the Water and Energy Transfer Processes in Large River Basins (WEP-L) model is 70.3% (Wang et al. 2023a), which is about 10% more than the relative contribution of climate change calculated by the SWAT model in this study, due to the fact that the WEP-L model separates the land use from the human activities, which results in a higher relative contribution of climate change. Guo et al. (2021) combined the Mann-Kendall test, IHA, Budyko's hypothesis, climate and subsurface changes in a comprehensive analysis of runoff in the Wu River Basin, and found that the precipitation contribution (P), the potential evapotranspiration (ET0), and the subsurface-to-runoff changes were 61.5, 11.4, and 26.9%, respectively. Climate change, as the dominant factor in runoff reduction, places higher demands on water resources management and climate change adaptation. There is a need to develop more refined and targeted water resources management and climate change adaptation strategies, and to strengthen inter-basin management as well as climate change monitoring and early warning systems in order to address the challenges of climate change on water resources and ecosystems.
This study also has limitations in terms of model uncertainty. The SWAT model requires parameter adjustments in all simulations of runoff. The magnitude and size of the parameter adjustments are uncertain and require consistent criteria. For example, the soil depth coefficients in the SWAT model were not investigated and surveyed in the field, but only trained empirically with the SWAT-CUP software. This leads to errors in runoff calculations and uncertainty in simulations. Meanwhile, the small training samples of the LSTM model do not provide more time-series information, which leads to lower simulation accuracy which will also affect the reliability of the study. During the SWAT model rate determination process, we selected 23 parameters, and more parameters can be added in the future to improve the accuracy of the simulation results. In the future, we can also use more accurate fully distributed hydrologic models for runoff simulation, such as the DHSVM (Distributed Hydrology Soil Vegetation Model) model, which takes into account multiple factors such as topography, soil, vegetation, rainfall, evaporation-transpiration, and groundwater flow, when dealing with hydrologic processes in a watershed, and delineates and simulates watersheds at a detailed spatial scale with complex parameterization schemes and high demand for input data, providing more refined and accurate runoff and hydrologic simulation results.
Climate change analysis
To this end, we should take the following measures: First, strengthen the early warning system and establish an effective rainfall and runoff monitoring network to ensure that rainfall and runoff data can be obtained in real time and accurately. By using modern weather forecast (such as GCM-SSP in CMIP6 framework) and runoff simulation technology (such as SWAT model and VIC model), the change trend of rainfall and runoff can be predicted in advance to provide a scientific basis for flood control and disaster reduction. The combination of global climate models (GCMs) and shared socioeconomic pathways (SSPs) provides powerful tools for understanding and predicting future climate change. GCM-SSP simulations can analyze climate trends such as temperature, precipitation, and extreme weather events under different socioeconomic development paths. This model helps to assess the impacts of different development paths on the climate system and develop adaptation and mitigation strategies. The second is to optimize reservoir scheduling. In view of the situation that rainfall is ahead of runoff, reservoir management departments should conduct reservoir scheduling in advance and rationally arrange reservoir storage and water release plans. Appropriately reduce the water level of the reservoir before rainfall to increase the storage capacity of the reservoir; the water level of the reservoir is gradually increased according to the runoff after rainfall to relieve the flood control pressure in the lower reaches of the Wu River.
Underlayment analysis
Land use shifts in the Wu River (a) and reservoir construction (b).
CONCLUSIONS
In this study, we used an integrated framework to attribute changes in runoff in the Wu River Basin. We used 32 IHA indicators to analyze the hydrological conditions of the Wu River Basin before and after 2004. Based on the SWAT model, we analyzed the runoff changes in different time scales in the Wu River Basin. To ensure the reliability of the simulation results, we also added the LSTM neural network model to analyze the runoff changes in the same time scale in the Wu River. The results show that the flow of the Wu River changed abruptly in 2004, and the hydrological condition of the Wu River Basin changed significantly after the abrupt change: the overall change was 56%, which was a medium degree of change, and 9 IHA indicators had high variability, 13 indicators had medium variability, and 10 indicators had low variability. The results also showed that on the annual scale, the results of both models indicated that climate change was the main driver influencing the changes in the runoff of the Wu River, and the relative contribution of climate change was 60 and 57% for the SWAT and LSTM models, respectively, on the annual scale. The inputs to SWAT models usually include parameters such as land use, rainfall, evaporation, and soil type, and the outputs are hydrologic variables such as soil moisture and runoff. The inputs to LSTM models are time series data, and the outputs depend on the specific task, and may be the predicted values for the next time step or other relevant information. SWAT models are more concerned with the physical factors of hydrologic processes, while LSTM models are more suited for dealing with time series data. Therefore, the characteristics and nature of the data may have an impact on the prediction results of the model. The attribution results of the two models on the seasonal scale showed that climate change was the dominant factor affecting runoff changes in summer, with absolute changes of 5,021.9 and 2,473.19 m3/s, respectively; human activities were the dominant factor affecting runoff changes in winter, with absolute changes of 263.04 and 296.84 m3/s, respectively; and on the monthly scale, the effect of climate change was more significant in flood season than in non-flood season, and its absolute changes peaked in July and May, with 1,976.2 and 1,417.56 m3/s, respectively.
DATA AVAILABILITY STATEMENT
Data cannot be made publicly available; readers should contact the corresponding author for details.
CONFLICT OF INTEREST
The authors declare there is no conflict.