Multi-scale flow regimes and driving forces analysis based on different models: a case study of the Wu River basin

Climate change and human activities have altered global hydrological-cycle processes (Dimitriadis et al. 2021). Climate change directly affects the hydrological cycle through the amount of precipitation or evapotranspiration (Remondi et al. 2015). Warmer temperatures will accelerate surface evaporation processes, releasing more vapors into the air and possibly increasing the probability of extreme events (Lau & Kim 2017). Human activities affect runoff by regulating reservoirs and altering water cycle patterns, with land use and land cover change being one of the most direct ways (Yang et al. 2019). As the climate warms and human activities continue to intensify, many rivers have experienced significant decreasing trends in runoff (Wang et al. 2012). These changes threaten global or regional water security.

The driving mechanisms influencing changes in runoff vary considerably across timescales. It is essential to quantify the impact of climate change and human activities on hydrological conditions at different timescales based on different models (Yuan et al. 2016). As the characteristics of runoff changes are important indicators for maintaining the balance between the supply and demand and the ecological stability in the basin, it is of practical importance to analyze the hydrological changes in the basin and their driving mechanisms (Song et al. 2021).

Many scholars have analyzed the hydrological condition qualitatively by analyzing the hydrological changes in the basin (Shahid et al. 2021). Previous studies have focused on three aspects to analyze runoff-driving mechanisms. One is the quantitative attribution of runoff changes based on different analytical formulations of the Budyko hypothesis. For example, Hou et al.’s (2022) attribution analysis of runoff changes in river basins along the Loess Plateau based on the elasticity coefficient method of the Budyko hypothesis found that the contribution of changes in substrate conditions to runoff reached 96.78%. Zheng et al. (2021) found that, based on the Budyko method for attributing runoff changes in the Kuye River basin, catchment surface-characteristic parameters were the most critical factors influencing runoff changes. However, the Budyko model could only quantify the contribution of runoff change drivers on an annual scale and could not quantitatively separate the effects of runoff change drivers on different timescales. The second aspect is the reconstruction of natural runoff based on the data-driven model, and then the causes of runoff changes being analyzed based on the differences between simulated and actual runoff. Fan et al. (2021) analyzed the runoff-driving mechanism of the Poyang Lake basin based on the long short-term memory model (LSTM) and found that the runoff changes were mainly influenced by climate change, with an average influence weight of 75.84%. Wang et al. (2023) proposed a monthly runoff interval prediction method based on whale optimization algorithm (WOA)-VMD-LSTM and nonparametric kernel density estimation, where VMD is variational mode decomposition. Dodangeh et al. (2021) proposed a new data intelligence model developed by mixing the adaptive neuro-fuzzy inference system with gray wolf optimizer, particle swarm optimizer, and WOA for the prediction of the daily river flow of the Talhan River, which is the main source of drinking water for Tehran, the capital of Iran. The success of the WOA model demonstrates its potential to be used for modeling highly nonlinear daily rainfall–runoff relationships. It indicates that the WOA-algorithm-optimized machine learning model coupled with hydrological modeling for runoff prediction has higher accuracy on daily and monthly scales and is more suitable for the analysis of drivers of runoff changes (Liu et al. 2022).

The third aspect is the restoration of the natural runoff series by physical hydrological models, including the soil and water assessment tool (SWAT) model and variable infiltration capacity (VIC) model. Measured and simulated runoffs were compared to quantify the weight of climate change and human activities on the driving mechanisms of runoff change. Chen et al. (2022) reconstructed the unaffected natural runoff series in the river basin by the SWAT model and found human activities to be the primary driver of runoff change by calculation. Yang et al. (2020) used the VIC model to analyze runoff variability in the Wei River basin and found that climate change was the leading cause of runoff reduction.

The principles and data collection methods for natural runoff restoration vary among different models, resulting in limitations and relativity in the attribution analysis of runoff changes. Therefore, more previous research that analyzes the driving factors of runoff changes in a watershed based on a single model at the annual scale is required. In addition, the current research needs more analysis at more detailed timescales (seasonal and monthly). It fails to consider multiple factors such as the seasonal distribution of precipitation and the operation of reservoirs, which are essential drivers of changes in runoff processes (Ning et al. 2022). In summary, this study aims to (1) use a whale-optimization-algorithm-optimized random forest model and a water erosion prediction (WEP-L) model to simulate and restore natural runoff at the annual, monthly, and seasonal scales; (2) analyze the ecological response of runoff using the indicator of hydrologic alteration (IHA) range of variation approach (RVA) method; and (3) employ a new framework that integrates models of different types to comprehensively analyze and quantify the relative contributions of climate change, human activities, and land use at multiple levels and scales, as well as considering the absolute changes at the monthly scale.

As can be seen from Table 1, in this study, we employed daily flow data from the Wulong hydrological station between 1990 and 2019, and daily meteorological information was obtained from 14 different meteorological stations, in addition to land-use data spanning six periods (1990, 1995, 2000, 2005, 2010, and 2015) and digital elevation model (DEM) data. The daily flow data were acquired from the Yangtze River Water Resources Commission (http://www.cjw.gov.cn/), meteorological information was obtained from the National Meteorological Science Data Center (https://data.cma.cn/), land-use data were sourced from the Environmental and Resource Science Data Center of the Chinese Academy of Sciences (https://www.resdc.cn), and the DEM data were obtained from HydroSHEDS (https://www.hydrosheds.org).

The cumulative anomaly method is mainly applied in hydrometeorological research. It verifies long-term trend changes in a sequence by assessing the degree of dispersion of the accumulation anomaly value X_t relative to the mean of the data. In the event of X_t decreasing, it signifies a value lower than the mean of the data, while its increase represents a value higher than the mean. The fluctuation chart of its long-term series can show the growth and decline of the data. Typically, the cumulative anomaly plot exhibits distinctive sections, where X_t increases or decreases, and the turning points where they occur reflect the inflection points of the data trend.

To quantify the variability of hydrologic conditions in rivers, the RVA is based on the hydrologic alteration index (IHA) proposed by Yin et al. (2015). Richter et al. (1996) adopted the RVA method to evaluate the hydrological regimes of rivers impacted by external factors. As shown in Table 2, it is divided into five categories: magnitude, timing, duration, frequency, and variability.

N_oi and N_e are the actual and desired numbers of years that fall within the RVA target threshold after the hydrological indicator alteration. To determine an objective standard for the degree of change in IHA indicators, it was established that values of D_i between 0% and 33% indicated low alteration. In comparison, values between 33% and 67% were considered a medium alteration, and values between 67% and 100% indicated high alteration. The overall degree of hydrologic alteration can be evaluated by averaging the hydrologic alterations of the 33 IHAs, but this does not reflect the weightings of individual indicators. More significant weightings were assigned to larger D_i values to reflect each indicator's weight.

Among them, a D_o value of less than 33% is considered low alteration, between 33% and 67% is considered medium alteration, and 67%–100% is considered high alteration (Zhang et al. 2019).

The random forest model represents a contemporary classification and regression technique of ensemble learning methods. The bootstrapped sample sets for this model are established through the random selection of variables and attributes. The classification regression decision tree is an ensemble learning model of individual learning machines within this model. Nevertheless, a challenge to effectively reshaping natural runoff with this model has been observed, which stems from an incomplete collection of driving factors and relatively significant simulation errors. In this context, the present study presents a solution in the form of a random forest model coupled with whale optimization algorithm, which enables a more accurate reconstruction of natural runoff. The whale optimization algorithm initializes a group of individuals in the feasible solution space. Each whale represents a potential optimal solution for extreme value optimization problems, and its position represents its characteristics. The fitness function calculates the fitness value. The quality of fitness represents the superiority or inferiority of the whale. The whales explore the solution space and update their positions by simulating the behaviors of enclosing prey, hunting, and searching for a game (Lv et al. 2020; Wang et al. 2023). Every time the whale population updates its position, it calculates the fitness value and compares it to the current optimal fitness value (Long et al. 2020). The input factors of the whale optimization algorithm and random forest (WOA-RF) model in this study include precipitation, air temperature, potential evapotranspiration, solar radiation, wind speed, and relative humidity, and the output factor is monthly runoff. The monthly mean flow before the mutation and the monthly mean rainfall, air temperature, potential evapotranspiration, and solar radiation after the mutation are used to restore the monthly mean natural runoff from 1990 to 2019 under the influence of only climate change based on the optimization of manual parameters.

This study developed a distributed hydrological model based on land use and meteorological data in the Wu River basin. Unlike data-driven models, distributed hydrological models provide a more comprehensive and accurate representation of the study area (Martin et al. 2017). To ensure the accuracy of natural runoff simulation, the WEP-L model employed a manual calibration approach for parameter tuning during the restoration process (Xu et al. 2021). The parameters of the WEP-L model mainly include basin storage capacity-related parameters, evaporation-related parameters, water-movement-related parameters, and surface interception-related parameters. The aquifer thickness correction coefficient specifically determines the basin storage capacity-related parameters. The aquifer thickness is the thickness between the ground and the impermeable layer. The suggested range for the aquifer thickness correction coefficient is between 0.1 and 20, and this study used a value of 2.5.

The evapotranspiration parameters mainly affect soil evaporation, vegetation transpiration, and runoff. These parameters include the thicknesses of the first three soil layers and the stomatal resistance correction coefficient. The recommended ranges for these values are as follows: 0.1–0.8, 0.2–2, 0.3–4, and 0–100, and in this study, the values were set to 0.2, 1, 1, and 100. The water-movement-related coefficients include the riverbed bottom material's hydraulic conductivity correction coefficient and the aquifer's lateral hydraulic conductivity correction coefficient. The suggested range for these values is 0.01–100, and in this study, the values were set to 50 and 3, respectively. Regarding surface-interception-related parameters, the main factor is the selection of the intercepted runoff storage capacity in the depression. The recommended values for different land-use types are as follows: forestland: 20 mm, grassland: 12 mm, bare land: 2 mm, sloping cultivated land: 5 mm, paddy fields: 60 mm, irrigated farmland: 50 mm, non-irrigated farmland: 50 mm, dam sites: 50 mm, and terraced fields: 80 mm.

The absolute change in runoff caused by climate change peaked in July at 801.8 m³/s, while the change in runoff in February, March, and December approached zero. As shown in Figure 9(b), the relative contribution rate of runoff is proportional to the absolute contribution amount, and in months with higher total contribution, the relative contribution rate is also relatively large. Among them, the relative contribution of climate change in August and September reached more than 80%, while the relative contribution of climate change in March, April, October, and November was only about 45%. The annual contribution rate of climate change was 81.8%, and the annual contribution rate of human activities was 18.2%. This indicates that climate change is the dominant factor in the changes of runoff in the Wu River basin, and the period with the most significant impact of climate change is the flood season. As shown in Figure 9, the contributions of climate change and human activities to runoff changes at monthly, seasonal, and annual scales have different expressions before and after the abrupt change. At the monthly scale, the relative contribution of climate change is high during the flood season and low during the nonflood seasons. On the seasonal scale, the relative contribution of climate change is considerable in spring and summer but small in autumn and winter. At the annual scale, climate change is the dominant factor of runoff change in the Wu River basin.

Climate change and human activities have jointly changed river runoff conditions, and quantifying the impact of climate change and human activities on river runoff changes is essential for water resources management and river ecological conservation (Duan et al. 2016; Yuan et al. 2018; Singh & Jain 2020). The analysis of the hydrological condition in the Wu basin before and after 2005 based on 32 IHA hydrological indicators revealed an overall change in the hydrology of 46% (moderate change). A noticeable change in the hydrological condition in the Wu River basin before and after the abrupt change provides a basis for subsequent analysis of runoff changes.

The WOA-RF model is based on a comparative analysis of the reconstructed natural runoff and the measured runoff and calculates the contribution values of climate change and human activities at the monthly scale and the impact weights at the annual, seasonal, and monthly scales. At the monthly scale, human activities and climate change increase simultaneously and peak during the flood season, with the contribution of climate change and human activities to runoff changes increasing by 893.1 and 702.8 m³/s, respectively, from May to July (Figure 9). This phenomenon is triggered by reservoir scheduling in the upper and middle reaches of the reservoir complex (Contreras et al. 2021). To prevent flooding, reservoir complexes store water at the end of the flood season to cope with the shortage of water for domestic and productive use during the nonflood season, while rainfall is also concentrated during the flood season, making the interaction between climate change and human activities maximum during the flood season (Sulis 2017), which in turn leads to a simultaneous increase and peak in the contribution of the different drivers. From May to July, the increase in the contribution of human activities to the change in runoff is higher than that of climate change, indicating that human activities during the flood season have intensified, and the degree of their changes is higher than that of climate change.

The WEP-L model is based on meteorological data and land-use data to learn about the meteorological geographic characteristics of a watershed. This study used the WEP-L model to explore the effects of climate change and land-use changes on runoff, using meteorological data in the change period and land-use data in the base period (Marhaento et al. 2017). The WEP-L model weights the impact of climate change on summer runoff at the seasonal scale at only 55.6%. On the one hand, the construction and operation of numerous reservoir clusters after the mutation year expanded human involvement in runoff changes. On the other hand, due to the high peak of natural runoff recovered based on meteorological data, the base flow is biased when calculating the impact weights of different drivers, resulting in a relatively small contribution of climate change during the flood season. At the annual scale, climate change was the main factor (70.3%) causing runoff changes. The WOA-RF model and the WEP-L model simulated runoff for the same watershed to show that WEP-L, due to the extra consideration of land use, had a lower peak runoff. The peak runoff in the simulation period is lower than that in the WOA-RF model. However, the simulation effects and correlations between the two models are generally consistent, indicating that the attribution of runoff changes based on simulation-based runoff calculations is reliable. Guo et al. (2021a) calculated the contribution of climate change and subsurface changes in the Wu River to the reduction of runoff at 73.1% and 26.9%, respectively, based on Buteyko, and the results support the conclusions of this study. The relative contribution of climate change at the annual scale calculated by the WEP-L model is about 10% higher than that of the Buteyko and WOA-RF models, which is because the Buteyko and WOA-RF models do not separate land use from human activities.

This study also has limitations in terms of model uncertainty. Both data-driven and distributed hydrological models require parameter adjustments when simulating runoff, and the magnitude and the size of parameter adjustments are uncertain and need consistent standards. For example, the depression storage coefficient in the WEP-L model did not undergo field surveys and investigations but was only trained based on experience. This resulted in errors in the runoff calculation and simulation uncertainty. However, the tension generated in the modeling process of this study had minimal impact on the model's output, and it reflected that our quantification of the driving factors of runoff changes in the Wu River basin at a delicate timescale is reliable.

In this study, we used IHA indicators to analyze changes in river hydrological conditions. We constructed an integrated framework to quantitatively assess and analyze the causal drivers of runoff changes in the Wu River basin at different timescales based on the strengths of the WOA-RF and WEP-L models to make the results more comprehensive and reliable. This study used an integrated framework to assess the attribution of runoff changes. The results show that the Wu River flow series changed abruptly in 2005, and the hydrological situation of the river changed dramatically after the abrupt change. Seven of the hydrological indicators in the IHA metrics have high variability, 16 have medium variability, and nine have low variability. The combined level of hydrological variability in the watershed is moderately altered. It was also found that at the annual scale, the results of both models showed climate change to be the main driver of runoff alteration. The relative contribution of climate change at the yearly scale was 81.8% and 70.3% for WOA-RF and WEP-L, respectively. The attribution results of the two models at the seasonal scale showed that climate change in summer was the dominant factor affecting runoff change, and its absolute change to runoff change was −1,502.9 and −1,499.9 m³/s for WOA-RF and WEP-L, respectively. Anthropogenic activities significantly impacted runoff changes in autumn, with absolute changes of −675.9 and −839.8 m³/s for WOA-RF and WEP-L, respectively. On the monthly scale, the results of the two models showed that the impact of climate change on runoff was more significant during the flood season than during the nonflood season, and their absolute changes peaked in July at −801.8 and −802.0 m³/s, respectively.

Not required as the study did not involve human or animal.

The authors have consent to participate any offer by the journal.

The authors are giving consent to publish the article in the submitted journal.

H.W.: funding acquisition, project administration, resources, investigation, supervision. S.C.: conceptualization, data curation, formal analysis, investigation, methodology, resources, software, validation, visualization, writing – original draft, and writing – review and editing. N.H.: investigation, formal analysis, methodology, validation, and visualization. L.H.: structure design and method determination. H.Y.: in-depth exploration and interpretation of the paper results. F.H.: data collection and processing. Y.M.: academic exchanges and paper discussions. W.C.: revision and review of papers. W.G.: funding acquisition and project administration.

This study was supported by the National Natural Science Fund of China (51779094); the 2016 Henan University Science and Technology Innovation Talent Support Plan (16HASTIT024); and the Guizhou Provincial Water Resources Department 2020 Water Conservancy Science and Technology Project (KT202008).

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

The authors declare there is no conflict.