Effects of Climate Change and Human Activities on Runoff in the Upper Reach of Jialing River, China

Abstract

In recent years, the runoff of numerous rivers has experienced substantial changes owing to the dual influences of climate change and human activities. This study focuses on the Lixian hydrological station’s controlled basin, located in the upper reaches of the Jialing River in China. The objective is to assess and quantify the impacts of human activities and climate change on runoff variations. This study analyzed runoff variations from 1960 to 2016 and employed the Soil and Water Assessment Tool (SWAT) model, the long short-term memory (LSTM) model, and eight Budyko framework formulations to assess factors influencing runoff. Additionally, it used the patch-generating land use simulation (PLUS) and SWAT models to simulate future runoff scenarios under various conditions. The results indicate the following. (1) The study area has witnessed a significant decline in runoff (p < 0.01), while potential evapotranspiration shows a significant upward trend (p < 0.01). Precipitation displays a nonsignificant decreasing trend (p > 0.1). An abrupt change point in runoff occurred in 1994, dividing the study period into baseline and change periods. (2) The Budyko results reveal that human activities contributed 50% to 60% to runoff changes. According to the SWAT and LSTM models, the contribution rates of human activities are 63.21% and 52.22%, respectively. Human activities are thus identified as the predominant factor in the decline in runoff. (3) Human activities primarily influence runoff through land cover changes. Conservation measures led to a notable increase in forested areas from 1990 to 2010, representing the most significant change among land types. (4) Future land use scenarios suggest that the highest simulated runoff occurs under a comprehensive development scenario, while the lowest is observed under an ecological conservation scenario. Among the 32 future climate scenarios, runoff increases significantly with a 10% increase in precipitation and decreases substantially with a 15% reduction in precipitation. These findings underscore the significant impact of human activities and climate change on runoff variations in the upper reaches of the Jialing River, highlighting the importance of incorporating both factors in water resource management and planning.

1. Introduction

Runoff is a critical factor influencing hydrological processes, playing a significant role in regional ecological and environmental transformations, as well as impacting economic development [1,2]. In the current era of global change, climate change [3] and human activities [4] are the predominant forces driving runoff dynamics. The evolution of hydrological elements and the distribution of water resources are greatly influenced by these factors [5,6]. Notably, Zhang et al. [7] documented a significant decline in runoff in six of China’s major river basins, attributing this variability to changes in rainfall patterns and land cover, with the contribution rates varying across different basins. Understanding the fluctuations in runoff and their underlying causes is a critical area of research in water resources. This study focuses on the upper Xihanshui River basin, aiming to elucidate the patterns of runoff evolution and the driving mechanisms at play. Such research is crucial for understanding the hydrological conditions of the area and provides valuable insights for the sustainable management of water resources within the basin [8].

In recent years, the attribution analysis of runoff variability has been pursued through diverse methodological approaches, encompassing traditional statistical techniques, hydrological modeling, data-driven hydrological methods, and hydrological sensitivity analysis based on the Budyko framework. Among these methodologies, traditional statistical methods offer a direct and straightforward means to quantify the influence of driving factors on runoff, making them relatively easy to implement [9]. Hydrological modeling, on the other hand, employs quantitative approaches to assess the impact of driving factors on runoff, utilizing hydrological or land surface process models such as the Soil and Water Assessment Tool (SWAT). The SWAT model is noted for its robust physical mechanisms and high accuracy, making it adaptable to a wide range of river basins [10,11]. Data-driven hydrological methods are primarily based on correlation analysis of the data itself, effectively processing time series data through techniques such as multiple linear regression and long short-term memory (LSTM) networks. These methods have demonstrated strong performance in runoff prediction and runoff attribution analysis. The LSTM model, with its unique structure, is capable of capturing long-term dependencies in time series data, thereby enhancing the accuracy of runoff predictions [12]. Meanwhile, the hydrological sensitivity analysis method based on the Budyko framework involves calculating runoff variations to explore the contribution rates of influencing factors in detail [13]. This method requires fewer parameters, offers a streamlined process, and demonstrates commendable accuracy and stability, making it highly suitable for conducting long-term runoff analysis in river basins [14].

Current research endeavors within the Jialing River basin have primarily focused on the characterization and attribution analysis of runoff in the Xihanshui River basin. These studies encompass water resources development, ecological and environmental transformations, development characteristics and their underlying causes, the interplay between climate change and runoff, and their reciprocal responses. Guan [15] conducted an extensive study in the upper Jialing River basin, revealing adverse trends in various hydrological parameters. These trends were attributed to global changes induced by human activities and their consequential effects on the basin’s land cover. Li et al. [16] examined the variability in runoff characteristics at Beibei Station using Spearman correlation analysis and the Mann–Kendall (MK) test, identifying precipitation as the principal cause of reduced runoff in the study area. Wang et al. [17] utilized six Budyko hypothesis equations to scrutinize the factors contributing to the decline in runoff in Jialing River tributaries, pinpointing human activities as the primary factor. Wang et al. [18] investigated the characteristics of runoff evolution in the Xihanshui River basin through cumulative anomaly analysis and hierarchical clustering, underscoring the substantial influence of human activities on basin runoff. Gu et al. [19] employed linear trends and linear regression to explore the evolution of hydro–sediment dynamics in the Xihanshui River basin, highlighting human activities as the predominant influencing factor. Despite these efforts, research on runoff variability within the upper Jialing River basin remains notably limited, especially regarding quantitative analysis. While some studies have attempted to quantify these changes, they often rely on individual methods, potentially introducing uncertainties and biases. To enhance the robustness and credibility of our findings, this study adopts a multimethod approach for runoff attribution analysis, enabling the comparison and cross-validation of outcomes.

Therefore, the primary objectives of this study are as follows:

(1) Historical Runoff Analysis: Utilizing meteorological and hydrological data from the Li County station in the upper reaches of the Jialing River from 1960 to 2016, we employ eight variations of water balance equations based on the Budyko method for comparative analysis. This is combined with the SWAT and LSTM models to conduct a thorough attribution analysis of runoff changes, aiming to uncover the patterns and influencing factors of runoff variations in the upper Jialing River over the past 57 years.

(2) Future Runoff Projections: Developing future climate scenario models using the SWAT model and land use change scenarios with the patch-generating land use simulation (PLUS) model, we aim to further investigate future runoff changes. This will provide a scientific basis for the sustainable development and utilization of water resources in the Jialing River basin. By focusing on the Xihanshui River basin, a critical part of the upper Jialing River, this research provides valuable insights that contribute to a broader understanding of the entire Jialing River basin.

2. Study Area and Data

2.1. Study Area

The Jialing River, a major tributary of the Yangtze River, encompasses a vast basin area of 160,000 km², traversing three provinces (Shaanxi, Gansu, and Sichuan) and one municipality (Chongqing). Ultimately, it converges with the Yangtze River, with a principal channel length of 1345 km and a basin area extending over 39,200 km². The Xihanshui River, located in the upper reaches, is one of the key tributaries contributing to the Jialing River’s flow [20].

The Xihanshui River originates from the western foothills of Mount Qishou in Tianshui City, Gansu Province, China, and has an average annual runoff of approximately 1.6 billion cubic meters [21]. This study focuses on the upper reaches of the Xihanshui River basin, particularly at the Li County hydrological station (coordinates: 105°11′18″E, 34°11′02″N) in Gansu Province, encompassing an area of roughly 3000 km². From 1960 to 2016, the average annual runoff in this region was around 257 million cubic meters, with an average annual precipitation of 541 mm and an average annual evapotranspiration rate of 895 mm. The study area is predominantly composed of grassland and cultivated land. The western part of the region is mainly covered by grassland and forest, while the central and eastern parts are primarily occupied by cultivated land. Built-up areas are mostly situated around river channels or in lowland regions. Water bodies and unused land are relatively scarce in the study area. Figure 1 illustrates the digital elevation model (DEM) of the control area managed by the Li County hydrological station within the Xihanshui River basin.

2.2. Data

In this study, annual runoff data from the Li County hydrological station in the Xihanshui River basin, spanning the years 1960 to 2016, was utilized for Budyko method calculations and runoff correlation analysis. Potential evapotranspiration data were calculated using the Penman–Monteith formula, as recommended by the Food and Agriculture Organization (FAO) in “Crop Evapotranspiration—Guidelines for Computing Crop Water Requirements” [22]. For the calibration process of the Soil and Water Assessment Tool (SWAT) model and the long short-term memory (LSTM) model, daily runoff data from the Li County station, covering the period from 1970 to 1980, were used. Meteorological data—including daily precipitation, average temperature, maximum temperature, minimum temperature, average atmospheric pressure, sunshine duration, daily average relative humidity, and average wind speed—were sourced from seven weather stations located near the study area. These data served as inputs for simulations conducted within both the SWAT model and the patch-generating land use simulation (PLUS) model. To facilitate basin delineation and simulations within the SWAT, LSTM, and PLUS models, digital elevation model (DEM) data specific to the Xihanshui River basin were utilized. Land use data for the years 1990, 2000, 2010, and 2020 within Gansu Province, were integral to model simulations. For the year 2019, natural data (including precipitation and temperature) and socioeconomic data (encompassing population, Gross Domestic Product (GDP), and road data) were incorporated into the PLUS model simulations. To derive average annual precipitation and potential evapotranspiration values for the study area, the inverse distance weighting (IDW) interpolation method was employed. Subsequently, the Budyko equation was applied for calculations and comparative analyses.

3. Methods

The analysis of runoff variations in this study involves the following steps: (1) employing the Mann–Kendall (MK) test and Pettitt test to analyze changes in hydrometeorological elements; (2) assessing contributions to runoff variations using the Budyko model, the long short-term memory (LSTM) model, and the Soil and Water Assessment Tool (SWAT) model; and (3) predicting future runoff variations using the SWAT and patch-generating land use simulation (PLUS) models (Figure 2).

3.1. Trend Analysis and Abrupt Change Point Detection

To analyze the trend of runoff variation within the controlled basin of the Li County hydrological station from 1960 to 2016, we applied the Mann–Kendall (MK) trend test method [23,24]. This method is widely used in hydrology for long-term time series analysis [25]. As a nonparametric approach, it is favored by scholars both domestically and internationally due to its robustness against outliers and its lack of reliance on specific distribution assumptions [26]. In this study, a standardized test statistic Z > 0 signifies an increasing trend in the data, while Z < 0 indicates a decreasing trend. The results are considered statistically significant when |Z| ≥ 1.96 or |Z| ≥ 2.58, corresponding to confidence levels of 95% or 99%, respectively.

To identify abrupt change points in the runoff sequence, we employed the Pettitt abrupt change point test method [27]. This statistical tool is widely recognized in hydrology for investigating the impacts of climate change and human activities on watershed runoff variations. It utilizes nonparametric analysis to detect abrupt shifts in data sequences. The method is valued for its straightforward and precise computation, allowing for the accurate determination of when data changes occur and effectively identifying abrupt change points within the data set.

3.2. Budyko Framework

For a watershed, based on the principle of water balance, the following equation can be obtained:

$$P=E+Q+\Delta S$$

(1)

where P represents the precipitation (mm) in the study area, Q is the runoff depth (mm), E is the actual evapotranspiration (mm), and ΔS is the change in water storage within the watershed, which can be neglected on a long-term scale. The dryness index ($\varnothing =E_0∕P$) as the ratio of actual evapotranspiration to precipitation can reflect the wetness or dryness condition within a watershed. In this research, eight Budyko equations were used (

Table 1. Eight equations based on the Budyko hypothesis.

Type	f(∅)
Budyko-Sch [28]	$f(\varnothing )=1-e^{-\varnothing }$
Budyko-Old [29]	$f(\varnothing )=\varnothing \tanh (1/\varnothing )$
Budyko [30]	$f(\varnothing )={(\varnothing \tanh (1/\varnothing )(1-e^{-\varnothing }))}^{1/2}$
Budyko-PT [31]	$f(\varnothing )=1/\sqrt{1+{\varnothing }^{-2}}$
Budyko-Fu [32]	$f(\varnothing )=1+\varnothing -{(1+{\varnothing }^m)}^{1/m}$
Budyko-Zhang [33]	$f(\varnothing )=(1+\omega \varnothing )/(1+\omega \varnothing +1/\varnothing )$
Budyko-CY [34]	$f(\varnothing )=1/{({(1/\varnothing )}^n+1)}^n$
Budyko-WT [35]	$f(\varnothing )=(1+\varnothing -\sqrt{{(1+\varnothing )}^{\partial }-4\partial (2-\partial )\varnothing })/2\partial (2-\partial )$

), with the latter four being parameter equations that offer higher precision and flexibility compared to the first four nonparameter equations.

The Budyko method posits that the actual evapotranspiration (E) and precipitation (P) ratio in a watershed is determined by the watershed water and heat balance and is expressed as a function of the watershed dryness index (E/P) according to the following equation:

$$E/P=f(E_0/P)=f(\varnothing )$$

(2)

where E₀ represents the multiyear average annual potential evapotranspiration in the watershed (mm), E is the multiyear average annual actual evapotranspiration in the watershed (mm), and P is the multiyear average annual precipitation in the watershed (mm).

Schaake et al. [36] first introduced the concept of the climate elasticity factor, defined as the degree of runoff variation influenced by changes in unit climatic elements. Arora [37] proposed a comprehensive method based on the Budyko method to analyze the long-term impact on runoff variations. Combining Table 1, precipitation (P) and potential evapotranspiration (E₀) are considered as the independent variables at the watershed scale. Therefore, the runoff variation caused by climate change can be expressed as follows:

$$d{\overline{Q}}_c=(\partial f/\partial P)dP+(\partial f/\partial E_0)d{E}_0$$

(3)

$$d{\overline{Q}}_c/Q={\epsilon }_{p}dP/P+{\epsilon }_{E_0}d{E}_0/E_0$$

(4)

$${\epsilon }_p=1+\varnothing f^{\prime }(\varnothing )/(1-f(\varnothing )),{\epsilon }_p+{\epsilon }_{E_0}=1$$

(5)

$$\Delta {\overline{Q}}_c=({\epsilon }_p\Delta P/P+{\epsilon }_{E_0}\Delta E_0/E_0)Q$$

(6)

where ${\epsilon }_P$ represents the elasticity factor coefficient for precipitation in the watershed and ${\epsilon }_{E_0}$ represents the elasticity factor coefficient for evapotranspiration in the watershed. P, Q, and $E_0$ are all known data.

3.3. Runoff Variation Attribution Analysis

To distinguish the impacts of climate change and human activities on runoff variations, using the Pettitt abrupt change point test method to analyze the runoff data, an abrupt change point is identified in 1994. Based on this, the period from 1960 to 1994 is considered the baseline period, and the period from 1995 to 2016 is considered the change period. Therefore, the total runoff variation can be calculated as follows:

$$\left\{\begin{array}{l}\Delta {\overline{Q}}_t=\Delta {\overline{Q}}_a-\Delta {\overline{Q}}_b\\ \Delta {\overline{Q}}_t=\Delta {\overline{Q}}_c+\Delta {\overline{Q}}_h\end{array}\right.$$

(7)

$$\left\{\begin{array}{l}{P}_{\mathrm{c}}=(\Delta {\overline{Q}}_c/\Delta {\overline{Q}}_t)\times 100\%\\ P_h=(\Delta {\overline{Q}}_h/\Delta {\overline{Q}}_t)\times 100\%\end{array}\right.$$

(8)

where $\Delta \overline{Q_t}$ represents the total runoff variation (mm), $\Delta \overline{Q_a}$ is the runoff in the changed period (mm), $\Delta \overline{Q_b}$ is the runoff in the baseline period (mm), $\Delta \overline{Q_c}$ is the runoff variation caused by climate change in the watershed (mm), $\Delta \overline{Q_h}$ represents the runoff variation caused by human activities (mm), P_c represents the contribution of climate change (%), and P_h represents the contribution of human activities (%).

3.4. SWAT Model

The Soil and Water Assessment Tool (SWAT) model is widely recognized as a prominent distributed hydrological model [38]. It leverages elevation data from the study area to delineate the basin and applies minimum drainage thresholds to partition the basin into subbasins, each assigned specific attribute values. Meteorological data are integrated into the SWAT model to construct the basin model, facilitating the simulation of runoff values across the calibration period of this study. Through iterative calibration and parameter fine-tuning using historical data, the SWAT model effectively quantifies the impacts of land use changes and climate variations on runoff. Operating at a daily temporal scale, it adeptly replicates diverse hydrological processes over prolonged durations. To refine model parameters and validate their performance, researchers commonly employ the SWAT Calibration and Uncertainty Programs software (SWAT-CUP 2012). Within SWAT-CUP, the Sequential Uncertainty Fitting algorithm (SUFI_2) is frequently utilized to optimize model parameters [39]. Model performance assessment hinges on metrics such as the Nash–Sutcliffe efficiency coefficient (NSE) and determination coefficient (R²). Higher values of NSE and R², approaching 1, signify enhanced simulation results. Successful calibration and validation of model parameters indicate robust model applicability. Subsequently, the calibrated and validated SWAT model stands ready for accurate runoff simulations in the study area. Following rigorous calibration and validation processes, the refined SWAT model proves instrumental for subsequent runoff simulations, ensuring reliable outcomes essential for hydrological assessments and water resource management strategies.

To determine the contribution rates, we employed well-calibrated hydrological model parameters and database information, making specific adjustments to climate data. The model utilized meteorological data from both the baseline period (1960–1994) and the change period (1995–2016) to simulate runoff. Subsequently, these simulated runoff values were compared against observed runoff data. The contribution rates were computed using the following expressions:

$$\Delta R={\overline{R}}_{obse}^2-{\overline{R}}_{simu}^1$$

(9)

$$\left\{\begin{array}{l}\Delta R_c={\overline{R}}_{simu}^2-{\overline{R}}_{simu}^1\\ \Delta R_h={\overline{R}}_{obse}^2-{\overline{R}}_{simu}^2\end{array}\right.$$

(10)

$$\left\{\begin{array}{l}{\eta }_h=\Delta R_h/\Delta R\times 100\%\\ {\eta }_c=\Delta R_c/\Delta R\times 100\%\end{array}\right.$$

(11)

where ∆R, ∆R_c, and ∆R_h represent the differences between observed runoff in the change period and simulated runoff in the baseline period, the runoff change due to climate variations, and the runoff change due to human activities, respectively; ${\overline{R}}_{simu}^1$, ${\overline{R}}_{simu}^2$, and ${\overline{R}}_{obse}^2$ represent the baseline period’s multiyear simulated average runoff, change period’s multiyear simulated average runoff, and change period’s multiyear observed average runoff, respectively; and η_h and η_c are the contributions of human activities and climate variations to the runoff changes, respectively.

3.5. PLUS Model

The PLUS model represents an advanced, patch-level land use prediction framework based on the principles of cellular automaton (CA). It employs multiple classes of random patch seeds and integrates an analysis module for land expansion strategies alongside a simulation model for future land use changes. Known for its efficacy, the PLUS model excels in forecasting upcoming trends in land use [40].

Utilizing the capabilities of the PLUS model and considering the specific characteristics of the study area, a comprehensive selection of 15 influential factors was carefully chosen as inputs. These factors, along with the land use expansion map and existing data for the year 2010, were methodically integrated into the PLUS model. The simulation process entailed configuring various parameters, particularly emphasizing the determination of neighborhood weights and meticulous adjustment of transferability matrix parameters. The transferability matrix is crucial because it determines the feasibility of transitions between different land types. A matrix value of 1 indicates a viable transition, whereas 0 signifies an infeasible one. To accurately reflect real-world land use dynamics within the study area, a transition matrix was meticulously defined (

Table 2. Land use transfer rule matrix under natural development scenario.

Land Use Types	Cultivated Land	Forest	Grassland	Water Area	Built-Up Land	Unused Land
Cultivated land	1	1	1	1	1	1
Forest	1	1	1	1	1	1
Grassland	1	1	1	1	1	1
Water area	1	1	1	1	1	0
Built-up land	1	1	1	1	1	1
Unused land	1	1	1	0	1	1

).

The model requires the definition of neighborhood weights customized to the specific characteristics of the study area. These weights are assigned values ranging from 0 to 1, and their detailed configurations are outlined in

Table 3. Neighborhood weights.

Land Use Types	Cultivated Land	Forest	Grassland	Water Area	Built-Up Land	Unused Land
Neighborhood weights	0.4915	0.0822	0.3862	0.0033	0.0345	0.0022

.

In this study, the 2010 land use data served as the baseline for simulating land use patterns in 2020 using the PLUS model. The simulated results were subsequently validated against real 2020 data, achieving a Kappa coefficient of 0.73, indicating high reliability [41]. To meet diverse developmental needs and prioritize objectives, adjustments were made to the transfer matrix settings based on the 2020 land use data. These adjustments aimed to forecast land use distribution scenarios for the study area by 2030, encompassing natural changes, ecological conservation, and farmland protection.

3.6. The Long Short-Term Memory Model

So far, when it comes to runoff research, long short-term memory (LSTM), as a powerful self-based data sequence model, shows its superiority in time series analysis. Compared with traditional hydrological statistical methods, LSTM is more flexible and adaptable. It can automatically learn and extract the hidden rules in the data, and thus performs well in dealing with nonlinear and dynamic hydrological processes [42]. In this paper, first, the relevant parameters of the LSTM model are adjusted, and the simulation accuracy in the study area reaches a high degree. Then, the LSTM with the best simulation effect is used to interpolate the missing runoff data and simulate the related runoff process in the study area, and it is applied to the subsequent contribution rate calculation.

In this study, the LSTM model was used to quantitatively analyze the runoff changes in the study area to accurately determine the contribution rate of climate change and human activities to runoff. Three scenarios are set, where P1 represents the monthly runoff process in the base period, P2 represents the monthly runoff process in the change period simulated by the LSTM model after parameter adjustment, and P3 is the measured monthly runoff process in the change period. By comparing the runoff data of P2 and P1, the contribution of climate change to the runoff process in the study area can be accurately determined. By comparing P3 and P2, the contribution of human activities to runoff change in the study area can be accurately evaluated. The following formula can be used to explain the average runoff variation and corresponding contribution rate caused by climate change and human activities in the study area [43]:

$$\left\{\begin{array}{l}\Delta Q_C=Q_2-Q_1\\ \Delta Q_H=Q_3-Q_2\end{array}\right.$$

(12)

where Q₀, Q₁, and Q₂ represent the multiyear average monthly flow process in the study area under P1, P2, and P3 scenarios, respectively; and ∆Q_C and ∆Q_H represent the absolute components of runoff changes caused by climate change and human activities, respectively.

The calculation method of the influence degree of climate change and human activities on runoff change in the study area is as follows:

$$\left\{\begin{array}{l}{\eta }_C=\frac{\left|\Delta Q_C\right|}{\left|\Delta Q_C\right| + \left|\Delta Q_H\right|}\times 100\%\\ {\eta }_H=\frac{\left|\Delta Q_H\right|}{\left|\Delta Q_C\right| + \left|\Delta Q_H\right|}\times 100\%\end{array}\right.$$

(13)

where η_C and η_H represent the relative impact of climate change and human activities on runoff change, respectively.

4. Results

4.1. Analysis of Climate and Runoff Trends

Based on Figure 3, the annual runoff series for the Xihanshui River at the Li County station from 1960 to 2016 exhibits a clear and consistent decreasing trend. The Mann–Kendall (MK) test yielded a calculated statistic of Z = −4.0133, indicating a statistically significant result that exceeds the 0.01 significance level (|Z| > 2.32). Furthermore, the Pettitt abrupt change point test applied to the annual runoff data, as illustrated in Figure 4, identifies a significant decrease in runoff starting around 1980. This analysis conclusively identifies 1994 as the year of an abrupt change, a finding also statistically significant at the 0.01 level. Therefore, 1994 is determined as the abrupt change point. Consequently, the study period logically divides into two distinct segments as follows: a baseline period (1960–1994), characterized by minimal human activities, and a change period (1995–2016), marked by a noticeable increase in human activities.

According to

Table 4. Watershed runoff depth, precipitation, evapotranspiration, and dryness index in different periods.

	Q (mm)	P (mm)	E0 (mm)	Dryness Index (∅)
Entire period (1960–2016)	74.64	540.53	894.80	1.72
Baseline period (1960–1994)	98.10	559.42	867.95	1.61
Change period (1995–2016)	37.30	510.48	937.57	1.90
Variation magnitude	−60.80	−48.95	69.61	0.29

, the average annual runoff depth during the baseline period and the change period is 98.10 mm and 37.30 mm, respectively, representing a significant decrease of 60.80 mm in runoff depth during the change period. Additionally, the baseline period exhibits a dryness index of 1.61, while the change period shows a dryness index of 1.90, indicating an increase in drought conditions during the change period and relative wetter conditions in the baseline period. Figure 5 illustrates the annual potential evapotranspiration, which demonstrates a fluctuating positive trend with Z = 3.558, exceeding the significance threshold (|Z| > 2.32) at the 0.01 level. This indicates that regional potential evapotranspiration has a negative impact on runoff variations. Conversely, the annual precipitation trend (Figure 5) displays a fluctuating negative trend with Z = −1.1909, where |Z| < 2.32, suggesting that the decreasing trend is not statistically significant. This indicates a positive influence of regional precipitation on runoff variations.

4.2. Runoff Sensitivity

The runoff elasticity coefficient indicates how changes in annual precipitation and evapotranspiration within the study area correlate with changes in runoff. These coefficients can either be positive or negative, signifying whether they have a positive or negative effect on runoff, respectively. During the change period relative to the baseline period, the region experiences increased aridity.

Table 5. Precipitation elasticity coefficient (ε_p) value in the study area.

Period	Budyko-Sch	Budyko-Old	Budyko	Budyko-PT	Budyko-Fu	Budyko-Zhang	Budyko-CY	Budyko-WT
Entire period (1960–2016)	2.72	2.74	2.74	2.58	2.82	2.54	2.88	2.48
Baseline period (1960–1994)	2.61	2.71	2.66	2.54	2.53	2.41	2.57	2.35
Change period (1995–2016)	2.90	2.78	2.87	2.65	3.29	2.74	3.37	2.68

clearly demonstrates that precipitation has a significant positive impact on runoff, with impact coefficients exceeding 2. Conversely,

Table 6. Evapotranspiration elasticity coefficient Evapotranspiration elasticity coefficient (${\mathsf{\epsilon }}_{E_0}$) value in the study area.

Period	Budyko-Sch	Budyko-Old	Budyko	Budyko-PT	Budyko-Fu	Budyko-Zhang	Budyko-CY	Budyko-WT
Entire period (1960–2016)	−1.72	−1.74	−1.74	−1.58	−1.82	−1.54	−1.88	−1.48
Baseline period (1960–1994)	−1.61	−1.71	−1.66	−1.54	−1.53	−1.41	−1.57	−1.35
Change period (1995–2016)	−1.90	−1.78	−1.87	−1.65	−2.29	−1.74	−2.37	−1.68

highlights that evapotranspiration adversely affects runoff, with impact coefficients ranging between 1 and 2. In summary, among the climate factors analyzed, precipitation has a more pronounced influence on runoff compared to evapotranspiration, making it the primary contributing factor.

4.3. Runoff Attribution

As shown in

Table 7. Factors and their contribution rates in runoff variation derived from eight equations under Budyko hypotheses.

	Budyko-Sch	Budyko-Old	Budyko	Budyko-PT	Budyko-Fu	Budyko-Zhang	Budyko-CY	Budyko-WT
ΔQc (mm)	−28.38	−28.59	−28.59	−26.65	−29.65	−26.07	−30.39	−25.36
ΔQh (mm)	−32.41	−32.21	−32.20	−34.15	−31.14	−34.72	−30.40	−35.44
Pc (%)	46.69	47.02	47.03	43.84	48.77	42.88	49.99	41.71
Ph (%)	53.31	52.98	52.97	56.16	51.23	57.12	50.01	58.29

, the overall trend in runoff exhibits a significant decrease. Notably, the reduction in runoff attributed to human activities, as analyzed through eight distinct methods, exceeds that attributed to climate factors. Despite variations among the eight equations, all consistently indicate that human activities are responsible for more than 50% of the runoff reduction, underscoring their dominant role. In contrast, climate factors collectively contribute less than 50%, positioning them as secondary factors. This highlights that the primary cause for the decline in runoff within the Li County hydrological station’s control area in the Xihanshui River from 1960 to 2016 can predominantly be attributed to human activities.

In summary, human activities emerge as the primary driver behind the observed decrease in runoff within the study area, with precipitation playing a secondary role, while evapotranspiration exerts the least influence. Human activities primarily affect runoff through alterations in land surface characteristics. It is noteworthy that the study area is affected by the “Natural Forest Protection Project,” an initiative aimed at conserving and restoring regional vegetation. This conservation effort has yielded positive outcomes since its inception. The increase in vegetation cover consequentially reduces evapotranspiration and runoff, contributing to the overall decrease in runoff volumes. Additionally, the impact of certain water conservancy projects cannot be overlooked because they can disrupt natural runoff generation and flow patterns, leading to decreased runoff volumes.

4.4. Land Use Change

The analysis of land use patterns across three distinct time periods (Figure 6 and

Table 8. Area and changes of different land use types in various years (km²).

Land Use Types	Area
	1990	2000	2010	Change from 1990 to 2000	Change from 2000 to 2010
Grassland	1223.24	1218.30	1154.73	−4.94	−63.57
Cultivated land	1597.19	1595.86	1579.91	−1.33	−15.95
Built-up land	86.29	92.00	100.08	5.71	8.08
Forest	244.12	244.13	312.54	0.01	68.41
Water area	10.95	11.51	11.45	1.56	−0.06
Unused land	0	0	3.08	0	3.08

) highlights that the study area is primarily dominated by grassland and cultivated land. Both grassland and cultivated areas have shown varying degrees of change, with notable reductions observed. The most significant transformation occurred in the forested area, which experienced relatively minor shifts between 1990 and 2000. However, from 2000 to 2010, there was a substantial increase in forested area. This significant expansion of forest cover and vegetation over the three decades can largely be attributed to diligent efforts in implementing forest conservation and protection projects.

During the period from 1990 to 2000, significant reductions were observed in the extents of both grassland and cultivated land, accompanied by expansions in areas designated as built-up land, forest, and water features. Particularly, notable shifts during this decade were observed in built-up land (

Table 9. Land use transition matrix for the period 1990–2000 in the study area (km²).

Land Use Types		1990					Total
		Grassland	Cultivated Land	Built-Up Land	Forest	Water Area
2000	Grassland	1218.22	0.08	0	0.01	0	1218.30
	Cultivated land	4.72	1591.14	0	0	0	1595.86
	Built-up land	0.19	5.51	86.29	0	0	92.00
	Forest	0.01	0.01	0	244.11	0	244.13
	Water area	0.10	0.46	0	0.00	10.95	11.51
Total		1223.24	1597.19	86.29	244.12	10.95	3161.80

).

During the period from 2000 to 2010, there were reductions observed in the areas allocated to grassland and cultivated land, along with a slight decrease in water features. Conversely, built-up land and forest areas continued to expand. The forested area, in particular, showed a significant increase of 68.41 km², largely due to the conversion of grassland and cultivated land (

Table 10. Land use transition matrix for the period 2000–2010 in the study area (km²).

Land Use Types		2000					Total
		Grassland	Cultivated Land	Built-Up Land	Forest	Water Area
2010	Grassland	1122.25	30.52	0.38	1.32	0.26	1154.73
	Cultivated land	35.61	1541.61	1.77	0.86	0.06	1579.91
	Built-up land	1.13	9.05	89.84	0.01	0.04	100.08
	Forest	57.14	13.58	0.01	241.81	0	312.54
	Water area	0.08	0.23	0	0	11.15	11.45
	Unused land	2.09	0.87	0	0.13	0	3.08
Total		1218.30	1595.86	92.00	244.13	11.51	3161.80

).

Upon reviewing Figure 7, it is apparent that the predominant land categories within the study area consist of grassland, cultivated land, and forest. In the land use change map spanning the period from 1990 to 2000, minor alterations were observed, primarily involving the expansion of built-up land and limited conversion of cultivated land into grassland. During the subsequent decade, from 2000 to 2010, more significant changes occurred compared to the previous period. In the eastern section of the study area, a substantial area of grassland transitioned into forest, while within the central sector, some cultivated land was converted into forested areas. This dynamic reflects the impact of national forest conservation and afforestation policies, leading to increased forest cover and vegetation resurgence. These changes underscore the profound influence of human activities on the terrestrial landscape.

4.5. SWAT Model Results

The SWAT model was implemented by integrating comprehensive databases encompassing land use, soil characteristics, and meteorological parameters. Initially, the study basin was partitioned into 11 subbasins based on elevation data, with a minimum drainage area set at 62.26 km². Subsequently, this division was further refined into 239 hydrological response units (HRUs), enabling a detailed and granular analysis of the study area. Daily meteorological station data were subsequently incorporated into the SWAT model to finalize the basin model construction.

The research design incorporated distinct time periods for model warm-up, calibration, and validation. Specifically, the years 1970 and 1971 were designated as the warm-up period, while the interval from 1972 to 1976 served as the calibration phase. The span from 1977 to 1980 was then assigned to validation. Land use data from 1980 were selected to represent the baseline period’s land use conditions. Observed monthly runoff data from 1970 to 1980 were utilized for parameter sensitivity analysis, model calibration, and validation. To facilitate this process, the SWAT-CUP software, implementing the SUFI_2 algorithm, was employed for both calibration and validation. Model parameters were meticulously assessed for sensitivity, considering the t-statistic and p-value. Subsequent enhancements in model accuracy were achieved through iterative parameter adjustments in the database.

The final selection of 29 parameters for sensitivity analysis was based on previous related studies [44]. Sensitivity analysis using the SUFI_2 algorithm identified 16 parameters significantly influencing runoff in the study area, with T sensitivity ≥ |0.70| and p significant value ≤ 0.50 for 16 parameters. The parameters were ranked in descending order of sensitivity as follows: CN2.mgt > CANMX.hru > SOL_BD(1).sol > SLSUBBSN.hru > HRU_SLP.hru > SOL_K(1).sol > SOL_AWC(1).sol > SFTMP.bsn > ALPHA_BNK.rte > RCHRG_DP.gw > SOL_Z(1).sol > SMFMN.bsn > GW_DELAY.gw > CH_K2.rte > ALPHA_BF.gw > EPCO.hru.

Typically, model results are deemed acceptable when both the coefficient of determination (R²) and Nash–Sutcliffe efficiency coefficient (NSE) exceed the threshold of 0.5. Notably, R² and NSE for the SWAT model were both 0.84 during the calibration period. During the validation period, both the R² and NSE reached 0.86, indicating a good simulation performance. This demonstrates that the SWAT model is suitably applicable to the study watershed and can be further utilized to investigate runoff variations under different scenarios. This robust agreement between simulated runoff and observed data for both the calibration and validation periods, as depicted in Figure 8, further attests to the SWAT model’s overall accuracy and its suitability for conducting runoff change attribution analyses at the Li County hydrological station.

The results of the analysis, as presented in

Table 11. Contribution analysis of SWAT model results in the study area.

Period	${\overline{\mathit{R}}}_{\mathit{s}\mathit{i}\mathit{m}\mathit{u}}^1$	${\overline{\mathit{R}}}_{\mathit{s}\mathit{i}\mathit{m}\mathit{u}}^2$	${\overline{\mathit{R}}}_{\mathit{o}\mathit{b}\mathit{s}\mathit{e}}^2$	∆R	∆Rc	∆Rh	ηc	ηh
	mm	mm	mm	mm	mm	mm	%	%
1960–1994	122.50	91.15	37.31	−85.19	−31.34	−53.85	36.79	63.21
1995–2016

, demonstrate a consistent decreasing trend in runoff. The reduction in runoff attributed to climate change is 31.34 mm, accounting for 36.79%. In contrast, human activities are found to be accountable for a larger decrease in runoff, amounting to 53.85 mm, and contributing significantly at a rate of 63.21%. Consequently, the comparison of simulated and observed runoff using the SWAT model unequivocally indicates that human activities are the primary driving force behind the observed reduction in runoff in the study area. This underscores the significance of human-induced factors in influencing the study area hydrological conditions.

4.6. LSTM Model Results

In this study, the LSTM model utilized data periods consistent with those of the SWAT model. Monthly meteorological and runoff data were employed, with seven meteorological variables selected as input features as follows: mean temperature, maximum temperature, minimum temperature, precipitation, mean relative humidity, mean wind speed, and sunshine duration. Runoff served as the output factor. Through preliminary calibration and hyperparameter tuning, an LSTM model tailored to the study area’s river basin was constructed.

The determination of optimal hyperparameters is crucial for enhancing the performance and accuracy of the LSTM model employed in this study. The final hyperparameters are as follows: the number of neurons is 300, the initial learning rate is 0.002, the number of iterations is 1000, and the packet loss rate is set to 0.3 after 700 times. The selection of these hyperparameters plays a key role in the performance improvement of the model and ensures the applicability and accuracy of the model in the study area.

In order to be consistent with the SWAT model, the runoff data from 1972 to 1976 were set as the training set, and the runoff data from 1977 to 1980 were set as the validation set. In the training set, the LSTM model in the study area performed well, the NSE and R² were 0.90, the NSE of the test set was 0.77, and the R² was 0.78 (Figure 9). The results indicate that the LSTM model has demonstrated well simulation performance, making it suitable for further application in runoff analysis within the basin. These performance indicators verify the excellent performance of LSTM in simulating monthly runoff, and lay a reliable foundation for subsequent runoff contribution rate analysis.

As shown in

Table 12. Contribution analysis of LSTM model results in the study area.

Period	Q0	Q1	Q2	∆QC	∆QH	ηc	ηh
	mm	mm	mm	mm	mm	%	%
1960–1994	98.18	82.76	65.91	−15.42	−16.85	47.78	52.22
1995–2016

, the reduction in runoff attributed to human activities is 16.85 mm, whereas the reduction caused by climate change is 15.42 mm. Subsequently, the adjusted LSTM model was used to calculate the contribution rate of runoff, and the final contribution rate of human activities was 52.22%, while the contribution rate of climate change was 47.78% (Table 12). This result is consistent with previous research results (Budyko and SWAT), both of which emphasize the dominant role of human activities in regional water resources changes. The research results provide a demonstration for the application of the LSTM model in hydrological simulation.

4.7. PLUS Model Land Use Scenario Simulation

In order to simulate land use changes in the control area of the Li County hydrological station and apply them to the runoff study, the research simulated land use data for four scenarios in the year 2030, based on the 2020 land use data as the baseline.

(1) Natural Development Scenario: This scenario essentially considers how land use might change without specific interventions or policies.

(2) Economic Development Scenario: This scenario examines the influence of economic development trends on land use.

(3) Ecological Conservation Scenario: The primary focus here is on simulating land use changes under the premise of ecological conservation.

(4) Comprehensive Development Scenario: This scenario considers various drivers of land use change and how they might interact to shape the future land use pattern.

The simulated land use data under these four scenarios will be applied in the subsequent runoff study. The specific transformation matrix settings are provided in

Table 13. Land use transfer rule matrix under policy-oriented scenarios.

Natural Development							Economic Development						Ecological Conservation						Comprehensive Development
	C	F	G	W	B	U	C	F	G	W	B	U	C	F	G	W	B	U	C	F	G	W	B	U
C	1	1	1	1	1	1	1	0	0	0	1	0	1	1	1	1	1	1	1	0	0	0	1	0
F	1	1	1	1	1	1	1	1	1	1	1	0	0	1	1	0	0	0	1	1	0	0	1	0
G	1	1	1	1	1	1	1	0	1	0	1	0	0	1	1	0	0	0	1	1	1	0	1	0
W	1	1	1	1	1	0	1	1	1	1	1	0	1	1	1	1	1	0	1	1	1	1	1	0
B	1	1	1	1	1	1	1	0	0	0	1	0	0	0	0	0	1	0	0	0	0	0	1	0
U	1	1	1	0	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

Note: C: cultivated land; F: forest; G: grassland; W: water area; B: built-up land; U: unused land.

.

The research has gone a step further to examine the potential impacts of various land use scenarios on runoff variations in the study area. Three extreme land use scenarios were established. The extreme forest scenario involved the conversion of all land cover types, except for water bodies and urban area, into forest land. Similarly, the extreme grassland and extreme cropland scenarios involved the conversion of all land cover types, except for water bodies and urban area, into grassland and cropland, respectively.

The research will use seven scenarios in the subsequent SWAT model analyses to assess how each land use pattern could impact future runoff conditions in the study area. This approach allows for a comprehensive understanding of the potential consequences of different land use decisions on the local hydrology.

4.8. Runoff Simulation in Future Change Scenarios

4.8.1. Land Use Change Scenarios

In the previous analysis, a total of seven simulation scenarios were employed within the PLUS model. These seven data sets pertaining to land use were subsequently utilized to simulate the underlying runoff dynamics in the context of future land use developments. The results of the simulated runoff magnitudes are presented in

Table 14. Simulated runoff depths under various future land use scenarios (mm).

	Policy-Oriented Land Use Scenarios				Extreme Scenario (Only Built-Up Land and Water Area Remain)
Land Use Scenarios	Natural Development	Economic Development	Ecological Conservation	Comprehensive Development	Extreme Forest	Extreme Grassland	Extreme Cultivated Land
Simulated Runoff Depth	89.81	90.44	89.76	90.56	75.78	74.46	101.50

. For the scenarios representing extreme forest and extreme grassland conservation policies, the computed runoff depths were 75.78 mm and 74.46 mm, respectively. The notable reduction in runoff depth observed in these cases can potentially be attributed to a reduction in evapotranspiration, stemming from an increased forest and grassland cover.

Under the natural development scenario, the computed runoff depth was 89.81 mm. This scenario aimed to maintain a hydrological cycle that closely aligned with historical local data trends. In the economic development scenario, the runoff depth was slightly higher at 90.44 mm. In this context, economic growth took precedence, potentially involving measures related to resource exploitation and overall economic progress. The marginal increase in runoff depth in this scenario may be attributed to the expansion of impermeable land cover. For the ecological conservation scenario, the computed runoff depth was 89.76 mm. This scenario prioritized ecological protection and sustainable development. The runoff depth was 90.56 mm in the comprehensive development scenario. This comprehensive scenario balanced economic growth, ecological preservation, and societal needs. In summary, the direction and policies governing future land use changes will impact runoff variations within the study area.

4.8.2. Climate Change Scenarios

Based on the content from the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report (AR6), the global average surface temperature is expected to reach or exceed 1.5 °C within the next 20 years. Furthermore, historical data for the study area indicates a consistent upward trend in temperatures and a concurrent decline in precipitation over the years. Drawing from the insights provided by the IPCC AR6 and the available historical climate data spanning the years from 1960 to 2016, the future climate change scenario for the study area is delineated in

Table 15. Setting of future climate scenarios.

Temperature\Precipitation	Decrease by 10%	Decrease by 15%	Increase by 10%	Unchange
Unchange	S1	S2	S3	S4
Increase by 1 °C	S5	S6	S7	S8
Increase by 2 °C	S9	S10	S11	S12
Decrease by 2 °C	S13	S14	S15	S16
Increase in Max Temperature by 1 °C	S17	S18	S19	S20
Increase in Max Temperature by 2 °C	S21	S22	S23	S24
Increase in Min Temperature by 1 °C	S25	S26	S27	S28
Increase in Min Temperature by 2 °C	S29	S30	S31	S32

. This approach results in a total of 32 distinct future climate change scenarios when combining the various temperature and precipitation change scenarios.

The calibrated and validated SWAT model was effectively employed to simulate runoff patterns within the study area under prospective climate scenarios. These simulations obtained the results of runoff changes under different climate change scenarios in the future.

Table 16. Future runoff depth simulations under climate scenarios (mm).

Temperature\Precipitation	Decrease by 10%	Decrease by 15%	Increase by 10%	Unchange
Unchange	66.31	55.10	119.13	91.16
Increase by 1 °C	67.19	55.92	119.76	92.00
Increase by 2 °C	67.55	56.36	119.80	92.17
Decrease by 2 °C	66.64	55.25	119.60	91.57
Increase in Max Temperature by 1 °C	66.72	55.52	119.20	91.46
Increase in Max Temperature by 2 °C	67.16	55.91	119.74	91.98
Increase in Min Temperature by 1 °C	66.74	55.53	119.23	91.47
Increase in Min Temperature by 2 °C	67.16	55.90	119.72	91.97

illustrates that regional runoff exhibits notable fluctuations in response to changes in precipitation and temperature. The following explanations provide a detailed account:

(1) Precipitation Impact: When temperature remains constant, a reduction of 10% and 15% in precipitation yields corresponding decreases in simulated runoff depth, reducing it from 91.16 mm to 66.31 mm and 55.10 mm, respectively. Conversely, a 10% increase in precipitation results in an amplified simulated runoff depth, raising it from 91.16 mm to 119.13 mm.

(2) Temperature Impact: When precipitation remains constant, an increase in temperature of 1 °C and 2 °C leads to an incremental rise in simulated runoff depth, with values of 91.16 mm increasing to 92.00 mm and 92.17 mm, respectively. Conversely, a decrease in temperature of 2 °C marginally increases the simulated runoff depth, changing it from 91.16 mm to 91.57 mm.

Among the 32 scenarios examined, Scenario S11 stands out with the highest simulated runoff depth. A 2 °C increase in temperature coupled with a 10% rise in precipitation results in a simulated runoff depth of 119.80 mm. This represents a significant increase of 28.64 mm compared to Scenario S4. In contrast, Scenario S2 displays the lowest simulated runoff depth, where temperature remains unchanged, and a 15% reduction in precipitation leads to a runoff depth of 55.10 mm. This reflects a substantial decrease of 36.06 mm compared to Scenario S4. Thus, precipitation exerts a significant influence on runoff dynamics in the study area. Under future climate scenarios, changes in precipitation patterns will play a crucial role in determining runoff variations, while the impact of temperature changes remains relatively minor.

5. Discussion

5.1. Results from Eight Budyko Equations

Table 7 presents the outcomes derived from employing eight distinct Budyko methods within the study area. These results consistently emphasize the significant influence of human activities as the primary driver behind observed hydroclimatic phenomena, albeit with varying specific contribution rates. Notably, the application of these diverse Budyko methods using identical data often yields disparate outcomes, a phenomenon highlighted by Shi et al. [45]. Such variability can be attributed to the unique assumptions, parameters, and model structures inherent in each Budyko method, leading to differential responses to various influencing factors in hydroclimatic assessments. Within the framework of Budyko analysis, the selection of critical parameters, such as the evaporative coefficient, becomes pivotal. Different methods may favor distinct parameter choices and adopt varied parameterization approaches, resulting in differences in calculated results, as observed by Shao et al. [46]. These discrepancies in parameterization among different Budyko methods inevitably translate into variations in derived contribution rates. Therefore, relying solely on a single approach may lead to limited accuracy and potential errors. However, integrating multiple Budyko methodologies holds promise for enhancing reliability and precision of outcomes, thereby mitigating inherent modeling uncertainties [47]. The amalgamation of diverse methods can mitigate individual biases and bolster overall predictive robustness. Moreover, comparative evaluation of these approaches contributes to a deeper understanding of their applicability and limitations, offering valuable insights for refining and optimizing future modeling endeavors.

5.2. The SWAT Model and the PLUS Model

The accuracy of both the Soil and Water Assessment Tool (SWAT) model and the patch-generating land use simulation (PLUS) model hinges significantly on the precise configuration of their parameters, as highlighted by Nguyen et al. [48]. However, it is crucial to acknowledge that achieving absolute model perfection may be unattainable due to inherent limitations in available temporal data and parameter settings. These constraints can result in disparities between model predictions and actual observations. Furthermore, the SWAT and PLUS models require extensive input data encompassing meteorological, land use, and hydrological information. The complexity and uncertainties associated with data collection can compromise the accuracy of these inputs, thereby influencing the outcomes of simulated models [49,50]. For instance, elevation data sourced from various providers may impact SWAT model construction differently across regions [51]. It is essential to recognize that simulated results are based on assumptions and model inferences, necessitating thorough analysis and validation using observed data and expert judgment in practical applications. To address these challenges, initial calibration and validation of the SWAT model were conducted using SWAT-CUP software, employing relevant metrics to assess model accuracy. Subsequently, runoff variations predicted by the SWAT model were validated using the Budyko method, reinforcing its reliability. Similarly, validation and simulation accuracy of the PLUS model primarily relied on the Kappa coefficient, as emphasized by Wang et al. [52]. In summary, despite potential challenges in simulation accuracy for both the SWAT and PLUS models, their reliability and applicability can be bolstered through rigorous validation against empirical data and information. This validation process is critical for enhancing the models’ credibility in subsequent analyses, thereby producing more trustworthy outcomes that can inform water resource management and decision-making processes effectively.

5.3. Setting of Future Scenarios

Currently, a significant number of studies utilize future climate scenario data with a spatial resolution of approximately 0.5°, a practice that has shown promising results, particularly for larger study areas, as highlighted by Hajjar et al. [53]. However, it is crucial to acknowledge that for relatively small study areas, the use of climate data with insufficient precision can introduce significant errors into the final results. Therefore, this study adopted an approach grounded in existing historical data within the study area to discern trends and variations. Subsequently, climate change scenarios were formulated based on this foundation. Anticipated temperature and precipitation ranges within the study area were projected to undergo changes within specific intervals, with various combinations of these changes constructed to represent future climate scenarios. This method is not only practical but also easily understandable and applicable. Importantly, it has garnered support from numerous scholars, as documented by Wang et al. [54]. Furthermore, the configuration of land use scenarios was established by assigning distinct probabilities to land cover transitions across different land classes. By setting these probabilities, it became feasible to simulate runoff conditions under various future land use scenarios and gain additional insights by examining runoff changes under alternative land use developments [55]. This dual-purpose approach enhances the study’s ability to forecast hydrological impacts under plausible future conditions while providing valuable insights into the potential consequences of different land use strategies.

5.4. The LSTM Model

The results obtained using the long short-term memory (LSTM) model are compared with those derived from other methods such as SWAT and the Budyko model. Remarkably, the outcomes of the LSTM model closely align with those of SWAT and Budyko. This consistency across methodologies underscores the reliability of the LSTM approach in simulating watershed runoff processes. While all three methods yield similar results, it is essential to acknowledge the strengths and limitations of the LSTM model compared to alternative approaches. One advantage of LSTM is its capability to capture complex temporal dependencies in the data, which makes it well-suited for modeling dynamic hydrological processes over time [56]. Additionally, LSTM can effectively handle nonlinear relationships between input variables and runoff, offering greater flexibility in modeling real-world hydrological systems. However, it is important to note that LSTM models require a substantial amount of data for training and can be computationally intensive, especially when dealing with large-scale watersheds or high-resolution data [57]. Furthermore, the interpretability of LSTM models may be limited compared to simpler models like Budyko, which provide more straightforward insights into the underlying hydrological processes. In summary, while LSTM demonstrates promising performance in simulating watershed runoff, researchers should carefully consider the tradeoffs between accuracy, computational complexity, and interpretability when selecting modeling approaches for specific applications in hydrological studies.

5.5. Uncertainty Analysis

The methodologies and data sets utilized in this study exhibit inherent limitations that warrant consideration. Specifically, the attribution analysis of runoff changes did not differentiate between the specific impacts of climatic factors such as variations in precipitation and temperature. While human activities were predominantly assessed through changes in land use, it is essential to recognize that the influence of human actions extends beyond land use changes alone. Factors like soil and water conservation practices, hydraulic engineering projects, and water usage in industrial and agricultural sectors [58] contribute significantly to hydrological processes and should be more comprehensively integrated into future analyses. Accurately quantifying the precise impacts of climatic factors and human activities, and further disaggregating their contributions to runoff changes, necessitates additional research efforts. Moreover, the adequacy of training samples used for modeling can significantly affect simulation accuracy [59]. This study relied solely on monthly runoff data for calibrating and validating the SWAT and LSTM models. Enhancing simulation accuracy in future investigations could involve simulating daily runoff data and comparing these results with those obtained from monthly simulations. Regarding future scenario simulations, the scenarios developed in this study were based on IPCC guidelines, focusing on average levels of precipitation and temperature changes. Given the regional scale of this study, inherent uncertainties may exist in the outcomes. Future studies should consider conducting more detailed and granular analyses to provide deeper insights into how specific changes in climate and human activities impact runoff dynamics within the study area.

6. Conclusions

Based on data collected from meteorological stations and the Li County hydrological station, this study employed a comprehensive approach to analyze runoff changes within the study area. The analysis utilized the Budyko framework, the SWAT model, and the LSTM model to investigate the evolutionary characteristics and influencing factors affecting these changes, with a specific emphasis on understanding the impact of human activities. Furthermore, this study projected future scenarios of runoff changes. The principal findings of this investigation are summarized as follows:

(1) The runoff in the controlled basin of Li County’s hydrological station on the Xihanshui River exhibited a pronounced decreasing trend spanning the period from 1960 to 2016. An abrupt change in runoff was notably identified in 1994, signifying a significant alteration in runoff patterns within the study area. The observed influence of human activities on runoff predominantly manifests through modifications in land cover. Specifically, there have been reductions in the areas of grassland and cropland, accompanied by expansions in land uses such as forestland and residential areas. Notably, forestland has experienced the most substantial increase in area.

(2) The results obtained from eight hydrological balance equations, all grounded in the Budyko hypothesis, consistently highlight a significant trend. Human activities emerge as the primary driver behind runoff reduction, accounting for 50% to 60% of the total reduction. Utilizing the calibrated SWAT model further underscores human activities as the predominant force responsible for runoff reduction, contributing 63.21% to the overall decrease. Additionally, employing the LSTM model tailored to the watershed of the study area reveals promising simulation capabilities. With NSE and R² values reaching 0.90 in the training set and exceeding 0.75 in the test set, the LSTM effectively captures watershed runoff processes. Specifically, the LSTM model estimates that human activities contribute 52.22% to runoff reduction, while meteorological factors contribute 47.78%. Overall, the congruence in results across these three methodologies solidifies the pivotal role of human activities in driving the observed reduction in runoff.

(3) In the future land use scenarios, the ecological protection scenario generated the least amount of runoff, while the comprehensive development scenario led to the highest runoff levels. This variation in runoff can be predominantly attributed to the substantial impacts of intensified human activities. In the 32 future climate scenarios, runoff simulations are notably elevated in the eight scenarios where precipitation increases by 10%, while they are significantly reduced in the eight scenarios where precipitation decreases by 15%. Precipitation remains the primary driving factor affecting runoff depth, while temperature exhibits limited sensitivity in this regard.