Hydrological Response to Climate Change: McGAN for Multi-Site Scenario Weather Series Generation and LSTM for Streamflow Modeling

Abstract

This study focuses on the impacts of climate change on hydrological processes in watersheds and proposes an integrated approach combining a weather generator with a multi-site conditional generative adversarial network (McGAN) model. The weather generator incorporates ensemble GCM predictions to generate regional average synthetic weather series, while McGAN transforms these regional averages into spatially consistent multi-site data. By addressing the spatial consistency problem in generating multi-site synthetic weather series, this approach tackles a key challenge in site-scale climate change impact assessment. Applied to the Jinghe River Basin in west-central China, the approach generated synthetic daily temperature and precipitation data for four stations under different shared socioeconomic pathways (SSP1-26, SSP2-45, SSP3-70, SSP5-85) up to 2100. These data were then used with a long short-term memory (LSTM) network, trained on historical data, to simulate daily river flow from 2021 to 2100. The results show that (1) the approach effectively addresses the spatial correlation problem in multi-site weather data generation; (2) future climate change is likely to increase river flow, particularly under high-emission scenarios; and (3) while the frequency of extreme events may increase, proactive climate policies can mitigate flood and drought risks. This approach offers a new tool for hydrologic–climatic impact assessment in climate change studies.

1. Introduction

The effects of global climate change on the water cycle are of significance [1,2]. Understanding how these changes impact hydrological processes is a crucial issue in current hydrometeorological research [3,4,5]. As the fundamental component of the hydrological cycle, watershed processes are particularly sensitive to climate change [6,7,8]. Quantitatively assessing the impacts of future climate change on watershed streamflow is crucial [9,10]. This not only helps us understand possible changes in the global water cycle but also provides support information for the development of regional adaptive water management strategies [11,12]. The main approach to this assessment involves constructing relationships between meteorological data and streamflow and then estimating watershed hydrological responses to future climate change scenarios [13,14,15]. The techniques for constructing these relationships vary, including physically based distributed hydrological models [16,17], data-driven statistical models [18,19], and semi-empirical models that combine both approaches [20]. Recently, machine learning techniques, particularly deep learning approaches such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), and long short-term memory networks (LSTMs), have shown potential and advantages in modeling watershed streamflow using time series weather data [21,22,23]. However, regardless of the method used, a key challenge in assessing future climate change responses is obtaining projected future climate data that can meet the requirements of the relational model assessment.

In climate change impact studies, global climate models (GCMs) are the primary tools for modeling future climates under different scenarios [24,25]. The latest shared socio-economic pathways (SSPs) scenarios presented by the Coupled Model Intercomparison Program Phase 6 (CMIP6) provide a common framework for inter-country scientists to use different GCMs and/or their outputs in an equivalent way for climate change studies [26]. However, GCM data have several major limitations. GCM outputs are often systematically biased and require correction. They also have very low spatial resolution (typically ~100-km), which is much lower than the accuracy required for watershed-scale or site-scale studies [27]. Additionally, simulations from different GCMs are often significantly different, and using a single GCM output increases the uncertainty of climate change impact assessments. To overcome these limitations and obtain highly accurate time-series data with short time steps, researchers have developed various bias correction and downscaling methods [28,29]. Kim et al. [30] proposed and implemented a new multivariate bias correction method to improve the accuracy of regional climate models. Islam et al. [31] developed a simple polynomial bias correction method for correcting GCM-derived forecast runoff biases, further demonstrating the importance of bias correction for hydrologic modeling. Elisabeth Vogel et al. [32] emphasized the importance of downscaling and bias correction methods in improving the spatial and temporal resolution of climate model outputs and removing systematic biases. These improvements are essential for accurately assessing the impacts of climate change on hydrological processes. Wamahiu et al. [33] highlighted the need for bias correction of GCMs prior to dynamic downscaling processes, which can significantly impact the simulation of precipitation and temperature for current and future climates. The study by David E. Robertson [34] examines how precipitation data output from GCM downscaling can be processed to improve hydrologic simulation results, focusing on the importance of data processing at different time scales. Preprocessing GCM data is common and necessary based on specific problems and application requirements prior to use, with weather generators being particularly effective in addressing time series related issues.

Weather generators (WGs), as commonly used statistical downscaling tools, are capable of generating long series of synthetic meteorological data, which have significant advantages especially for the analysis of small-area or site-scale future meteorological data problems. Sylvie Parey et al. [35] proposed a bivariate generator of daily temperature and rainfall based on Hidden Markov Models, which can be used not only for analyzing current climatic conditions but also for predicting future climates, providing a new way of thinking about the generation of multivariate and multisite weather series. Li et al. [36] evaluated the application of the multisite weather generator MulGETS in the Yangtze River Basin, and the results showed that the model can effectively generate spatially correlated precipitation series while maintaining the spatial and temporal distribution characteristics of precipitation. This provides a new tool for precipitation simulation and climate change impact assessment in large-scale basins. Evin et al. [37] proposed an important extension to the Wilks multisite daily precipitation model, which integrates the heavy-tailed distribution, spatial tail dependence, and temporal dependence using decomposition methods, effectively reproducing the statistical characteristics of extreme rare events across various temporal and spatial scales and demonstrating a clear advantage in simulating and assessing large-area extreme precipitation events. However, existing downscaling methods still face challenges in dealing with multi-site data, especially in modeling extreme events and maintaining consistency of precipitation occurrence among sites. Additionally, effectively integrating outputs from multiple GCMs while maintaining the spatial consistency of the downscaled data remains a critical issue [38].

In recent years, artificial intelligence and deep learning techniques have shown great potential in reanalysis, downscaling, and the generation of synthetic meteorological data [39,40]. Among them, generative adversarial networks (GANs) are increasingly being applied to this field due to their advantages in generating complex distributed data [41,42]. Sha et al. [43] proposed a conditional deep convolution generative adversarial nets (cDCGAN)-based downscaling method that employs image super-resolution techniques to generate high-resolution daily temperature data from low-resolution GCM output. While this method performs well in maintaining spatial distribution features, it primarily targets univariate and raster data. Hong Kang Ji et al. [44] highlight that GANs are effective at capturing the spatial correlations of extreme events and accurately reflecting the original data’s spatial and temporal distributions, rather than simply replicating their statistical features. This ability makes GANs particularly effective in modeling extreme rainfall events and reflecting real spatial correlation patterns. The GAN-based weather generators utilize semi-supervised learning mechanisms to better capture potential spatial distribution features of historical data, generating new data that are similar to, but not identical to, historical data, avoiding the rigidity of traditional regression methods. Meanwhile, by introducing noise inputs, GANs can effectively simulate the inherent uncertainty of meteorological data, especially excelling in reflecting the stochastic nature of precipitation distributions, thus generating simulation results more aligned with the characteristics of real meteorological distributions. However, previous studies have mainly focused on applying GANs to single-site weather generation or spatial downscaling of gridded data, while the application of GANs to multi-site weather generation, especially maintaining both temporal and spatial consistency among multiple stations, remains largely unexplored.

Accordingly, this study proposes an innovative technical framework for assessing the impacts of climate change on watershed hydrological processes. The framework consists of three key components: (1) a future scenario weather generator (FS-WG), established and validated with historical site observations, that adjusts probability density function parameters based on monthly-scale climate change information from multiple GCM ensemble predictions to generate regional average synthetic daily weather series; (2) a multi-site conditional generative adversarial network (McGAN) model that captures spatial correlation characteristics from historical data to transform regional average weather data into multi-site daily scale data while maintaining spatial consistency; and (3) a long short-term memory (LSTM) network model for establishing nonlinear relationships between multi-site meteorological data and watershed outlet streamflow. Using the Jinghe River Watershed in west-central China as a case study, this framework evaluates the impacts of future climate change on watershed hydrological processes under four shared socioeconomic pathways (SSPs) scenarios. Unlike traditional spatial downscaling methods, this approach focuses on generating site-scale weather sequences with temporal continuity and spatial consistency, particularly suitable for watershed hydrological impact assessments that require continuous daily-scale station data. The overall technical framework is illustrated in Figure 1.

2. Materials and Methods

2.1. Study Area and Data

This study is carried out in the Jing River watershed (Figure 2), located in central-western China, at the junction of Shaanxi, Gansu, and Ningxia provinces. The Jing River, a secondary tributary of the Yellow River, joins the Wei River before eventually flowing into the Yellow River. In this study, the Tao-yuan hydrological station, situated upstream of the confluence of the Jing and Wei Rivers, is used as the watershed outlet. The total catchment area is 45,442 km². The average elevation of the Jing River Basin is 1387 m, with the highest point reaching 2906 m and the lowest at 352 m above sea level. Cropland and grassland are the primary land use types within the watershed, with forested areas dominating in the headwater regions. There are four national meteorological stations within the watershed, with detailed information provided in

Table 1. Characteristics of meteorological stations in the Jinghe River watershed.

Name	ID	Longitude	Latitude	Elevation	Periods
Huan Xian	53821	36.583	107.300	1255.6	1957-01-01 to 2020-12-31
Ping Liang	53915	35.550	106.667	1346.6	1957-01-01 to 2020-12-31
Xi Feng Zhen	53923	35.733	107.633	1421.0	1957-01-01 to 2020-12-31
Chang Wu	53929	35.200	107.800	1206.5	1957-01-01 to 2020-12-31

. The average annual temperature is 9.2 °C, and the average annual precipitation is 524.5 mm. Details as to the data sources are provided in

Table 2. Summary of the original data source.

Name	Source and Description	Resolution	Remark
Historical Weather Records	Climatic Data Center, National Meteorological Information Center, China Meteorological Administration (http://data.cma.cn, accessed on 1 October 2024)	Daily	1957–2020
Historical Hydrological Data	Annual Hydrological Report P. R. China, Volume 4(8), from National Library of China	Monthly	2006–2020
Historical GCM Data	Coupled Model Intercomparison Project Phase 6 Dataset (https://pcmdi.llnl.gov/CMIP6/, accessed on 1 October 2024)	1 degree	~100 km × 100 km
Future Climate Data	Global Historical and Future Climate Grid Dataset Downscaled Based on GCMs (www.worldclim.org, accessed on 1 October 2024)	10 min	18.5 km × 18.5 km

.

2.2. Hydrological Long Short-Term Memory (LSTM) Design

In this study, a long short-term memory (LSTM) network model was developed to simulate daily streamflow at the watershed outlet. LSTM networks are particularly effective for processing time series data, making them suitable for capturing the relationships between meteorological conditions and watershed discharge. The model uses daily precipitation and temperature data from four meteorological stations within the watershed to predict the discharge at the outlet. The training and testing datasets were created using daily records from 2006 to 2020, including both meteorological data and discharge measurements. All data were normalized to a range of 0 to 1 using the min–max scaler method to ensure numerical stability during training. A 30-day window period was established, where meteorological data within this window serve as input features and the discharge at the end of the window serves as the prediction target, enabling the model to learn streamflow patterns based on recent meteorological variations.

The model architecture consists of four LSTM layers, each with 64 hidden units, and incorporates dropout to prevent overfitting. A fully connected layer follows the LSTM layers to produce the final predictions. This design captures both short-term and long-term dependencies in the time series data. The model was trained using mean squared error (MSE) as the loss function and the Adam optimizer, with L2 regularization implemented for better generalization. Following the temporal nature of the data, the dataset was chronologically divided, with the first 80% used for training and the remaining 20% for testing. This split ensures that only past data are used to predict future streamflow. Model performance was evaluated using both the mean absolute percentage error (MAPE) and Nash–Sutcliffe efficiency coefficient (NSE). The trained LSTM model provides an effective tool for assessing climate change impacts on watershed hydrology. By inputting synthetic weather data generated under different climate scenarios, the model can simulate potential future streamflow patterns. The methods for generating these synthetic weather series are detailed in Section 2.3 and Section 2.4.

2.3. Future Scenario Weather Generator (FS-WG) Design

The future scenario weather generator (FS-WG) developed in this study treats the target study area as a whole, representing the average conditions of the four meteorological stations within the area. The FS-WG model distinguishes between dry days (no precipitation) and wet days (with precipitation) to generate two key meteorological variables: daily average temperature and daily precipitation. Assuming regional meteorological characteristics remain consistent within each month, separate model parameter sets are established for each of the 12 months. A two-state first-order Markov chain model is employed to generate sequences of dry and wet weather status.

The model parameters are calibrated using historical data from 1957 to 2020. Specifically, if the daily average precipitation from the four meteorological stations on any given day is greater than zero (i.e., if at least one station records precipitation), that day is classified as wet. Two sets of transition probabilities are calculated for each month: dry-to-wet and wet-to-dry. The first day’s status is determined by sampling based on historical precipitation probability, with subsequent days determined using these transition probabilities. The FS-WG model generates daily precipitation and temperature data separately for dry and wet days using probability density functions established through kernel density estimation (KDE) with a Gaussian distribution kernel. On dry days, precipitation is set to zero and temperature is sampled from the dry-day temperature distribution. On wet days, both variables are sampled from their respective wet-day distributions.

For future climate change scenarios, this study utilizes four shared socioeconomic pathways (SSPs: 1-26, 2-45, 3-70, and 5-85) proposed by CMIP6, focusing on four time periods (2030s, 2050s, 2070s, and 2090s). To minimize uncertainties, an ensemble of GCM results was employed using the WorldClim version 2.1 database [45]. This dataset offers global raster data of monthly temperature and precipitation projections for the four SSPs across the four future time periods, as predicted by different GCMs. By extracting the raster data for the target study area, monthly change rates in precipitation and temperature were calculated by comparing future projections with baseline data for each scenario and period. These rates were used to adjust the KDE probability density functions for generating scenario-based synthetic weather series. The FS-WG outputs, representing regional averages, are further processed by the McGAN model to generate spatially consistent multi-site data, which then serve as inputs for the LSTM model to assess climate change impacts on streamflow.

This study initially identified 11 GCMs from WorldClim version 2.1 that provide complete data sets for each of the four SSPs (SSP1-26, 2-45, 3-70, 5-85) and corresponding time periods (2030s, 2050s, 2070s, 2090s). These models include ACCESS-CM2, BCC-CSM2-MR, CMCC-ESM2, EC-Earth3-Veg, GISS-E2-1-G, INM-CM5-0, IPSL-CM6A-LR, MIROC6, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL. To evaluate their performance in representing local climate conditions, we used the original CMIP6 outputs (100 km × 100 km resolution) from these GCMs. For each GCM, historical monthly precipitation and temperature data at the four meteorological station locations were extracted from their CMIP6 simulations for the period 1957–2015. These values were compared with observed historical data from the four meteorological stations to estimate the performance of each GCM in simulating historical climate conditions for the study area. Historical accuracy was estimated using metrics such as mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), the coefficient of determination (R²), and the Nash–Sutcliffe efficiency coefficient (NSE). K-means clustering analysis was performed based on the suite of estimation metrics, and the optimal number of clusters was determined using the Davies–Bouldin index (DBI). The GCMs in the cluster with the highest historical accuracy were selected as the final models for assessing future climate changes.

For these selected GCMs, we then used their WorldClim data (18.5 × 18.5 km resolution), which provides downscaled and bias-corrected monthly values for both historical period and future periods under different SSP scenarios. Monthly change rates were calculated by comparing the historical data with future projections for each combination of time periods (2030s, 2050s, 2070s, 2090s) and SSP scenarios (SSP1-26, 2-45, 3-70, 5-85). Using consistent resolution data for both historical and future periods avoided potential biases from resolution differences. The ensemble average of these change rates across the selected GCMs was used as the adjustment factor for the KDE in our weather generator, with different adjustment factors applied for each SSP scenario and future period. This ensemble approach helps reduce uncertainties associated with individual GCM projections while maintaining the temporal and spatial consistency of climate change signals across different scenarios. Using these updated parameters, the FS-WG model, which was originally developed and validated with historical data, was applied to generate synthetic daily weather series for each combination of SSP scenarios and future time periods (2030s, 2050s, 2070s, and 2090s). For each scenario-time period combination, a 20-year daily temperature and precipitation series was generated. Since the output of the FS-WG model represents the average values for the four meteorological stations in the study area, a multi-conditional generative adversarial network (McGANs) model was further employed to generate daily temperature and precipitation data for each of the four stations, based on the daily average temperature and precipitation values. The design and training process of the McGANs model are detailed in Section 2.4.

2.4. Multi-Conditional Generative Adversarial Networks (McGANs) Design

In this study, a multi-conditional generative adversarial networks (McGANs) model was developed to generate sets of temperature and precipitation data for four stations within the study area based on the regional average temperature and precipitation. The model architecture is specifically designed for transforming regional average values into 2 × 2 grid data representing meteorological values for four target stations (Figure 3).

The generator employs a deconvolutional–convolutional neural network architecture. As shown in Figure 3a, an initial random 512 × 1 × 1 noise vector is transformed into a 128 × 16 × 16 noise tensor through four layers of deconvolution. Concurrently, a 1 × 2 × 2 conditional tensor is constructed where all four values are identical and equal to the target regional average. This conditional tensor undergoes three layers of deconvolution to form a 128 × 16 × 16 conditional tensor. The noise tensor and conditional tensor are merged along their dimensions to form a 256 × 16 × 16 coupled tensor, which is further processed through three convolutional layers to produce a 1 × 2 × 2 output tensor. Each layer is followed by BatchNorm2d and uses LeakyReLU activation function, while the final output layer employs a Sigmoid activation function to ensure the outputs are within the range of 0–1. A bias correction ensures the mean of the generated 2 × 2 grid equals the input conditional number, while the spatial distribution pattern is determined by the network based on learned historical characteristics.

The discriminator is constructed based on a conditional convolutional neural network architecture. As shown in Figure 3b, both the input 2 × 2 grid tensor and its corresponding condition tensor are processed through convolutional layers, each yielding intermediate tensors of size 32 × 2 × 2. These tensors are then concatenated along the first dimension to form a 64 × 2 × 2 tensor, which is further processed through two additional convolutional layers to produce a final 1 × 1 × 1 tensor output. Each convolutional layer is followed by BatchNorm2d and uses LeakyReLU activation function, with the final layer employing a Sigmoid activation function to ensure the output value is constrained between 0 and 1, representing the probability that the discriminator estimates the input tensor as real.

The training process follows the standard GAN framework with several stability enhancements: (1) initial 100 epochs for discriminator-only training; (2) dynamic learning rate adjustment for the generator; and (3) early stopping mechanism based on Nash equilibrium state detection. Training stops when the discriminator’s accuracy for both real and fake samples stabilizes around 0.5 for 50 continuous epochs after the 500th epoch.

More details about the model architectures and implementation procedures described in Section 2.2, Section 2.3 and Section 2.4, including LSTM configuration, FS-WG parameter calibration processes, and McGAN training strategies, along with all model codes used in this study, are available in the Supplementary Materials.

3. Results

3.1. LSTM Results for Hydrological Modelling

Using an LSTM model, the daily river streamflow at the watershed outlet was simulated based on 30 days of prior daily temperature and precipitation data from four meteorological stations, as illustrated in Figure 4. The results indicate that the mean absolute percentage error (MAPE) between the simulated and observed streamflow during the test period is 81.10%, while the Nash–Sutcliffe efficiency (NSE) is 0.79. The test dataset was entirely independent from the training dataset, having never been used during the LSTM training process. This suggests that the LSTM model demonstrates good extrapolation capabilities and is well-trained.

The main hyperparameters used for training the LSTM included setting the model with four LSTM layers, a hidden layer dimension of 64, a learning rate of 0.0005, and a dropout ratio of 0.5. Using the LSTM model developed in this study, it is possible to analyze daily river streamflow at the watershed outlet based on daily weather data from the four stations. By using synthetic daily weather data series that represent future climate change scenarios as input into the LSTM model, future river streamflow can be modeled and predicted, allowing for an assessment of the potential impacts of future climate change on the watershed’s hydrological processes. Detailed discussions about the results and reliability of the LSTM model can be found in Section 4.1.

3.2. Future Scenario Weather Generation

3.2.1. GCM Selection and Projected Future Climate Change

The simulation accuracy of the 11 alternative GCMs for historical monthly precipitation and average daily temperature in the target study area is shown in

Table 3. Simulation accuracy and clustering analysis results for the 11 selected GCMs for the study area.

GCM\Items	Monthly Precipitation					Monthly Average Daily Temperature					Clusters
	MSE	RMSE	MAE	R2	NSE	MSE	RMSE	MAE	R2	NSE
ACCESS-CM2	166.24	10.48	7.83	0.92	0.82	163.34	10.03	7.57	0.92	0.82	1
BCC-CSM2-MR	706.26	20.33	17.13	0.94	0.41	668.18	18.78	15.71	0.94	0.45	4
CMCC-ESM2	3974.80	46.01	37.56	0.93	−2.06	3439.84	40.03	32.54	0.93	−1.50	2
INM-CM5-0	1818.75	31.04	23.43	0.80	−0.41	1781.39	28.96	21.87	0.82	−0.34	3
IPSL-CM6A-LR-INCA	212.56	12.09	9.31	0.92	0.72	150.14	10.18	8.07	0.93	0.77	1
MIROC6	1317.29	27.18	19.55	0.80	−0.07	1364.52	26.58	19.03	0.82	−0.09	0
MRI-ESM2-0	319.70	14.22	9.76	0.84	0.69	301.29	13.23	8.90	0.85	0.71	1
UKESM1-0-LL	101.27	8.69	6.08	0.94	0.84	101.73	8.51	5.93	0.94	0.83	1
EC-Earth3-Veg	124.43	9.27	7.48	0.94	0.85	119.55	8.79	6.98	0.93	0.85	1
GISS-E2-1-G	994.07	23.30	20.44	0.94	0.21	1020.39	22.56	19.62	0.94	0.20	5
MPI-ESM1-2-HR	315.05	14.05	11.27	0.86	0.70	287.91	12.85	10.29	0.87	0.73	1

. Based on the accuracy indicators, a K-means clustering analysis was performed, and the Davies–Bouldin index (DBI) was used to determine the optimal number of clusters. The DBI assesses the quality of clustering by comparing the compactness within each cluster and the separation between clusters; a lower DBI value indicates better clustering effectiveness. The DBI results for the clustering analysis of the historical data fitting accuracy of the 11 GCMs are shown in Figure 5. The analysis indicated that clustering into six groups was a locally optimal choice, with the effectiveness improving with more categories. When the number of categories reached 10, the DBI value was at its lowest, suggesting optimal internal compactness and separation between groups at this classification. However, since there are only 11 datasets in total in this study, dividing into 10 groups clearly contradicts the purpose of conducting clustering analysis and does not meet the application demands. Therefore, this study adopted a locally lower DBI value and performed a K-means clustering analysis of the 11 GCMs into six groups, selecting the group with the highest result accuracy to be the chosen GCMs.

Six GCMs that have historically performed well in the study area were selected, including ACCESS-CM2, EC-Earth3-Veg, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL. These GCMs have a coefficient of determination (R²) above 0.84 and a Nash–Sutcliffe efficiency coefficient (NSE) above 0.69 for historic data, indicating that their simulation results significantly outperform the average of actual measurements, proving to be both positive and effective. The extracted data include monthly precipitation and average daily temperature from these six GCMs across four future time periods (2030s, 2050s, 2070s, and 2090s) and four climate scenarios (SSP 1-26, SSP 2-45, SSP 3-70, and SSP 5-85). These values are compared with current baseline levels, and the mean of the change rates from the six GCMs is used as the ensemble outputs, representing projected climate changes. Since the future GCM data provided by WorldClim includes only daily maximum and minimum temperatures, their average is used to compute the daily average temperature to meet the meteorological data requirements necessary for LSTM-based streamflow modeling. These projected climate change data are then integrated into the FS-WG model to generate synthetic daily temperature and precipitation data series that reflect the characteristics of various time periods and scenarios projected for the study area in the future. Detailed results of this integration are described in the following Section 3.2.2.

3.2.2. FS-WG for Reginal Weather Generation

The FS-WG model developed in this study is calibrated based on averages of the historical data from four meteorological stations in the study area from 1957 to 2020, with dry day ratios and dry-to-wet and wet-to-dry transition probabilities as shown in

Table 4. Dry–wet transition probability parameters of the FS-WG model.

Month	Probability Dry	Probability Wet to Dry	Probability Dry to Wet
Jan.	0.82	0.36	0.14
Feb.	0.74	0.35	0.22
Mar.	0.70	0.35	0.27
Apr.	0.62	0.33	0.35
May.	0.56	0.31	0.44
Jun.	0.51	0.27	0.50
Jul.	0.38	0.19	0.54
Aug.	0.38	0.17	0.49
Sep.	0.40	0.19	0.46
Oct.	0.53	0.28	0.39
Nov.	0.72	0.35	0.21
Dec.	0.86	0.42	0.13

. Here, “Probability Dry” represents the probability that any given day in a month is a dry day (no precipitation), which is used to generate the dry or wet status for the first day of the month. “Probability Wet to Dry” and “Probability Dry to Wet” represent the probabilities of transitioning from a wet day to a dry day and from a dry day to a wet day, respectively, based on which the daily dry or wet status throughout the month is generated. After generating the daily dry/wet statuses, samples are taken from the kernel density functions corresponding to dry and wet states to obtain daily precipitation and average daily temperatures. The kernel density functions for monthly precipitation amounts and average temperatures on dry and wet days are derived from historical data analysis.

Using the calibrated FS-WG model, a 64-year series of synthetic daily weather data was generated and compared with the historical weather records from weather stations in the study area. This comparison was achieved to assess whether the FS-WG model could produce synthetic weather data series with statistical characteristics similar to those of the historical data in the target region. The results, as shown in Figure 6, indicate that the synthetic daily temperature series generated by the FS-WG closely matches the actual monthly mean temperatures, with both the correlation coefficient and the Nash–Sutcliffe efficiency coefficient reaching as high as 0.9999. The monthly variance of the synthetic temperature series is generally higher than that of the actual series, indicating a slightly greater dispersion in the synthetic data, yet the similarity between them is still high, with a correlation coefficient of 0.9988 and a Nash–Sutcliffe efficiency coefficient of 0.9634. The synthetic daily precipitation series generated by the FS-WG showed relatively lower accuracy compared to the temperature series. The correlation coefficient between the synthetic and actual monthly mean precipitation was 0.9989, and the Nash–Sutcliffe efficiency coefficient was 0.9837. The correlation coefficient for the monthly variance between the synthetic and actual precipitation series was 0.9980, with a Nash–Sutcliffe efficiency coefficient of 0.9858. These results demonstrate that the FS-WG model is capable of generating synthetic daily weather data series that have similar statistical characteristics to those observed in the study area.

Using the FS-WG model calibrated and validated with historical observations (1957–2020), synthetic daily weather data series for the years 2021–2100 were generated to serve as baseline data. This baseline scenario maintains the original model parameters without any climate change adjustments, representing a hypothetical future where climate conditions remain similar to the historical period. In contrast, for climate change scenarios, the ensemble mean changes in monthly precipitation and average daily temperature relative to current levels, based on the six selected GCMs across four future time periods and four scenarios, were calculated to characterize future climate changes, as shown in

Table 5. Future climate change scenarios: temperature changes (°C, absolute difference from current) and precipitation changes (ratio to current).

Period	Scenario	Item	Jan.	Feb.	Mar.	Apr.	May.	Jun.	Jul.	Aug.	Sep.	Oct.	Nov.	Dec.
2030s	SSP1-26	temperatures	1.96	1.71	1.46	1.63	1.56	1.98	2.24	2.63	2.36	1.94	1.93	2.11
		precipitation	1.00	1.06	1.04	1.04	1.10	1.01	1.01	0.95	1.11	1.03	1.04	1.00
	SSP2-45	temperatures	1.87	1.42	1.54	1.62	1.58	1.95	2.13	2.46	2.38	1.76	1.74	1.88
		precipitation	1.00	1.04	1.05	1.05	1.06	1.02	1.04	0.98	1.05	1.04	1.02	1.02
	SSP3-70	temperatures	1.53	1.48	1.25	1.24	1.44	1.83	2.04	2.37	2.19	1.71	1.58	1.75
		precipitation	1.00	1.01	1.06	1.05	1.05	1.01	1.04	0.96	1.04	1.08	1.01	1.01
	SSP5-85	temperatures	1.82	1.85	1.47	1.59	1.77	2.08	2.32	2.67	2.42	1.90	1.95	1.98
		precipitation	1.00	1.03	1.07	1.02	1.06	1.04	1.04	0.97	1.08	1.06	1.03	1.02
2050s	SSP1-26	temperatures	2.27	2.33	2.12	1.91	1.96	2.55	2.91	3.30	2.99	2.20	2.20	2.43
		precipitation	1.05	1.08	1.08	1.09	1.09	0.99	1.04	1.02	1.15	1.14	1.08	1.02
	SSP2-45	temperatures	2.78	2.00	2.33	2.27	2.39	2.93	3.13	3.49	3.46	2.76	2.66	2.57
		precipitation	1.03	1.10	1.06	1.09	1.09	1.00	1.09	0.96	1.11	1.11	1.04	1.01
	SSP3-70	temperatures	2.69	2.31	2.14	2.13	2.40	2.99	3.17	3.59	3.43	2.75	2.59	2.63
		precipitation	1.00	1.06	1.12	1.07	1.06	1.01	1.06	0.98	1.08	1.09	1.02	1.00
	SSP5-85	temperatures	3.41	2.99	2.90	2.69	2.89	3.59	3.81	4.35	4.07	3.49	3.09	3.39
		precipitation	1.00	1.10	1.12	1.13	1.06	1.03	1.05	0.96	1.18	1.10	1.07	1.00
2070s	SSP1-26	temperatures	2.72	2.46	2.24	2.36	2.11	2.72	3.03	3.54	3.10	2.49	2.36	2.59
		precipitation	1.06	1.12	1.10	1.13	1.14	1.04	1.05	0.96	1.08	1.14	1.15	1.05
	SSP2-45	temperatures	3.31	2.81	3.03	2.97	3.02	3.71	3.94	4.56	4.29	3.54	3.26	3.32
		precipitation	1.04	1.12	1.09	1.10	1.12	1.01	1.05	0.96	1.14	1.14	1.11	1.05
	SSP3-70	temperatures	3.93	3.67	3.43	3.36	3.45	4.18	4.53	5.05	4.90	3.97	3.71	3.96
		precipitation	1.01	1.08	1.11	1.14	1.11	1.05	1.05	0.96	1.11	1.22	1.06	1.01
	SSP5-85	temperatures	4.68	4.22	3.98	3.95	4.29	5.12	5.47	6.30	6.02	4.77	4.70	4.70
		precipitation	1.15	1.20	1.20	1.18	1.11	1.05	1.04	0.96	1.13	1.29	1.10	1.08
2090s	SSP1-26	temperatures	2.61	2.54	2.31	1.92	2.11	2.70	3.01	3.20	3.02	2.31	2.36	2.54
		precipitation	1.03	1.11	1.11	1.12	1.09	0.99	1.04	1.06	1.14	1.12	1.10	1.06
	SSP2-45	temperatures	3.82	3.49	3.24	3.20	3.29	4.11	4.23	4.97	4.81	3.89	3.61	3.80
		precipitation	1.05	1.16	1.17	1.12	1.18	1.03	1.05	0.96	1.14	1.25	1.15	1.06
	SSP3-70	temperatures	5.04	4.98	4.61	4.46	4.60	5.46	5.72	6.47	6.53	5.43	5.21	5.17
		precipitation	1.10	1.15	1.20	1.15	1.16	1.07	1.09	0.95	1.19	1.27	1.14	1.04
	SSP5-85	temperatures	6.59	6.61	5.80	5.70	5.91	7.05	7.16	8.22	8.27	6.82	6.53	6.40
		precipitation	1.19	1.20	1.29	1.25	1.18	1.07	1.11	0.94	1.20	1.44	1.27	1.19

. For each SSP scenario, changes in monthly temperatures and precipitation rates for the four periods are used as scenario factors to adjust the probability density functions of the FS-WG model. This modification enables the generation of synthetic daily weather data series for the target study area from 2021 to 2100 under four SSPs climate change scenarios. These scenario-based data, along with the baseline scenario data, are then used as inputs for the McGANs model, which generates synthetic weather data series for the four weather stations in the study area under both baseline and future climate change scenarios. More details about the McGANs model application are described in Section 3.2.3.

3.2.3. McGANs for Multi-Site Weather Generation

Separate McGANs models were trained for temperature and precipitation, aiming to reach a Nash equilibrium where the loss functions converge around 0.5. This convergence point marks the conclusion of the training process, as illustrated in Figure 7. The McGANs model for temperature required 1005 epochs to converge, while the precipitation model converged in 830 epochs. More stringent early stopping conditions were applied to temperature training due to the more regular spatial distribution of temperatures across the four sites, which facilitated convergence. In contrast, precipitation’s spatial distribution exhibits greater randomness, leading to less stability in the loss functions during training. Excessive iterations could result in gradient explosion or divergence issues in the precipitation generator. Consequently, more robust generator hyperparameters and relatively simpler discriminator hyperparameters were set for precipitation generator to achieve a balanced architecture and ensure successful GAN training.

The main hyperparameters for both the generator and discriminator are listed in

Table 6. Summary of hyperparameters for the McGANs model.

Hyper-Parameters	Temperature		Precipitation
	Generator	Discriminator	Generator	Discriminator
Noise dimension	512	NA	512	NA
Batch size	5844	5844	2331	2331
Learning rate	0.0001~0.0005	0.0001	0.0001~0.0005	0.0001
Initial features	64	32	64	32
Early stop interval	0.491~0.509	0.491~0.509	0.483~0.517	0.483~0.517

. Using the trained McGANs model’s generator in inference mode, the output from the FS-WG model is input as conditional data to generate synthetic daily temperature and precipitation series for the four stations over the next 80 years under different SSPs climate change scenarios. These synthetic data series can then serve as input for the LSTM model to simulate daily streamflow at the watershed outlet under various scenarios, thereby assessing the impact of future climate changes on the watershed’s hydrological processes. Detailed descriptions of these simulations are provided in Section 3.3.

3.3. Future Streamflow Scenario Analysis

Using the synthetic daily weather data for the four stations generated by the McGANs model under various future climate change scenarios, the LSTM model was used to simulate daily streamflow under different SSPs scenarios, as illustrated in Figure 8. Overall, compared to the current baseline level, future climate changes are projected to have a positive impact on watershed streamflow yields. Under the SSP 1-26 and SSP 2-45 scenarios, the increase in streamflow is relatively modest, while under the SSP 3-70 and SSP 5-85 scenarios, the increase in streamflow is more significant. This suggests that proactive climate policies can positively influence the mitigation of climate change impacts on watershed hydrological processes and reduce variability in streamflow. Detailed discussions of the impact of future climate changes on streamflow can be found in Section 4.2.

4. Discussion

4.1. Models Accuracy and Reliability

This study applied a long short-term memory (LSTM) model to simulate watershed streamflow, demonstrating the effectiveness of using daily weather data from multiple sites to model river flow at the watershed outlet. According to the model results, the Nash–Sutcliffe efficiency (NSE) value of 0.79 indicates that the model accurately simulates streamflow dynamics and captures the overall patterns effectively. Compared to traditional hydrological models, the LSTM model proposed in this study shows unique advantages, especially in handling long-term dependencies and nonlinear time series data. The mean absolute percentage error (MAPE) of 81.10% during the validation period requires careful interpretation. This seemingly high value is primarily due to the nature of the MAPE calculation itself, where even small absolute errors during low flow periods can result in large percentage errors. For example, an absolute error of 1 m³/s would represent a 50% error when the observed flow is 2 m³/s, but only a 1% error when the flow is 100 m³/s. Additionally, this characteristic may be related to our choice of loss function during model training. When using mean squared error as the loss function, the squared differences between simulated and observed values during high flow periods may disproportionately influence the optimization process, as the same relative errors at higher flow values produce larger absolute differences and thus larger contributions to the loss function. This mathematical property of the loss function might lead to an optimization bias towards high flow periods. However, this is common in hydrological modeling and does not necessarily indicate poor model performance during low flow periods. The absolute errors remain within acceptable ranges across all flow conditions, and importantly, the predictions for low and medium flow periods do not show order-of-magnitude discrepancies, making the model suitable for analyzing future climate change impacts during dry seasons and assessing drought risks. The model shows consistent performance across different flow magnitudes, with particularly good results in capturing the dynamics of both high and low flows. This comprehensive performance is validated by comparing our results with related daily-scale flow simulation studies [46,47,48], demonstrating that our model achieves accuracy levels suitable for climate change scenario analysis. The LSTM model’s ability to capture both the overall patterns and specific flow characteristics makes it a reliable tool for assessing the impact of future climate changes on watershed hydrology.

The results of the FS-WG weather generator are comparable to those of similar weather generator models based on the concept of statistical downscaling [49]. For this study in particular, the averages of generated daily temperature are almost identical to the observed values, demonstrating high accuracy, a characteristic also observed in other weather generator models [50]. This similarity may be attributed to the relative regularity of temperature data, and because this study models different months separately, resulting in high consistency in monthly temperatures, which enhances the outcomes. The standard deviation of the generated daily temperature dataset is greater than that of the observed dataset, indicating that the data variability produced by the weather generator is higher, probably due to the use of the KDE method for sampling temperature values. The averages of daily precipitation data generated by the FS-WG are slightly higher than the observed data, with an overestimation more noticeable during the rainy season from May to July. This may be because the KDE sampling for wet day precipitation is set with a minimum limit of zero to prevent negative values, but without a maximum limit control. Thus, removing low negative values without similarly eliminating high values can shift the sampled data distribution towards higher values relative to the original distribution. However, overall, the consistency between the FS-WG weather generator’s simulation results and the original data is acceptable. In practical applications, the FS-WG is employed to generate synthetic weather data series for both an unchanged baseline scenario and various climate change scenarios. These synthetic series are then used to model future streamflow to estimate impacts of climate change on watershed hydrological processes. This approach effectively mitigates systematic errors introduced by the FS-WG, thereby enhancing the reliability of comparative scenario analysis.

In this study, the McGAN model was used to generate temperature and precipitation data for four stations based on regional average values as conditional numbers. Since the generator model was obtained through semi-supervised training, the results essentially represent one possible outcome that maintains similar spatial distribution characteristics to the historical temperature and precipitation distributions at these four stations. The output of the McGAN model is not a regression to a specific set of values and therefore cannot be evaluated using common metrics such as correlation coefficients and various errors. The primary purpose of using the McGAN model in this study was to generate input data for the LSTM model to simulate streamflow. Consequently, a streamflow prediction comparison experiment was designed to test the suitability of McGAN. The actual measurement group used real daily temperature and precipitation data from the four stations from 1957 to 2020 as input data for the LSTM model, simulating daily streamflow. In contrast, the simulation group used the average of real daily temperature and precipitation data from the four stations from 2006 to 2020 as conditional inputs for the McGAN model, generating synthetic daily temperature and precipitation data for these stations from 2006 to 2020, which was then used as input for the LSTM model to simulate daily streamflow. The consistency of streamflow results obtained using output from the McGAN model and original historical measured data as inputs for the LSTM model is shown in Figure 9. It was found that the Nash–Sutcliffe efficiency coefficient for the simulated and actual measurement groups was 0.92, and the linear correlation coefficient was 0.94. This indicates that the streamflow modeled using the four stations synthetic data generated by the McGAN model for LSTM model inputs are highly similar to those obtained using real weather data of the four stations as LSTM inputs. Thus, it can be concluded that using McGAN-generated data for four stations, based on regional average temperature and precipitation for LSTM input to simulate watershed streamflow, yields reliable results.

In summary, the results of the LSTM model, FS-WG, and McGAN model used in this study all fall within an acceptable range of accuracy. The modeling results obtained from these models are sufficient for estimating the impacts of future climate change on watershed streamflow, and the related response assessments are reliable. The feature analysis and discussions of climate change impacts on watershed streamflow are detailed in the following Section 4.2.

4.2. Impacts of Climate Changes on Streamflow

From 2021 to 2100, the average annual streamflow under various climate scenarios for each 20-year period (2030s, 2050s, 2070s, and 2090s) is shown in Figure 10. A notable trend of generally increasing streamflow is observed from the 2030s to the 2090s, suggesting that climate change may alter the hydrological cycle, potentially leading to increased streamflow yields. Compared to the baseline scenario, most climate scenarios show higher annual streamflow, particularly under SSP 3-70 and SSP 5-85 scenarios. This may indicate that high greenhouse gas emission scenarios could result in more frequent precipitation events in the future.

SSP 3-70 and SSP 5-85 consistently show higher streamflow than other scenarios across all periods, particularly in the 2070s and 2090s, indicating that under these high-emission scenarios, more precipitation events and increased flood risks may occur. Specifically, the 2030s reflect the changes in the near term, with streamflow slightly increasing across all scenarios relative to current levels. This trend may be attributed to the gradual rise in precipitation, coupled with the fact that temperature increases have not yet become pronounced. By the 2050s, streamflow further increases under all scenarios, especially in SSP 3-70 and SSP 5-85, where streamflows are significantly higher than those in the baseline and other lower emission scenarios (SSP 1-26 and SSP 2-45), indicating that the impacts of climate change are beginning to intensify, causing more precipitation and potential extreme hydrological events. In the 2070s, streamflow continues to rise across all scenarios, with notable increases in SSP 3-70 and SSP 5-85 compared to the 2050s, reflecting the long-term effects of climate change on the water cycle, including more frequent and intense precipitation events. By the end of the century, streamflow peaks in all scenarios, particularly under high-emission scenarios, indicating that the impacts of climate change on streamflow have reached their peak, potentially leading to more severe flood risks, especially under scenarios where greenhouse gas emissions are uncontrolled.

Average monthly streamflow under different climate change scenarios for each month is calculated as shown in Figure 11. The analysis indicates that all scenarios reach peak streamflow between May and September, suggesting that these months constitute the primary rainy season for the region. During these months, the SSP 3-70 and SSP 5-85 scenarios exhibit higher peak precipitation, which may be attributed to the more intense rainfall events projected under these more aggressive climate change scenarios. The baseline scenario generally shows the lowest streamflow, especially during the dry seasons (January to April and October to December), suggesting that future climate change scenarios could potentially increase streamflow, primarily concentrated in the rainy season from May to September. The SSP 1-26 and SSP 2-45 scenarios show mild increases in monthly streamflow compared to the baseline, reflecting the controlled greenhouse gas emissions under positive climate policies. In contrast, the SSP 3-70 and SSP 5-85 scenarios show substantially higher streamflow in almost all months, especially during the rainy season, reflecting the significant impact of uncontrolled greenhouse gas emissions on precipitation-streamflow patterns.

By identifying events of extreme high and low streamflow under different climate scenarios, and analyzing the frequency, magnitude, and temporal distribution of these extreme events, the study further discusses the differences in flood and drought risks under various future climate change scenarios. The 95th percentile is defined as the threshold for identifying events with a higher risk of flooding, and the 5th percentile as the threshold for identifying events with a higher risk of drought. The frequency of extreme events over the next 80 years under each scenario is shown in Figure 12. The results show that there are significant differences in the frequency of extreme events across scenarios. SSP 3-70 and SSP 5-85 exhibit a notably higher number of extreme high flow events compared to other scenarios, indicating a potentially higher flood risk under these scenarios. Conversely, the baseline scenario shows the highest number of extreme low flow events, suggesting that drought risks are more pronounced under current climate conditions, implying that future climate changes may mitigate drought risks. However, it is important to notice that the SSP 5-85 scenario also shows a higher number of extreme low flow events compared to other climate change scenarios, although less than the baseline level. This suggests that under uncontrolled future climate extremes, there might be a scenario where both flood and drought risks are elevated. This could be related to the coupling increase in precipitation and temperature under high greenhouse gas emission conditions: high precipitation may lead to an increased risk of floods during intense short-duration rainfall, while high temperatures may enhance evapotranspiration during dry seasons, elevating drought risks. Relatively, under proactive climate policies represented by SSP 1-26, a balance between flood and drought risks can be achieved, resulting in the lowest risk of extreme flow events.

4.3. Features and Limitations of the Study

The FS-WG-McGANs technique proposed in this study represents a novel approach for generating synthetic weather data series at multiple locations. This technique ensures that the generated weather data retain statistical characteristics similar to historical observations and simultaneously produces multi-site artificial weather data series with related attributes. The conditional generative adversarial network architecture guarantees that the data generated not only follow the overall regional averages (conditional numbers) but also reflect the spatial correlations (through the deconvolutional-convolutional neural network architecture) and randomness (through random noise inputs) between multiple sites. Compared to traditional single-location weather generators that iterate multiple times to produce results, the FS-WG-McGANs model provides more consistent multi-site data series on a daily scale, making it better suited for input in hydrological scenario analysis. Moreover, the inherent properties of generative models that blend rationale with randomness allow for a better representation of potential extreme meteorological conditions, thereby more comprehensively reflecting potential hydrological process impacts.

The limitations of this study primarily include the challenges in effectively using the FS-WG-McGANs model, the limited reliability of generated weather data, and uncertainties surrounding the accuracy of streamflow simulations produced by the LSTM model. First, training the McGANs model requires specific skills, especially for precipitation processes. This is partly due to the inherent limitations of the GAN method, which requires a delicate balance between generator and discriminator capabilities to prevent mode collapse, thus demanding considerable modeling expertise. Additionally, the spatial distribution of precipitation data is highly random. Temperature is mainly influenced by topography and underlying surface conditions, which are relatively fixed, making the spatial distribution of temperature data more regular and easier to model. In contrast, precipitation is mainly influenced by cloud cover, with less impact from topography and underlying conditions, making its spatial distribution less regular. This is also why we utilize generative models to address this issue, but it also presents challenges in training, with precipitation generators often achieving lower stability than temperature generators.

Secondly, FS-WG’s approach to generating future climate change scenario data is deficient in accounting for changes in precipitation patterns. The current assumption in FS-WG is that the mechanisms of regional dry–wet transitions remain constant, and future changes in monthly precipitation are mainly reflected in the daily precipitation amount on wet days. In reality, changes in monthly precipitation could result from maintaining the frequency of precipitation with more or less rainfall each time or from changes in the frequency of rainfall events (continuous rain or dry spells). The monthly-scale data from GCMs currently used lacks support for this analysis, and direct use of GCM daily data is challenging due to the low spatial resolution, which makes it difficult to reflect changes in precipitation frequency at the station scale. Future studies need to further refine the FS-WG model to address this issue.

Lastly, while this study focuses specifically on climate change impacts on streamflow, it is important to acknowledge that future hydrological changes will likely result from multiple interacting factors. These include land use changes, water resource management policies, human activities (such as water withdrawal and reservoir regulation), vegetation dynamics, and changes in watershed characteristics. The coupling effects between climate change and these factors present an important direction for future research. Additionally, the LSTM model used in this study is a classical deep neural network, and recent advances in machine learning offer potential improvements. Various novel technologies have been introduced to address meteorological–hydrological time series issues, including improved recurrent neural network models, combinations of RNNs and CNNs, transformer models focusing on spatiotemporal attention, hybrid models coupling data pattern decomposition with deep neural network predictions, and even applications of general large models on time series problems. These technologies have shown advantages in addressing specific challenges, and future research could explore state-of-the-art time series methods to establish more robust relationships between regional site-scale meteorological data and target river streamflow, potentially leading to more precise predictions and improved decision-making support for water resource management.

5. Conclusions

This study has developed an innovative LSTM-FS-WG-McGANs approach for assessing the impact of climate change on hydrological processes in medium to small-scale watersheds. The key conclusions are as follows:

(1)

The proposed FS-WG-McGANs method effectively generates synthetic multi-site weather data series that preserve both temporal and spatial characteristics of historical observations. The FS-WG model demonstrates high accuracy in reproducing historical patterns, with correlation coefficients and NSE values of 0.9999 for temperature and 0.9989 for precipitation. The McGAN model successfully maintains spatial consistency among stations, as validated by streamflow simulation comparison tests (NSE = 0.92; R = 0.94). This integrated approach enables the robust generation of daily temperature and precipitation data representing various climate change scenarios at multiple sites simultaneously based on ensemble outputs from multiple GCMs.

(2)

The LSTM model effectively simulates river discharge at the watershed outlet, achieving a Nash–Sutcliffe efficiency coefficient (NSE) of 0.79 on daily scale streamflow simulation during the validation period. While the model shows varying prediction accuracy across different flow magnitudes (MAPE = 81.10%), the absolute errors remain within acceptable ranges comparable to other daily-scale flow simulation studies. The model maintains reliable performance across both high and low flow conditions, making it suitable for comprehensive climate change impact assessment including both flood and drought risk analysis.

(3)

Under future climate scenarios, streamflow in the study area is projected to increase, particularly under high emission scenarios (SSP 3-70 and SSP 5-85). Quantitative analysis shows varying degrees of change across different SSPs: compared to the baseline scenario, SSP 1-26 and SSP 2-45 show mild increases in monthly streamflow, while SSP 3-70 and SSP 5-85 exhibit substantially higher flows, especially during the rainy season (May to September). The frequency of extreme hydrological events is likely to increase under high emission scenarios, with analysis indicating higher risks of both floods and droughts. However, proactive climate policies, such as those represented by the SSP 1-26 scenario, can effectively reduce future flood and drought risks in the Jing River watershed and decrease the probability of extreme streamflow events, serving as an effective approach to address the impacts of climate change.