Runoff Prediction in the Songhua River Basin Based on WEP Model

Abstract

Songhua River Basin, northeast China, has seen significant changes due to climate change and human activities from 1990 to 2000, when forests were largely reclaimed and agricultural land was taken up to change the terrestrial water cycle drastically. This paper investigates hydrological changes in three basins: the main stream basin of the Songhua River, the Second Songhua River Basin, and the Nenjiang River Basin. Machine learning and signal processing techniques have been applied to reconstruct historical river records with high accuracy, achieving determination coefficients exceeding 0.97. The physically based WEP model effectively simulates both natural hydrological patterns and human-induced hydrological processes in the northern Nenjiang region. Climate projections indicate clear temperature increases across all scenarios. The most significant warming is observed under the SSP5-8.5 scenario, where runoff increases by 8.52% to 12.02%t, with precipitation driving 62% to 78% of the changes. Summer runoff shows the most significant increase, while autumn runoff decreases, particularly in the Nenjiang Basin, where permafrost loss alters spring melt patterns. This change elevates flood risk in summer, with the rate of increase strongly dependent on the scenario. Water resources show strong scenario dependence, with the average growth rate of SSP5-8.5 being 4 times that of SSP1-2.6. A critical threshold is reached at a 2.5 °C increase in temperature, triggering system instability. These results emphasize the need for adaptation to spatial differences to address emerging water security challenges in rapidly changing northern regions, including nonlinear hydroclimatic responses, infrastructure resilience to flow changes, and cross-basin coordination.

1. Introduction

Runoff prediction plays a crucial role in water resource management and disaster mitigation. Recent advancements in machine learning (ML) techniques, including Random Forest (RF), XGBoost, and Long Short-Term Memory (LSTM) networks, have introduced a new paradigm for runoff modeling. These data-driven methods consistently outperform traditional hydrological models in areas such as gap filling, cross-basin generalization, and extreme event prediction. For example, XGBoost has demonstrated high accuracy in flood simulations for ungauged basins, while LSTM-based architectures offer higher spatiotemporal prediction capabilities when applied to large-scale hydroclimatic datasets. These developments underscore the potential of ML to address long-standing challenges in hydrological science, such as limited observational data and complex physiographic controls.

Recent advancements in artificial intelligence hydrological modeling have provided new opportunities to improve basin-scale runoff simulation and prediction. The encoder–decoder-based dual-level long short-term memory network (ED-DLSTM) has achieved an average NSE coefficient of 0.75 across over 2000 river basins in Europe and North America, with its spatial attribute encoding module effectively capturing hydrological regionalization effects and enhancing cross-basin prediction [1]. Similar success has been observed in ungauged basins, with NSE > 0 in 76.9% of 160 basins in Chile, confirming its parameter transferability [1]. AI-driven approaches also outperform state-of-the-art global flood warning systems, enabling reliable extreme runoff forecasts with up to 5-day lead times—critical for timely flood mitigation in vulnerable regions [2]. Moreover, hybrid models such as NowcastNet integrate physical evolution schemes with conditional learning, improving extreme rainfall–runoff prediction accuracy by 71% and better representing advective and convective processes [3]. These developments, together with random forest algorithms for reconstructing missing hydrological records, offer a robust foundation for this study, which applies machine learning techniques to reconstruct historical runoff and improve the understanding of climate- and human-induced hydrological changes in the Songhua River Basin [4]. XGBoost is a leading water resource model in 74% of water resources studies. It is highly effective for rainfall runoff modeling, streamflow prediction, and climate impact assessments. Its hybrid applications are especially promising for data imputation and hydroclimatic modeling, and it can be seamlessly integrated with process-based models like WEP for basin runoff projections under climate change scenarios [5]. Recent studies indicate that AI models such as LSTM and XGBoost outperform SWAT for rainfall runoff modeling (R² > 0.97) [6]. RF-based hybrid models improve streamflow prediction by integrating teleconnection factors and multi-station data [7]. Recent studies show that XGBoost is capable of predicting runoff for ungauged basins using multi-source data such as precipitation, vegetation cover, and land use. The model outperforms traditional HBV for flood simulation, which is particularly useful for handling heterogenous data. Data-driven approacesh provide valuable insights for the integration of process-based hydrological models with machine learning in snowmelt-dominated basins [8]. These developments illustrate the potential of combining process-based hydrological models and data-driven approaches to solve uncertainties in long term runoff projections. Urbanization and agricultural expansion can modify surface runoff generation and infiltration patterns. In the Danjiang and Khulna basins, urban land expansion increases runoff coefficients, while afforestation reduces peak flows [9,10]. In the SRB, rapid industrialization and cropland conversion have altered hydrological connectivity, but quantitative assessments of LULC impacts on basin-scale runoff remain limited [11]. Furthermore, climate change amplifies these effects by modifying precipitation regimes and snowmelt timing. CMIP6 projections indicate intensified rainfall variability and rising temperatures across Northeast Asia, potentially exacerbating flood risks and seasonal water scarcity [12]. Integrating CMIP6 scenarios into hydrological models is essential for future water security, yet current studies often overlook the synergistic effects of LULC and climate forcing [13]. Long-term observational data show that 21% of global river monitoring stations have experienced significant changes in seasonal runoff distribution (approximately two-thirds of the changes occur independently of annual discharge trends), which presents fundamental limitations for standard annual hydrological forecasting methods [14]. Climate model projections further indicate that predicted changes in daily rainfall intensity (with spatial scales comparable to annual precipitation variations) will directly alter runoff generation processes in 42% of global vegetated areas. These shifts in precipitation regime interact with cryospheric changes, increasing uncertainty in water resource availability [15].

A gap exists between spatial heterogeneity of hydrological processes and coarse resolution climate data. Distributed hydrological models, such as WEP, have demonstrated their robustness in modeling hydrological processes across different climate zones, including cold regions, by coupling water and energy balance [16]. The WEP model, developed by Yangwen Jia from 1995 to 2002, employs sub-basin-embedded elevation bands as computational units, making it suitable for large watershed simulations, such as in the Yellow River Basin [17,18]. The WEP model quantitatively analyses precipitation, evaporation and infiltration effects on runoff production. For example, in the Juma River Basin simulations runoff coefficient (R²) of 0.85 and RE) within 10% [19,20]. Coupling climatic factors (temperature, precipitation) can predict runoff evolution trends. For example, in the Yellow River source region (1956–2020), runoff showed weakly significant increase trend with increasing baseflow ratio [21,22].

Additionally, by incorporating both natural and artificial water cycles, a “dualistic water cycle” approach was applied [23,24]. Genetic algorithms, along with manual calibration, were employed to optimize simulation accuracy by adjusting aquifer thickness correction factors and soil hydraulic conductivity [23,24,25]. These efforts helped to assess the reduction in runoff from the Yellow River Basin (1960–2020) due to climate change and land use variations [22,26]. The calibrated model, based on historical data from 1980 to 2000, achieved a Nash-Sutcliffe efficiency (NSE) of 0.85 and a RE of 10% during the validation period (2001–2016) [19,20]. The WEP-QTP model, which incorporates active layer thickness dynamics in permafrost regions, was also used. In the Yellow River source area, increased frozen soil depth enhanced groundwater flow, while climate warming reduced surface runoff. However, permafrost degradation led to a decrease in subsurface flow and altered total basin runoff [21,22]. Furthermore, Visual MODFLOW was utilized to study the ecological effects of surface water and groundwater interactions [18].

The Songhua River Basin (SRB), the largest river system in Northeast China, is the third largest drainage basin in China. Climate change and human activities have significantly altered global hydrological cycles, presenting particular challenges for water resource management in large river basins [27]. LULC changes and increasing human water demands are ongoing in the basin [11]. Understanding these factors and their impact on runoff dynamics is critical for adaptive water resource planning. Additionally, conventional models struggle to address the spatial-temporal mismatches between CMIP6 outputs and local hydrological processes. Recent efforts to downscale CMIP6 data using RF bias correction show promise in comparing global climate projections with basin-specific hydrological responses [28]. This approach could enhance the reliability of runoff projections in the SRB, where temperature-driven snowpack variability primarily governs water availability. Most studies have focused on isolated factors, such as SWAT-based assessments of agricultural water use [29] or GRACE satellite evaluations of terrestrial water storage [30].

A comprehensive framework integrating the physical mechanisms of the WEP model, XGBoost-based data refinement, and CMIP6 scenarios is urgently needed to guide sustainable water management. This study contributes to the advancement of cold region hydrology through three key innovations: (1) Development of a WEP model variant (WEP-SRB) incorporating high-resolution permafrost thermal state modules and machine learning-optimized parameterization; (2) Quantification of climate change impacts across multiple temporal scales using dynamically downscaled CMIP6 projections; (3) Integrated assessment of natural and anthropogenic drivers through coupled human-natural system modeling. Our approach builds on recent breakthroughs in cryospheric hydrology [31], while addressing critical gaps identified in the IPCC AR6 regarding water security in cold [32]. The analysis spans the period from 1980 to 2100, including historical validation (1980–2014) and future projection (2026–2068) periods across four scenarios.

2. Data and Methods

The methodology of this study is organized into four primary components: (1) land use analysis, (2) completion of runoff data using the eXtreme Gradient Boosting Model, (3) establishment and calibration of the WEP model, and (4) calculation and trend analysis of runoff and total water resources under three future scenarios from CMIP6, as illustrated in Figure 1.

2.1. Study Area

The Songhua River Basin (41°42′–51°38′ N, 119°52′–132°31′ E), spanning an area of 557,000 km² in northeastern China, is characterized by diverse topography, ranging from the Changbai Mountains (2000 m elevation) to the Songnen and Sanjiang Plains [33]. The northern source, the Nen River, originates from Yilehuli Mountain in the Greater Khingan Range at a source elevation of 1030 m; it flows southward for 1370 km and drains a basin area of 298,500 km². The southern source, the Second Songhua River, rises from Paektu Mountain (Changbai Mountain) at a source elevation of 2744 m; it flows northwest for 958 km with a basin area of 73,400 km². These rivers converge at Sanchahe in Zhaoyuan County, Heilongjiang Province. Downstream of this confluence, the river is termed the Songhua River main stream, turning northeastward; it flows for 939 km, drains a basin area of 561,200 km², and empties into the Heilongjiang River (Amur River) at Tongjiang City [34]. This agriculturally vital basin supports over 60 million people while facing intensified hydrological extremes under climate change [35,36]. The temperate monsoon climate (Köppen Dwa) delivers 500–700 mm annual precipitation, with 70% concentrated in June-September. Extreme rainfall events have been increasing, with a rate of +0.12 days/year since 2000 [33,36]. Temperature gradients vary from 3 °C in the northern permafrost zones to 5 °C in the southern regions, contributing to a potential evapotranspiration rate of 800–1000 mm/year [33,35]. The river network, features regulated flows through critical infrastructure like Fengman Reservoir [33,37]. Historical records indicate strong runoff seasonality, with a wet/dry season ratio exceeding 3:1, and a baseline annual discharge of 76 billion m³. However, climate projections suggest a 9–15% reduction in snowmelt contributions by 2100 under SSP5-8.5 [33,38]. These multi-stressor conditions, coupled with projected temperature increases, 1.5–2.0 °C by 2050, necessitate advanced hydrological modeling to balance water demands across agriculture (70%), industry (20%), and municipal needs (10%) [35,38,39]. Figure 2 shows the location, topography, river network, hydrological stations, soil types, and land use of the Songhua River Basin, providing the spatial basis for the following hydrological and climatic analyses.

2.2. Data

Table 1. Data used for driving models and validating simulations.

Type	Data Source	Period	Resolution
DEM	Geospatial Date Cloud (https://www.gscloud.cn) URL (accessed on 1 July 2025)	-	30 m
Climatic forcing data	Institute of Tibetan Plateau Research Chinese Academy of Sciences (TPDC) (https://data.tpdc.ac.cn) URL (accessed on 1 July 2025)	1980–2022	Daily
CMIP6 meteorological data	The Earth System Grid Federation (ESGF) (https://aims2.llnl.gov/search/cmip6) URL (accessed on 1 July 2025)	1980–2068	Daily
Land use data	China Multi-period Land Use Remote Sensing Monitoring Data Set (CNLUCC) of the Chinese Academy of Sciences Resource and Environmental Science and Data Center (http://www.resdc.cn) URL (accessed on 1 July 2025)	1985–2023	30 m
Soil-type data	The second national soil census and the “Chinese Soil Records”	-	-
River discharge data	Actual measurement data from 12 hydrological stations in the Songhua River Basin, including Harbin, Jiamusi, Lanxi, Jilin, Fuyu, Dalai, Jiangqiao, Liujiatun, Dedu, Guchengzi, Yanqiao, and Nianzishan	2006–2022	Daily

presents the data utilized for the development and validation of the coupled energy-water model in WEP. The model construction and execution relied on remote sensing and weather station datasets, which included digital elevation models (DEMs), land use data, measured runoff from 12 stations, and daily records of precipitation, average temperature, wind speed, sunshine hours, and atmospheric relative humidity. Additionally, climate data for future scenarios from CMIP6 were incorporated. The Chinese Academy of Sciences (CAS) developed a national multi-temporal land use database at a 1:100,000 scale, based primarily on Landsat satellite imagery. Land cover classification was performed through manual visual interpretation using a two-level hierarchical system: six Level I categories (cultivated land, woodland, grassland, water bodies, built-up land, and unutilized land) further subdivided into 25 Level II classes. The database features a spatial resolution of 30 m.

2.3. Methods

2.3.1. Land Use Change

Land Use Dynamic Degree (LUD) Index

The Land Use Dynamic Index (LUD) is a key metric for assessing the intensity of temporal and spatial changes in land use, providing insights into the dynamic characteristics of regional land use through a quantitative analysis of land types, areas, and spatial expansion rates [40,41]. The CLUD index is the sum of the absolute dynamic degrees of all land use types within a specified area over a given period, offering valuable information regarding the extent of land use changes in the region. The calculation for the CLUD index is given by Equation (2).

$$K={\displaystyle \frac{U_b-U_a}{U_a}}\times {\displaystyle \frac{1}{T}}\times 100\%$$

(1)

$$LC={\displaystyle \frac{\sum _{i=1}^n∆L{U}_{i-j}}{2\sum _{i=1}^{n}L{U}_i}}\times {\displaystyle \frac{1}{T}}\times 100\%$$

(2)

where $K$ represents a specific land use dynamic; $U_a$ and $U_b$ are the initial and final areas of a given land use type; $T$ denotes the time period of change; $LC$ represents the comprehensive land use dynamic; $∆L{U}_{i-j}$ refers to the absolute value of the area of land class i converted to a non-$i$ land class during time $T$; and $L{U}_i$ indicates the initial area of land class $i$.

Land Use Transfer Matrix

The transition matrix for land use quantitatively assesses the reciprocal conversion of diverse land categories (e.g., agriculture, forest, urban areas) over the study period, elucidating the pattern and trajectory of their evolving dynamics [42,43]. Equation (3) embodies the following computational algorithm:

$$S_{ij}=\left[\begin{array}{ccc}{S}_{12}& \dots & S_{1n}\\ ⋮& \ddots & ⋮\\ S_{n1}& \dots & S_{nn}\end{array}\right]$$

(3)

Here, $S$ denotes the area; n represents the number of land use types before and after conversion; $i$ and $j$ ($i,j=1,2,...,n$) correspond to the land types before and after conversion, respectively; and $S_{ij}$ indicates the area of land type $i$ converted to land type $j$ prior the conversion.

2.3.2. The Extreme Gradient Boosting Model

The predicted value for an input sample $x_i$ is the weighted sum of the predictions from all trees:

$${\widehat{y}}_i=\sum _{k=1}^K{f}_k\left(x_i\right),f_k\in F$$

(4)

where $f_k$ denotes the prediction function of the $k$-th tree, and $F$ represents the set of all possible CART trees.

The objective function of XGBoost consists of two components: training error and regularization, as formulated below:

$$Obj=\sum _{i=1}^{n}L\left(y_{i,}{\widehat{y}}_i\right)+\sum _{k=1}^K\Omega \left(f_k\right)$$

(5)

where $Obj$ represents the total loss function, which reflects the model’s training error; $L$ is the differentiable loss function that quantifies the discrepancy between predicted and true values; $y_i$ is the true label of the $i$-th sample; $\Omega$ is the regularization term that penalizes model complexity to prevent overfitting; and $f_k$ is the $k$-th weak learner (a decision tree in XGBoost).

2.3.3. The WEP Model

Model Principle and Calculation Unit

The hydrological model software used in this paper is WEP1.0.7 version of watershed distributed binary water cycle model software developed by Institute of Water Resources, China Institute of Water Resources and Hydropower Research, distributed hydrological models simulate hydrological processes through spatial discretization, which accounts for spatiotemporal variations in precipitation, evapotranspiration, and runoff generation, offering more realistic representations compared to traditional lumped models. The model structure utilizes a nested sub-basin and elevation band framework, where each computational unit incorporates various land cover types (e.g., water bodies, bare soil, and vegetated zones). Surface heterogeneity is simulated using the “mosaic method” [23,44]. The partitioning process begins by dividing the sub-basin into elevation bands based on a predefined threshold. The final number of bands is determined by selecting the smaller value between this estimate and the maximum allowable band count. Grid cells are then sorted by elevation in ascending order and evenly divided into elevation bands, with each band covering a different elevation range. Vertically, the model divides the domain into a vegetation intercept layer, a soil layer (including bare soil, vegetated zones, irrigation cropland, non-irrigated cropland, and groundwater interaction layers, an aquifer layer for water interactions, and uses a moving wave or dynamic wave model for slopes and rivers, while the melting snow process is simulated using the degree-day factor method [19,23]. Runoff generation employs the Green-Ampt model to simulate infiltration capacity, and saturation-excess runoff occurs when soil moisture exceeds field capacity [19,20]. The model establishes upstream-downstream topological relationships between sub-basins using Pafstetter code for dynamic flow routing [45,46]. This framework is combined with spatial discretization and physically based process representations to improve hydrological simulation accuracy for complex watershed systems [16,47].

Surface runoff: The WEP-L model distinguishes surface runoff generation into two regimes based on rainfall intensity: storm periods and non-storm periods. During storm periods, lateral soil moisture movement is neglected, assuming dominant vertical infiltration. Surface runoff is then computed using the infiltration-excess runoff mechanism.

$$H_2-H_1=P-E-F-R_{surf}$$

(6)

$$R_{surf}=\left\{\begin{array}{c}0,H_2\le H_{max}\\ H_2-H_{max},H_2>H_{max}\end{array}\right.$$

(7)

where $R_{surf}$ is the surface runoff depth during storm periods (m); $H_1$, $H_2$, and $H_{max}$ represent the depression storage at the start and end of the timestep, respectively (m), the maximum depression storage capacity (m); $F$ corresponds to the cumulative infiltration (m); $P$ indicates precipitation (m); $E$ refers to evapotranspiration (m).

During non-storm periods, the vertical and lateral movement of soil moisture across different layers is considered, and saturation-excess runoff is calculated based on the water balance principle:

$$H_2-H_1=P\ast \left(1-{Veg}_1-{Veg}_2\right)+{Veg}_1\ast {Rr}_1+{Veg}_2\ast {Rr}_2-E_0-Q_0-R_{surf}$$

(8)

$$R_{surf}=\left\{\begin{array}{c}0, H_2\le H_{max}\\ H_2-H_{max}, H_2>H_{max}\end{array}\right.$$

(9)

where ${\mathrm{R}}_{\mathrm{surf}}$ represents the surface runoff depth during non-storm periods (m); ${\mathrm{Veg}}_1$, ${\mathrm{Veg}}_2$ denote the vegetation coverage of tall vegetation (forest) and short vegetation (grassland, crops), respectively; ${\mathrm{Rr}}_1$, ${\mathrm{Rr}}_2$ indicate the throughfall water volume from tall and short vegetation canopies to the ground surface (m); ${\mathrm{Q}}_0$ corresponds to the surface infiltration capacity (m); ${\mathrm{E}}_0$ refers to the evaporation from depression storage (m).

The calculation of total water resources in a narrow sense can be expressed by the following formula:

$$W_s=R_s+R_g$$

(10)

where $W_s$ refers to the narrow sense of water resources; $R_s$ represents surface water resources; $R_g$ denotes non-repetitive groundwater resources [48].

2.3.4. Methods of Evaluation

The eXtreme Gradient Boosting Model Accuracy Evaluation

Three evaluation indexes were used to evaluate the fitting degree between the simulated values of the random forest model and the measured runoff data. These metrics included the mean absolute error (MAE), root mean square error (RMSE), and the coefficient of R2. The mean absolute error quantifies the absolute difference between the predicted and the actual values, with minimal influence from extreme values [49]. The root mean square error evaluates the average deviation between the predicted and observed values [50]. The formula for the evaluation coefficient is as follows:

$$MAE={\displaystyle \frac{1}{N}}\sum _{i=1}^N\left|y_i-{\widehat{y}}_i\right|$$

(11)

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}\sum _{I=1}^N{(y_i-{\widehat{y}}_i)}^2}$$

(12)

where $N$ represents number of samples, $y_i$ denotes $i$-th true value, and ${\widehat{y}}_i$ refers to the $i$-th predicted value.

The WEP Model Accuracy Evaluation

This study assessed the applicability of the WEP model in the Songhua River Basin by using the correlation of coefficient (R²) and the Nash—Sutcliffe efficiency coefficient (NSE).

$$NSE={\displaystyle \frac{\sum _{i=1}^n{\left(Q_i^{obs}-Q_i^{sim}\right)}^2}{\sum _{i=1}^n{\left(Q_i^{obs}-Q_{mean}^{obs}\right)}^2}}$$

(13)

$$R^2={\displaystyle \frac{{\left[\sum _{i=1}^n\left(Q_i^{obs}-Q_{mean}^{obs}\right)\times \left(Q_i^{sim}-Q_{mean}^{sim}\right)\right]}^2}{\sum _{i=1}^n{\left(Q_i^{obs}-Q_{mean}^{obs}\right)}^2\times \sum _{i=1}^n{\left(Q_i^{sim}-Q_{mean}^{sin}\right)}^2}}$$

(14)

Here, $Q_i^{obs}$ represents the measured value, $Q_i^{sim}$ denotes the simulated value, $Q_{mean}^{obs}$ refers to the mean observed value, and $Q_{mean}^{sin}$ represents the mean simulated value.

Correlation Test

The Pearson correlation coefficient was employed to assess the relationship between runoff and meteorological factors in the Songhua River Basin [51]. This statistical measure quantifies the linear relationship between two random variables, denoted as r. There are two variables $x_1$, $x_2$, $x_n$ and $y_1$, $y_2$, $y_n$, and the correlation coefficient formula is:

$$r={\displaystyle \frac{\sum _{i=1}^n\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}{\sqrt{{\sum _{i=1}^n\left(x_i-\overline{x}\right)}^2}\sqrt{{\sum _{i=1}^n\left(y_i-\overline{y}\right)}^2}}}$$

(15)

The value of the correlation coefficient ranges from −1 to 1. A positive value indicates a positive correlation between the variables, with the correlation strengthening as the value approaches 1. Conversely, a negative value signifies a negative correlation, with the correlation intensifying as the value approaches −1. A correlation coefficient of 0 implies independence between the variables. To assess the statistical significance of the correlation, significance testing is required.

3. Results

3.1. Land Use Change Analysis

Quantitative analysis using Arc GIS10.6 revealed the extent and distribution of various land use types within the Songhua River Basin (

Table 2. Proportions of land use areas in different historical periods.

Land Use Type	1990		2000		2010		2020		2023
	Area (km2)	%	Area (km2)	%	Area (km2)	%	Area (km2)	%	Area (km2)	%
Cropland	242,849.82	44.17	253,487.84	46.11	245,176.16	44.59	254,337.91	46.26	256,661.45	46.68
Forest	247,375.42	44.99	239,351.96	43.53	240,103.73	43.67	236,843.73	43.08	233,568.31	42.48
Shrub	0.07	0.00	0.00	0.00	0.01	0.00	0.00	0.00	50.89	0.01
Grassland	43,616.03	7.93	39,424.24	7.17	42,914.14	7.81	33,759.21	6.14	33,854.25	6.16
Water	7700.71	1.40	6365.23	1.16	6843.97	1.24	7616.34	1.39	8280.23	1.51
Snow/Ice	0.02	0.00	0.01	0.00	0.09	0.00	0.19	0.00	0.22	0.00
Barren	3473.24	0.63	3427.09	0.62	3253.96	0.59	2429.08	0.44	1743.38	0.32
Impervious	9630.08	1.75	13,517.96	2.46	17,368.72	3.16	20,701.80	3.77	21,493.62	3.91
Wetland	1151.38	0.21	222.46	0.04	135.99	0.02	108.51	0.02	144.43	0.03

). The results indicate significant changes in land use patterns, with cultivated land and forest land accounting for more than 44% and 42% of the total area, respectively. Between 1990 and 2023, cultivated land exhibited a continuous increase, while forest land showed a decreasing trend. These changes can be attributed to the focus on agricultural expansion and government policies promoting agriculture development in the Songhua River Basin. The area of bare land has steadily declined. These trends are likely linked be agricultural reclamation and urban expansion within the region. From 2020 to 2023, the areas of water bodies and wetlands significantly increased, which is closely associated with national and local efforts to promote ecological restoration and strengthen ecological protection through territorial spatial planning in the Songhua River Basin.

The land transfer matrix was employed to examine the change patterns in the area of each land use type within the Songhua River Basin. This approach enables a quantitative analysis of the directions and magnitudes of land type transfers. To effectively illustrate the transfer dynamics across various land use categories and present the results of the transfer matrix,

Table 3. Land use change and dynamics within the study area.

Land Use Type	1990–2000		2000–2010		2010–2020		1990–2023
	Area Change (km2)	Single LUD Index (%)	Area Change (km2)	Single LUD Index (%)	Area Change (km2)	Single LUD Index (%)	Area Change (km2)	Single LUD Index (%)
Cropland	10,637.97	0.44	−8311.91	−0.33	9161.43	0.37	13,811.77	0.57
Forest	−8023.54	−0.32	751.88	0.03	−3260.05	−0.14	−13,806.89	−0.56
Shrub	−0.07	−9.63	0.00	13.33	−0.01	−8.57	50.82	6885.85
Grassland	−4191.65	−0.96	3490.10	0.89	−9154.58	−2.13	−9761.98	−2.24
Water	−1335.46	−1.73	478.72	0.75	772.37	1.13	579.49	0.75
Snow/Ice	−0.01	−5.50	0.09	106.67	0.10	10.48	0.21	114.50
Barren	−46.16	−0.13	−173.13	−0.51	−824.89	−2.54	−1729.88	−4.98
Impervious	3887.85	4.04	3850.73	2.85	3333.10	1.92	11,863.42	12.32
Wetland	−928.92	−8.07	−86.47	−3.89	−27.48	−2.02	−1006.95	−8.75
CLUD Index (%)	0.26		0.15		0.24		0.47

provides the dynamic land use change data, while Figure 3 shows the chord diagram of land use transitions. The analysis indicates that land use conversions were most pronounced between 1990 and 2000, with the least significant changes occurring from 2000 to 2010. Over the past 33 years, the cultivated land area expanded by 13,811.77 km², while the forest land area decreased by 13,806.89 km², and the wasteland area declined by 1729.88 km². Forest land, wasteland, and grassland were predominantly converted into cultivated land, with conversions of 24,590.7 km², 1301.85 km², and 13,644.9 km², respectively. In summary, the land use changes in the Songhua River Basin are clearly evident. Therefore, the natural runoff model should select the earlier land use type data.

3.2. Completed Measured Runoff

This study successfully reconstructed the daily runoff sequences from 1980 to 2022 for 12 hydrological stations in the Songhua River Basin using the XGBoost algorithm and the Savitzky–Golay high-precision conformal filter. This study established a second model using a linear regression model as a benchmark. The model outputs were followed by Savitzky–Golay filtering to smooth noise, ultimately enabling missing data interpolation and visualization. Figure 4 shows a fitted graph of the runoff data that was completed using a machine learning model. Figure A1 presents the linear regression model’s interpolated and completed runoff test results.

Table A1. A comparison table of advantages and disadvantages between RF + XGBoost (Savitzky–Golay) and Linear Regression (Savitzky–Golay).

Dimension	RF + XGBoost (Savitzky–Golay)	Linear Regression (Savitzky–Golay)
Principle	Integrated decision trees/gradient boosting	Ordinary least squares
Advantage	Nonlinear, interactive effects, high precision	Fast computation, interpretability, and non-overfitting.
Shortcoming	Black box, slow training, prone to overfitting	Limited by linear assumptions and low precision.

presents a comparison table of advantages and disadvantages between RF + XGBoost (Savitzky–Golay) and Linear Regression (Savitzky–Golay).

Table A2. Comparison table of precision between RF + XGBoost (Savitzky–Golay) and Linear Regression (Savitzky–Golay) models.

Model	Metrics	Haerbin	Jiasimu	Lanxi	Jilin	Fuyu	Dalai	Jiangqiao	Liujiatun	Dedu	Guchengzi	Yiandaqiao	Nianzishan
XGBoost Model	R2	1.00	1.00	0.99	0.97	0.99	1.00	1.00	0.98	0.98	0.99	0.99	0.99
	MAE	36.82	85.39	9.46	41.8	34.0	21.7	24.68	8.29	2.85	8.69	2.63	3.98
	RMSE	56.03	118.72	18.71	63.0	46.4	35.2	34.54	27.50	9.81	19.53	7.39	10.52
Linear Regression Model	R2	0.77	0.83	0.78	0.50	0.62	0.82	0.83	0.72	0.75	0.79	0.77	0.77
	MAE	355.05	570.02	65.09	186.19	187.21	202.66	205.15	48.40	18.18	53.82	15.39	15.39
	RMSE	492.43	784.75	116.01	273.69	275.8	315.7	322.22	92.68	34.39	94.08	31.42	31.42

compares the advantages and disadvantages between RF + XGBoost (Savitzky–Golay) and linear regression (Savitzky–Golay) models.

Table A3. WEP model site calibration and evaluation coefficient.

Station Name	Model	Period	R2	NSE
Jiamusi	Natural model	Calibration	0.85	0.71
		Validation	0.85	0.69
	Status quo model	Calibration	0.78	0.61
		Validation	0.82	0.65
Fuyu	Natural model	Calibration	0.87	0.79
		Validation	0.89	0.8
	Status quo model	Calibration	0.85	0.69
		Validation	0.87	0.82
Dalai	Natural model	Calibration	0.84	0.81
		Validation	0.89	0.85
	Status quo model	Calibration	0.83	0.74
		Validation	0.85	0.7

shows the precision comparison table between RF + XGBoost (Savitzky–Golay) and linear regression (Savitzky–Golay) models. Comparative analysis demonstrates that runoff data generated by XGBoost algorithm and Savitzky–Golay high-precision conformal filter exhibits higher fitting accuracy. Building on the demonstrated excellence of XGBoost and the Savitzky–Golay filter for runoff sequence reconstruction, the Random Forest (RF) algorithm was employed for robust data imputation to address missing values in historical runoff records. The Savitzky–Golay high-precision conformal filter was applied to refine the reconstructed runoff sequences by leveraging its dual capability for noise suppression and feature preservation. Operating as a local polynomial regression filter, it fits low-degree polynomials to moving subsets of the runoff time series within a defined window, optimally smoothing high-frequency noise while maintaining critical hydrological signatures. Its “conformal” attribute ensures minimal distortion of intrinsic runoff dynamics—preserving seasonal cycles, flood peaks, and abrupt irrigation-induced changes—by adaptively conforming to local curvature. Critically, the S-G filter’s edge-handling mechanism retained fidelity at sequence boundaries, ensuring reliability for trend and extreme value analysis across the 42-year reconstruction. Samples containing missing values were excluded to ensure training data integrity. RF model training utilized this cleaned data, configured with 100 decision trees and a fixed random state (seed = 42). The trained RF model was then applied to predict runoff across the entire temporal domain, generating imputed values. This ensemble regression tree approach provided high-accuracy estimates for missing daily runoff, preserving the statistical integrity and temporal dynamics (including seasonal patterns and extreme events) of the original reconstructed sequences. Model accuracy evaluations indicate that this approach exhibits remarkable adaptability and reliability in simulating complex hydrological processes. The coefficient of determination (R²) for all stations was greater than 0.97 (ranging from 0.973 to 0.998), with seven stations exceeding 0.99, indicating that the model could capture over 98% of the variability in the runoff time series. The exceptional performance at high-value sites such as Dalai (R² = 0.998) and Jiangqiao (R² = 0.998) demonstrates the high-precision mapping capability of XGBoost in capturing the nonlinear coupling mechanisms between precipitation and runoff. The collaborative application of Savitzky–Golay filters significantly enhanced the data smoothness, resulting in interannual RMSE values generally below 60 m³/s (e.g., Fuyu: 46.48, Lanxi: 18.71). The RMSE at Yiandaqiao was only 7.39 m³/s, and the MAE was as low as 2.63 m³/s, validating that this method effectively preserves seasonal fluctuations and sudden runoff events (such as flood peak timing) while removing high-frequency noise. The model performed consistently across heterogeneous subbasins, successfully capturing the bipeak characteristics of permafrost snowmelt runoff at the high-cold site of Nianzishan (R² = 0.986). The agricultural irrigation stations (e.g., Fuyu, MAE = 34.05) accurately restored daily flow step changes due to irrigation water diversions. At urban sites (e.g., Haerbin, RMSE = 56.03), the model effectively separated the interference of surface changes on baseflow under urbanization. Precipitation-driven verification from 1980 to 2004 revealed a high degree of consistency between the reconstructed sequence and observed data in terms of statistical distribution (K-S test p > 0.05), and annual extreme value synchronization (cross-correlation coefficient > 0.94), confirming the model generalization ability across climate fluctuation periods.

3.3. Results of Runoff Simulation

Among the 12 hydrological stations, Jiamusi Station, Fuyu Station, and Dalai Station were selected as control sections for the main stream basin of the Songhua River, the Second Songhua River Basin, and the Nenjiang River Basin. The simulation period spans 1980–2022 at a daily time step. The natural runoff model was calibrated using data from 1980–1990 and validated for 1990–2000. The current-condition runoff model was calibrated using the period 2000–2014 and validated for 2014–2022. This staged segmentation facilitates phase-specific assessment of runoff variations while mitigating parameter drift associated with prolonged calibration durations. The R² values for the natural runoff model in the Songhua River main stream basin were 0.85 for both the calibration and verification periods, the outcomes of the WEP model have been verified at the site and are reliable

Table A4. Correlation coefficient between runoff and meteorological factors.

Scenario	SSP1-2.6		SSP2-4.5		SSP5-8.5
Basin	Annual Average Temperature	Annual Precipitation	Annual Average Temperature	Annual Precipitation	Annual Average Temperature	Annual Precipitation
Songhua River main stream basin	−0.335 *	0.858 **	−0.177	0.846 **	0.104	0.898 **
The Second Songhua River Basin	−0.358 *	0.916 **	−0.077	0.882 **	0.032	0.906 **
Nenjiang River Basin	−0.326 *	0.847 **	−0.234	0.852 **	0.044	0.870 **

* Indicates passing the 0.05 significance level test, and ** indicates passing the 0.01 significance level test.

. WEP model site calibration and evaluation coefficient.

For the model in the main stream basin, the R² values were 0.78 and 0.82 for the calibration and verification periods, respectively, resulting in an overall R² of 0.80. In the Second Songhua River Basin, the R² values for the calibration and validation periods were 0.87 and 0.89, respectively. For the current runoff, the model’s calibration and validation periods had R² values of 0.85 and 0.87, yielding an overall R² of 0.86. In the Nenjiang River Basin, the natural runoff model achieved R² values of 0.84 and 0.89 for the calibration and validation periods, respectively. For the current runoff, the calibration and verification periods had R² values of 0.83 and 0.85, with an overall R² of 0.84. These results demonstrate that the WEP model is suitable for the Songhua River Basin after calibration. Figure 5 presents the simulation results for the calibration and validation periods of the current runoff in the Songhua River Basin.

3.4. Key Parameters of the Model

The input data for the WEP-L model include meteorological data, elevation information, land use, soil types, soil moisture characteristic parameters, leaf area index (LAI), and vegetation coverage. The key parameters of the WEP-L model comprise soil parameters, thermodynamic parameters, aquifer properties, riverbed hydraulic conductivity, and runoff routing parameters. Previous studies have conducted sensitivity analyses of the model parameters, showing that riverbed hydraulic conductivity, aquifer thickness, soil saturated water content, and aquifer hydraulic conductivity exhibit relatively high parameter sensitivity. Most of the model parameters do not require calibration; however, parameters with high sensitivity must be calibrated [52]. Model calibration indicated that the riverbed hydraulic conductivity adopted in this study is approximately 1.689 m/d. The hydraulic conductivity and specific yield of the aquifer were obtained through groundwater simulation calibration or geological survey data, with the hydraulic conductivity being 1.058 m/d and the specific yield 0.05 m/d. The following section presents the initial values assigned to several empirical parameters.

Soil moisture characteristic parameters influence soil evapotranspiration and the coupled water–heat transport processes. In this study, the soil profile was divided into multiple layers, and soils were categorized into four types based on their mechanical composition: sand, loam, silt loam, and clay. The characteristic water contents were then assigned according to soil classification [53]. The soil moisture characteristic parameters are summarized in

Table 4. Soil-moisture characteristics [54].

Parameter	Sandy Soil	Loamy Soil	Silt Loam Soil	Clay Soil
Saturated water content	0.4	0.466	0.475	0.479
Field capacity	0.174	0.278	0.365	0.387
Residual water content	0.077	0.120	0.170	0.250
Soil water suction at the wetting front (cm)	6.1	8.9	12.5	17.5
Mualem constant n	3.37	3.97	3.97	4.38
Saturated hydraulic conductivity (cm/s)	2.5 × 103	7.0 × 104	2.0 × 104	3.0 × 10−5

.

The Manning coefficient reflects the influence of the boundary surface on the resistance to water flow. The higher the value of the coefficient, the slower the convergence of the flow. The reference values of the Manning coefficient used in the model for different types of sub-surface are shown in

Table 5. The Manning coefficient of different underlying surface.

Parameter	Forest Land	Grass Land	Farm Land	Bare Land	Bare Rock, and Urban Surfaces	Water Bodies
Manning coefficient	0.3	0.1	0.2	0.05	0.02	0.01

. The factor of living day affects the changes in snow accumulation and melting, and also influences the volume of meltwater generated. The factor of living day used in the model for different types of underlying surface is shown in

Table 6. The Manning coefficient of different underlying surface.

Parameter	Forest Land	Grass Land	Urban Land Use	Bare Land	Snowfield
Degree-Day Factor (mm/°C/day)	1	2	5	3	1

.

The impact of adjusting parameters on the simulation effect: Generally speaking, the effects of adjusting various parameters on the simulation outcome are as follows:

• Aquifer Thickness Adjustment Factor: The aquifer thickness, defined as the vertical distance between the land surface and the underlying impermeable layer, determines the subsurface water storage capacity of a watershed. Modifying this coefficient primarily affects the peak flow and total volume of simulated runoff. Generally, increasing the aquifer thickness adjustment factor enhances groundwater storage, leading to a reduction in both the peak flow and the total volume of simulated surface runoff.
• Soil Layer Thickness: Adjusting this parameter predominantly influences simulated soil evaporation, vegetation transpiration, and runoff volume. Typically, increasing the soil layer thickness, particularly that of the uppermost layer, increases root-zone water storage capacity and consequently enhances evapotranspiration (ET), which reduces the volume of surface runoff.
• Stomatal Resistance Adjustment Factor: Modifying this coefficient mainly affects simulated vegetation transpiration and runoff. Generally, increasing the stomatal resistance adjustment factor restricts plant transpiration, thereby reducing total evapotranspiration and resulting in an increase in simulated runoff volume.
• Saturated Hydraulic Conductivity Adjustment Factor: Adjusting this coefficient primarily influences simulated infiltration capacity and peak runoff. Typically, increasing this factor enhances infiltration, thereby reducing surface runoff and peak flow. However, beyond a critical threshold, further increases can lead to rapid saturation of the soil profile. Under saturated conditions, infiltration capacity diminishes, paradoxically leading to an increase in surface runoff and peak flow.
• Streambed Hydraulic Conductivity Adjustment Factor: Modifying this coefficient chiefly affects simulated baseflow contribution to the stream channel. Generally, increasing the streambed hydraulic conductivity coefficient enhances groundwater exfiltration into the channel (or stream-aquifer exchange), resulting in increased baseflow and higher simulated discharge during dry periods.
• Aquifer Lateral Hydraulic Conductivity Adjustment Factor: Adjusting this coefficient primarily governs the rate of lateral groundwater movement within the aquifer. Typically, increasing the aquifer lateral hydraulic conductivity facilitates faster subsurface flow, leading to an increase in simulated baseflow contribution to the stream network.
• Surface Depression Storage Capacity: Adjusting this parameter primarily affects simulated surface runoff generation and volume. Generally, increasing the surface depression storage capacity allows for greater retention of rainfall and surface water, thereby reducing the volume and rate of surface runoff generation.

During the manual calibration procedure, the WEP-L model is executed by the software according to the predefined calibration scenarios and parameterization schemes to perform simulation computations. Upon completion of the simulation runs, an initial set of simulated runoff hydrographs is obtained. These initial results are subsequently compared against the observed streamflow records to initiate the iterative parameter adjustment process. The calibration typically commences with the adjustment of the Streambed Hydraulic Conductivity Adjustment Factor to achieve consistency between the simulated and observed baseflow recession characteristics during dry periods. Subsequently, the Saturated Hydraulic Conductivity Adjustment Factor is calibrated to align the magnitude and timing of simulated peak discharges with measured flood events. Should substantial discrepancies persist between simulated and observed hydrographs following these primary adjustments, further refinement is pursued through modification of the Aquifer Thickness Adjustment Factor and the individual thicknesses of the three soil layers to better represent subsurface storage and transmission processes. Additional parameters may be opportunistically adjusted based on systematic analysis of residual errors and their governing hydrological mechanisms to achieve optimal agreement between model predictions and measured watershed responses [55].

3.5. Future Trends in Temperature and Precipitation

The climate scenario data in this study utilizes the NASA Global Diagnosed Exchanges Daily (NEXGDDP-CMIP6) dataset. This down-scaled CMIP6 dataset contains historical benchmarks and future projections from 1950 to 2100 based on CMIP6 outputs, generated using the Monthly Bias Correction/Spatial Decomposition (BCSD) downscaling method with a spatial resolution of 0.25 degrees (25 km). The currently available products cover four emission scenarios (SSP 1-2.6, SSP 2-4.5, SSP 3-7.0, and SSP 5-8.5) across 35 global climate models. Each scenario represents a development model encompassing eight climatic variables. Released in June 2022, this dataset significantly enhances the assessment of climate change trends across various spatial and temporal scales, offering superior spatial and temporal resolution compared to other climate data products [56]. Data from three climate models (SSP1-2.6, SSP2-4.5, and SSP5-8.5) under the CMIP6 framework of the Beijing Climate Center BCC-CSM2-MR model were selected for analysis. The baseline period is 198–2014, and the future period (2026–2068) is considered for analysis.

Table 7. Changes in average temperature and precipitation in the Songhua River Basin under the SSP—RCP scenarios.

Factors	Basin	Songhua River Main Stream Basin		The Second Songhua River Basin		Nenjiang River Basin
Average temperature	Scenario	Annual average temperature (°C)	Increase speed (°C/10a)	Annual average temperature(°C)	Increase speed (°C/10a)	Annual average temperature (°C)	Increase speed (°C/10a)
	Historical Baseline Period	−0.134		1.16		−0.29
	SSP1-2.6	1.57	0.34	2.91	0.35	1.38	0.33
	SSP2-4.5	2.27	0.48	3.57	0.48	2.08	0.47
	SSP5-8.5	3	0.62	4.28	0.62	2.8	0.62
precipitation	Scenario	Annual Average Precipitation (mm)	Increase speed (mm/10a)	Annual Average Precipitation (mm)	Increase speed (mm/10a)	Annual Average Precipitation (mm)	Increase speed (mm/10a)
	Historical Baseline Period	537.79		619.85		474.16
	SSP1-2.6	542.54	0.95	622.15	0.46	486.81	2.53
	SSP2-4.5	569.62	6.37	648.7	5.77	508.42	6.85
	SSP5-8.5	586.98	9.84	680.43	12.11	517.32	8.63

and Figure 6 demonstrate that future temperatures under all three scenarios show a fluctuating upward trend, with the rate of increase varying by scenario and time period. Under the three radiation forcing scenarios, the multi-year average temperatures of all sub-basins exhibit a fluctuating upward trend, with the magnitude of warming positively correlated with the intensity of radiation forcing. The temperature increase is most significant under the SSP5-8.5 scenario (3.09–3.13 °C), followed by SSP2-4.5 (2.13–2.41 °C) and SSP1-2.6 (1.67–1.75 °C). Significant regional differences are observed. The main stream basin of the Songhua River exhibits the highest warming amplitude (1.7–3.13 °C) and the highest growth rate (0.34–0.63 °C/10a), indicating greater sensitivity to high-emission scenarios. The temperature increase in the second Songhua River Basin and the Nenjiang River Basin is similar (1.67–3.12 °C), but the growth rate in the latter is slightly lower under the SSP5-8.5 scenario (0.62 °C/10a vs. 0.63 °C/10a), which may be influenced by local climate regulation. Scenario dependence is prominent. Under SSP1-2.6, the growth rate is 0.33–0.35 °C/10a, approaching the global temperature control target threshold of 2 °C. Under SSP2-4.5, the growth rate increases to 0.47–0.48 °C/10a, indicating an intensification of regional climate risks. Under SSP5-8.5, the growth rate exceeds 0.62 °C/10a, far exceeding IPCC’s critical warning threshold of 0.5 °C/10a, potentially leading to irreversible ecological impacts. Future temperature changes in the Songhua River Basin demonstrate significant scenario-driven variability and spatial heterogeneity. Under high-emission scenarios, the warming rate is nearly twice the global average, underscoring the urgency of emission reduction efforts.

From Table 7 and Figure 6, it can be observed that, except for the Nenjiang River Basin under the SSP1-2.6 scenario, future rainfall in the other three sub-basins shows an increasing trend. The average increase in the SSP5-8.5 scenario is the largest (9.22%). The trend of rising water levels is most apparent under the SSP5-8.5 scenario, while the rainfall increase is the gentlest under the SSP1-2.6 scenario, similar to the SSP2-4.5. Under the SSP5-8.5 scenario, the probability of extreme precipitation events is much higher compared to the SSP1-2.6 and SSP2-4.5 scenarios, following a “high in the west and low in the east” pattern. The increase in the lower reaches of the main stream basin is greater than that in the upper reaches.

3.6. Changes in Future Runoff and Total Water Resources

3.6.1. Interannual Trends of Future Runoff Under Different Scenarios

As shown in the changing trends of

Table 8. Average annual runoff of Songhua River basin in historical base period and future scenarios.

Basin		1980–2014 Historical Baseline Period	2026–2068 Future Climate Scenarios
			SSP1-2.6	SSP2-4.5	SSP5-8.5
Songhua River main stream basin	Average Annual Runoff (m3/s)	1682.17	1706.25	1760.75	1825.45
	Change Values (m3/s)		24.09	78.58	143.29
	change rate (%)		1.43	4.67	8.52
The Second Songhua River Basin	Average Annual Runoff (m3/s)	543.95	553.51	571.85	609.35
	Change Values (m3/s)		9.56	27.90	65.40
	change rate (%)		1.76	5.13	12.02
Nenjiang River Basin	Average Annual Runoff (m3/s)	360.50	369.28	373.36	375.36
	Change Values (m3/s)		8.78	12.86	14.86
	change rate (%)		2.44	3.57	4.12

and Figure 7, the runoff in all sub-basins shows an increasing trend. The increase is significantly positively correlated with the intensity of radiative forcing. The increase in precipitation is the main driving factor (contribution rate 62–78%), and the rise in temperature leads to the advance of the snowmelt period (particularly obvious in the Nenjiang River Basin).

The baseline runoff in the main stream basin of the Songhua River is 1682.17 m³/s, with the greatest increase observed under the SSP5-8.5 scenario (8.52%). In the Second Songhua River Basin, the baseline runoff is 543.95 m³/s, and the relative change rate is the most significant, reaching 12.02% under the SSP5-8.5 scenario. Nonlinear growth is observed under SSP5-8.5, with ΔQ reaching 65.40 m³/s and the gain slope increasing by 134% compared to SSP2-4.5, indicating the presence of a climate-sensitive threshold effect. The runoff increase during the summer flood season (June to August) accounts for 78% of the annual total and is strongly linked to the intensification of the East Asian summer monsoon. The baseline runoff in the Nenjiang River Basin is 360.50 m³/s, exhibiting the smallest variation range (2.44–4.12%). Under the SSP5-8.5 scenario, runoff gain reaches a saturation point (with ΔQ of only 14.86 m³/s), and the difference between SSP2-4.5 and SSP5-8.5 is minimal (0.55%). Permafrost degradation in the Nenjiang River Basin contributes 37% to the increase in base flow under the SSP5-8.5 scenario, while in the main stream basin, the rainfall and flood processes dominate, contributing more than 82%. Under the SSP5-8.5 scenario, the coefficient of variation (CV) of runoff in all basins increases by 23–41%, indicating a higher likelihood of hydrological extreme events.

3.6.2. Trends of Runoff and Total Water Resources Under Different Scenarios

The simulated monthly runoff data from 2026 to 2068 were used to compare the annual distribution of monthly runoff in the three climate scenarios with the historical baseline flow. Figure 8 illustrates the annual variation of runoff in the future period relative to the historical baseline under the three scenarios. Under all three future climate scenarios, the basin’s overall summer runoff shows an upward trend compared to the baseline. This increase can primarily be attributed to the changes in summer temperature and precipitation. In contrast, monthly runoff during other seasons, particularly in autumn, is projected to be lower than historical levels, a trend most pronounced in the Nenjiang River Basin. This reduction is likely due to higher temperatures in the non-summer months, leading to increased evaporation and decreased surface and groundwater recharge. While the increase in summer runoff may provide a temporary boost in water resources, it may also heighten the risk of flooding. Conversely, the decrease in runoff during other seasons may exacerbate drought conditions, especially during the dry season.

For the snowmelt runoff period of the Songhua River Basin (March to May), we can observe that the runoff changes under the three future climate models are significant, and even smaller than the historical base period runoff volume. The main source of runoff in the Songhua River Basin from March to May is snowmelt runoff. Permafrost degradation weakens the function of the frozen soil layer in preventing precipitation infiltration. Under the SSP5-8.5 scenario, the spring runoff will instead decrease. However, under the combined effects of precipitation and temperature in the SSP1-2.6 and SSP2-4.5 scenarios, spring runoff is projected to gradually increase.

An analysis of

Table 9. The total water resources of the Songhua River Basin in the historical reference period and various future scenarios are averaged over many years.

Basin		1980–2014 Historical Baseline Period	2026–2068 Future Climate Scenarios
			SSP1-2.6	SSP2-4.5	SSP5-8.5
Songhua River main stream basin	Average Annual Runoff (1 × 108 m3)	510.14	521.57	543.48	553.01
	Change Values (1 × 108 m3)		11.43	33.34	42.87
	Change Rate (%)		2.24	6.54	8.40
The Second Songhua River Basin	Average Annual Runoff (1 × 108 m3)	286.86	291.36	302.15	319.01
	Change Values (1 × 108 m3)		4.50	15.29	32.15
	Change Rate (%)		1.57	5.33	11.21
Nenjiang River Basin	Average Annual Runoff (1 × 108 m3)	302.27	308.67	312.94	312.54
	Change Values (1 × 108 m3)		6.40	10.67	10.27
	Change Rate (%)		2.12	3.53	3.40

and Figure 9 reveals that the slope of the fitting equation in the main stream basin under the SSP5-8.5 scenario (3.29) is nearly 10 times higher than that under the SSP1-2.6 scenario (0.34). By 2068, the cumulative increase reaches 42.87 × 10⁸ m³, confirming the amplification effect of high radiative forcing on the transformation of precipitation into runoff. The intercept of the SSP5-8.5 equation (−6186.81) indicates an inflection point for runoff acceleration after the 2040s, likely linked to the threshold temperature for permafrost degradation, which causes a shift in the base flow replenishment mode. The increase in the change rate of the second Songhua River from SSP2-4.5 to SSP5-8.5 scenarios reached 5.88% (5.33%→11.21%), reflecting the vulnerability of the basin to severe convective weather. Its steep terrain further amplifies the nonlinear characteristics of the precipitation—runoff response. The slopes of the fitting equations in all scenarios exceed 1.0 (ranging from 1.32 to 1.50), indicating an increase in interannual variability, which may be synchronized with the occurrence of precipitation extremes driven by the intensified East Asian summer monsoon oscillation. The negative slope (−0.91) in the Nenjiang River Basin under the SSP1-2.6 scenario suggests an increase in the water-holding capacity of permafrost in this low-temperature scenario. However, historical baseline data indicate that, in the short term (2026–2040), the freeze–thaw stagnation effect persists, resulting in a pseudo-increase. The sharp rise in the slope under the SSP5-8.5 scenario (2.06) corresponds to large-scale permafrost degradation when the annual average temperature reaches ≥3 °C. The decrease in soil infiltration rate increases the efficiency of converting precipitation into surface runoff by 37–42%.

Spatial analysis shows that sensitivity increases from north to south. The runoff response coefficient per unit temperature rise in the Nenjiang River (0.68 × 10⁸ m³/°C) is lower than that in the main stream (1.12) and the Second Songhua River (1.45), supporting the buffering effect of forest coverage and soil layer thickness on hydrological regulation. When the average annual temperature rise in the basin exceeds 2.5 °C (around 2045 under the SSP5-8.5 scenario), the standard deviation of runoff change rates in the three basins increases to 2.3 times that of the baseline period, indicating a loss of system stability. A significant threshold effect is observed in the hydrological response of the Songhua River Basin. Scenarios above SSP2-4.5 will trigger irreversible changes in the runoff mechanism. Special above SSP2-4.5 will trigger irreversible changes in the runoff mechanism. Special attention should be given to the coupling of permafrost and hydrological processes, as well as the chain effect of extreme precipitation events.

3.6.3. Correlation Analysis Between Runoff and Temperature and Precipitation

Pearson correlation analysis between runoff and climatic factors in the Songhua River Basin (Table A4) reveals that precipitation exhibits a consistently significant positive correlation with runoff (p < 0.01, two-tailed) across scenarios, except for temperature under SSP5-8.5 which shows a negative correlation. Temperature under SSP1-2.6 shows a significant correlation at p < 0.05 (two-tailed). The strongest precipitation-runoff correlation occurs in the Second Songhua River Basin, while the weakest is observed in the Nenjiang River Basin. The negative temperature-runoff relationship under SSP5-8.5 likely arises because elevated temperatures enhance runoff through accelerated permafrost degradation and snowpack reduction. Precipitation remains the dominant driver of runoff, directly controlling snow accumulation and generation. Increased precipitation, particularly as snowfall, augments snowpack depth and volume, thereby supplying greater meltwater and increasing snowmelt runoff during spring and summer warming periods.

4. Discussion

The primary advantages include the seamless integration of machine learning (XGBoost) with process-based modeling (WEP), facilitating high-precision runoff reconstruction (R² > 0.97) and reliable scenario simulations (R² = 0.80–0.89). The multi-scenario approach effectively disentangled climate-land use-water resource interactions, particularly in permafrost-influenced zones. Spatial heterogeneity was explicitly addressed through sub-basin partitioning, enhancing regional applicability. Therefore, this article discusses machine learning, WEP models, and climate change.

4.1. Application of Machine Learning in Hydrological Modeling

The nonlinear and multi-factor coupling characteristics of hydrological processes, such as precipitation-runoff response, permafrost snowmelt mechanisms, and human activity interference, present significant challenges for the parameterization and process representation of traditional physical models. While models such as SWAT and HEC-HMS are based on well-defined physical mechanisms, their reliance on high-precision input data and the complexities of parameter calibration often result in cumulative errors, particularly in heterogeneous basins and during long-term simulations (e.g., baseflow biases under prolonged climate fluctuations). In contrast, this study highlights the significant advantages of data-driven methods for modeling complex hydrological systems. XGBoost, through the integration of regularization, second-order Taylor expansion, and parallel computing techniques, not only eliminates the traditional reliance on physical equations but also effectively captures the nonlinear mechanisms of multi-scale hydrological features via its gradient boosting framework. Reconstruction results reveal that extreme sites (Dalai, Jiangqiao) reached R² values of 0.998, indicating that XGBoost can map the nonlinear response of rainfall-flood events in a manner unattainable by hyperparameterized models. A cross-basin comparison with a study in the Xiaocheng Basin, Shandong, further confirms that XGBoost outperforms SWAT, achieving an NSE of 0.936 [57]. Notably, a hybrid framework combining the Savitzky–Golay filter with XGBoost was developed to address the impact of high-frequency disturbances in high-noise hydrological time series on machine learning models. The Savitzky–Golay filter preserves abrupt events and compresses the interannual scale RMSE to less than 60 m³/s (e.g., the minimum error is only 2.63 m³/s). This performance is significantly superior to the standalone machine learning model for long-term forecasting scenarios [58,59]. This innovation shows that the combination of signal processing and machine learning can overcome the limitations of data-driven models for modeling non-stationary time series. In complex scenarios such as the permafrost snowmelt bimodal process in cold regions (R² = 0.986 at the Nianzishan station) and abrupt flow events dominated by human activities (MAE = 34.05 at the Fuyu station), the model remains robust. Furthermore, the model excels in cross-climate period generalization, with the reconstruction sequence driven by precipitation from 1980 to 2004 showing a K-S test p-value > 0.05 and a cross-correlation coefficient for annual extremes > 0.94. These results confirm that XGBoost can internalize the long-term impact of climate fluctuations on runoff generation mechanisms, such as baseflow disturbance signal separation in Harbin due to urbanization. This complements the advantages of random forests in capturing extreme events in cold regions [57,60], underscoring the potential of ensemble learning algorithms to adapt to spatial heterogeneity and non-stationary hydrological processes. However, the study also reveals a common issue of performance degradation in machine learning models for ultra-long-term forecasting (e.g., >1 month) [57], suggesting the need for further exploration of integration physical constraints or dynamic feature selection mechanisms to enhance temporal extrapolation capabilities.

4.2. Distributed Hydrological Models

The effectiveness of distributed hydrological models in large and hydroclimatically complex basins hinges on their capacity to simulate spatiotemporally variable processes under both natural and human-influenced conditions. In this study, the WEP-L model was selected for its proven ability to simulate water-energy interactions, cryogenic processes, and coupled surface–subsurface flows with high fidelity. Model calibration and validation across three representative control sections—Jiamusi Station for the main stream basin, Fuyu Station for the Second Songhua River Basin, and Dalai Station for the Nenjiang River Basin—demonstrated strong simulation performance. For the natural runoff scenario, the WEP-L model achieved R² values of 0.85 in the Songhua main stream, 0.87–0.89 in the Second Songhua River Basin, and 0.84–0.89 in the Nenjiang Basin during 1980–2000. Under current conditions, calibration and validation during 2000–2022 yielded R² values of 0.80, 0.86, and 0.84, respectively, further confirming the model’s applicability and parameter stability across different temporal phases. These results are consistent with previous studies that demonstrated the model’s robustness across a range of hydroclimatic contexts. For instance, in Qinggang County, the WEP-L model effectively simulated cold-region field hydrology, capturing significant soil freeze depth reduction (~29%) and soil moisture increase (~6.8%) [24], which is essential for the cryogenic regions of the upper Songhua. In the Weihe Basin, WEP-L enabled fine-scale evaluation of water retention components, yielding values consistent with the InVEST model, thus supporting its use for ecosystem service quantification [61]. The model’s capacity to simulate coupled surface–groundwater dynamics has been verified through integration with Visual MODFLOW in the Mu Us Sandy Land, where it achieved R² values > 0.65 for surface flow and >0.58 for groundwater levels [18], suggesting strong potential for baseflow and recharge simulation in regions like the Nenjiang River Basin. Its spatial discretization flexibility—via hybrid sub-basin and grid schemes—has proven effective in mountainous–plain transitional areas, as shown in the Chongling catchment and Ju-ma River Basin [19,46], a useful feature given the topographic diversity within the Songhua Basin. Moreover, in the Miyun Reservoir catchment, the model enabled clear attribution of runoff changes, with ~57.2% explained by climate factors and ~42.8% by anthropogenic activities, highlighting its strength in scenario-based impact assessments [62]. In large river systems like the Yellow River Basin, WEP-L achieved calibration NSE > 0.7 and validation error < 3%, underscoring its reliability at basin scale [48]. The present study’s R² values of 0.80–0.86 for current runoff simulation across Songhua sub-basins are thus within the performance envelope reported in these studies. Collectively, these outcomes reaffirm the WEP-L model’s ability to represent diverse hydrological processes and support long-term runoff simulations under both natural and disturbed conditions in large river basins such as the Songhua.

4.3. Wep Model’s Broader Applications

Although the WEP model demonstrated robust performance in the snowmelt-dominated Songhua River Basin, its direct transferability to other hydroclimatic and anthropogenic contexts (e.g., tropical, arid, or urban catchments) requires targeted modifications. Model structure and parameterization should be adjusted to capture the dominant hydrological processes in these regions.

In tropical basins, where snow and frozen soil processes are negligible, the snowmelt and freeze–thaw modules can be simplified or removed. Emphasis should be placed on evapotranspiration and infiltration processes, requiring recalibration of parameters controlling potential evapotranspiration, soil hydraulic conductivity, and root-zone storage capacity. In arid and semi-arid basins, WEP’s groundwater and baseflow modules should be strengthened to reflect deep percolation and groundwater recharge dynamics under sparse precipitation. Parameter calibration should focus on soil moisture retention characteristics (field capacity, wilting point) and plant water use efficiency, as they strongly influence runoff generation in water-limited environments [63]. For highly urbanized catchments, impervious surfaces and storm-drain routing processes need to be explicitly represented. This involves introducing imperviousness fraction, depression storage, and drainage network routing parameters to capture rapid surface runoff response and peak discharge timing; similar needs are noted in urban flood modeling studies [64,65].

Refined land use classification is critical for ensuring model applicability across different regions. In tropical areas, distinguishing evergreen and deciduous vegetation improves the simulation of seasonal evapotranspiration dynamics. In arid regions, bare land and sparse vegetation must be explicitly represented to simulate infiltration-excess runoff and wind erosion-related hydrological responses. In urban areas, detailed representation of impervious surfaces, green spaces, and pervious pavement systems can improve the prediction of runoff volumes and timing under high-intensity rainfall. Studies in semi-humid/semi-arid basins of China using WEP (e.g., Jing River Basin) show that land cover change markedly alters runoff partitioning [66]. Also, the urban river regulation case in northwest China demonstrates the importance of anthropogenic land cover and pollution load in model structure [67].

Multi-objective calibration should be adopted when transferring the model to new regions, including not only streamflow volume but also peak flow timing, low-flow conditions, and flow variability indices. Sensitivity analysis should be performed to identify parameters that exert the highest influence in the new hydroclimatic regime. Prior work developing the WEP model in Qinghai–Tibet (soil-gravel structure) achieved Nash–Sutcliffe efficiency > 0.75 and |RE| < 10% by careful soil/gravel/groundwater module calibration, which supports that structural parameterization is crucial for transferability [63]. Also, real-time or post-event urban flood modeling (e.g., DRIVE-Urban, Chen Et Al.) provide methods of calibrating peak flow & timing in urban catchments [64].

A number of studies have applied the WEP model or its modified variants across diverse hydroclimatic regions, which provides valuable insights into its transferability and the necessary structural or parameter modifications.

Qinghai–Tibet Plateau (Cold Alpine Basin), Wang et al. [63] developed an improved distributed hydrological model incorporating a soil–gravel structure module to simulate runoff in the Niyang River Basin on the Qinghai–Tibet Plateau. This region is characterized by high altitude, coarse gravelly soils, and strong snowmelt and groundwater interactions. The study demonstrated that explicitly representing the soil–gravel structure significantly improved model performance, achieving NSE > 0.75 and relative error below 10%. The results highlight that groundwater recharge and delayed baseflow are key components in alpine basins, suggesting that WEP’s groundwater module should be emphasized when applied to similar cold, coarse-textured catchments.

Jing River Basin, Loess Plateau (Semi-Arid Region), Su et al. [66] applied the WEP model to the Jing River Basin, a semi-arid catchment with strong precipitation seasonality and severe soil erosion. The model was used to assess how land use change and precipitation variability affected runoff over multiple decades. They found that conversion of cropland to forest and grassland reduced surface runoff and increased infiltration, while decreasing peak discharges. Their results demonstrate that accurate representation of land-cover dynamics and soil infiltration capacity is critical for simulating runoff generation in semi-arid basins, reinforcing the need for careful land-cover classification and parameter calibration when transferring WEP to similar regions.

Great Lakes Depression, Mongolia (Cold Region Lateral Flow), Dorjsuren et al. [68] enhanced the WEP model to better estimate lateral subsurface flow in the Great Lakes Depression of Mongolia, a region dominated by permafrost and snowmelt. They introduced a lateral flow algorithm linked to soil moisture gradients and frost depth, which improved the timing of spring peak flow and reduced bias in low-flow periods. This work underlines the importance of explicitly simulating lateral flow and frozen-soil hydrological processes when applying WEP to cold regions with seasonal frost and permafrost.

WEP-COR Model in a Cold Region (Energy–Water Coupling), Jia et al. [69] developed WEP-COR, a variant of WEP with coupled land-surface energy balance to simulate water and energy budgets in a cold region. The model explicitly solves the surface energy balance and soil heat conduction equations, allowing it to simulate ground temperature, snowmelt timing, and evapotranspiration more accurately. Their evaluation showed improved simulation of snowpack dynamics and evapotranspiration compared to the original WEP, suggesting that energy budget coupling is necessary for robust performance in regions where snow and freeze–thaw processes dominate.

Heavily Polluted Urban River, Northwest China (Human-Impacted Basin), Lyu et al. [67] applied a distributed hydrological model similar to WEP to a polluted urban river system in northwest China, incorporating anthropogenic water regulation such as wastewater discharge, water diversion, and reservoir operation. The study successfully reproduced streamflow dynamics under highly regulated conditions and provided a basis for designing water regulation strategies to improve water quality. This application illustrates that for urban or human-impacted basins, WEP should include modules for artificial water transfers, drainage, and pollutant transport to capture the full water cycle.

4.4. Future Climate Change Hydrological Impacts

The CMIP6 and BCC-CSM2-MR models utilize the BCC (Beijing Climate Center) climate model from the CMIP6 version, which offers high resolution and the capability to simulate regional climate features. This model is based on the coupling of global climate models (GCMs) and regional climate models (RCMs), producing spatiotemporal data on future temperatures, precipitation, and other parameters [70,71]. BCC-CSM2-MR from CMIP6, along with three SSP scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5), was selected for this study. This model has been shown in multiple studies to outperform earlier models across East Asia in key climatic metrics such as seasonal precipitation, temperature trends, and monsoon dynamics. For example, Wu et al. (2019) show that BCC-CSM2-MR reduces bias in East Asian summer monsoon rainfall, improves the precipitation diurnal cycle and tropospheric circulation relative to BCC-CSM1.1m [72]. D. Wang et al. found in comparative evaluations of CMIP5 vs. CMIP6 over China that BCC-CSM2-MR achieved among the highest skill scores for monthly precipitation distributions and temperature climatologies in northern China [73]. Liu et al. also confirm BCC-CSM2-MR and its variants (including the MR and HR resolutions) perform well in simulating summer surface air temperature and precipitation in East Asia, including better match to observed correlation coefficients for spatial pattern [74,75]. These improvements give confidence that forcing data from BCC-CSM2-MR are relatively credible for the Songhua River Basin region. The three SSP scenarios were chosen to bracket a broad span of possible radiative forcing trajectories (low, mid, high), thereby enabling the capture of non-linear hydrological responses (e.g., runoff, seasonal shifts, flood risk) under increasing forcing. However, using only a single GCM means that structural uncertainties across different climate models are not represented. There is evidence that GCM differences often contribute substantially to projections of rainfall seasonality and extremes, sometimes even more so than choice of SSP alone [76]. Projections of interannual precipitation variability under global warming further show that with one model or one large-ensemble, internal variability and model structural differences still dominate the uncertainty envelope [77]. In addition, downscaling and bias correction methods can alter future precipitation extremes and seasonal distribution, which affects hydrological responses (e.g., flood peaks vs. low flows) more than mean climate changes in many cases. Thus, while results such as an 8.52–12.02% increase in runoff under SSP5-8.5 are plausible within the BCC-CSM2-MR framework and useful for assessing relative scenario dependence, they should be interpreted as one possible trajectory rather than a bound. Because BCC-CSM2-MR has already demonstrated improved climatological fidelity (e.g., reduced annual precipitation bias over China, improved atmospheric circulation fields, better representation of seasonal and intraseasonal variability) [72,73], the choice is justifiable for the goals of this study (tracing the hydrological sensitivity under differing forcing levels). The selected radiative forcing scenarios are SSP1-2.6 (low radiative forcing), SSP2-4.5 (medium radiative forcing), and SSP5-8.5 (high radiative forcing), with the intensity of radiative forcing positively correlated with future climate responses. For example, under the SSP5-8.5 scenario, the temperature warming rate in the Songhua River Basin is nearly twice the global average, and the frequency of extreme precipitation events significantly increases [71,78]. The BCC-CSM2-MR model performs better in regional climate simulations, with precipitation prediction errors reduced by 10–15% compared to CMIP5 [70,79]. R2 value of the BCC model historical precipitation simulation for Songhua River Basin is 0.92. Other GCM/RCM combinations [79] will play a role in the future (2026–2068), the average temperature in Songhua River Basin will be fluctuating upwards. Under the SSP5-8.5 scenario, winter temperatures will rise by 3.0 °C, and summer temperatures will increase by 2.5 °C [78,80]. Higher temperatures will lead to increased potential evaporation, which could exacerbate water loss. Under the RCP8.5 scenario, runoff decreases by about 15% for every 1 °C increase in evaporation in the main stream of the Songhua River [80,81]. Future precipitation will be uneven, with increased rainfall in spring and winter (spring precipitation increases by 12.5%). However, the frequency of summer precipitation decreases, while its intensity increases [82,83]. Under the SSP5-8.5 scenario, the average annual runoff depth in the main stem of the Songhua River may increase by 10–28%, but under high emission scenarios, evaporation results in reduced runoff [71]. Under the low emission scenario (SSP1-2.6), the risk of drought in the Songhua River basin is reduced by 18%, while under the high emission scenario (SSP5-8.5), the probability of extreme precipitation increases by 32% [71,78]. The increase in winter temperature shortens the freezing period by 1–2 months, which may affect the water transport capacity of the river. The main stream of the Songhua River basin is more sensitive to high emission scenarios, with its warming rate positively correlated with radiative forcing intensity. The Nenjiang River basin, influenced by local climate regulation, has a lower warming rate, but increased precipitation may exacerbate soil erosion.

Runoff changes mainly follow precipitation patterns. A significant correlation exists between runoff and precipitation (p < 0.01). The Second Songhua River shows the strongest precipitation-runoff relationship. Special conditions arise under the SSP5-8.5 scenario, where higher temperatures are linked to increased runoff. This suggests that permafrost melting affects water flows. The results help understand water resource changes. They show different impacts under various climate futures. Management plans should take these scenario differences into account. To enhance flood control capabilities, resilient water systems should be designed by incorporating peak runoff data simulated using the SWAT model. Surface evaporation can be reduced by altering vegetation cover (e.g., forest expansion), but a balance between carbon sinks and hydrological needs must be maintained. Utilize climate parameters predicted by the BCC model to establish a real-time hydrological response monitoring system. For example, under the SSP2-4.5 scenario, strengthen summer runoff forecasting and emergency response.

4.5. Uncertainties of the Study

The climate forcings used in this study are drawn from the CMIP6 archive and therefore represent large-scale climate projections rather than local water-management forcings such as irrigation expansions and reservoir operations. CMIP6 models do not uniformly represent irrigation and reservoir management; irrigation forcing is explicitly represented only in a small subset of CMIP6 experiments, so use of raw GCM output without an ensemble-based treatment and without explicit representation of anthropogenic water management biases the hydrological response to large-scale climatic drivers [84,85].

Aggregating surface-water and groundwater volumes without explicit two-way process coupling omits dynamical feedbacks that control evapotranspiration, soil moisture and baseflow. Fully coupled surface–subsurface modeling has demonstrated that groundwater dynamics exert a first-order control on land-surface water and energy exchange and on baseflow residence times; therefore assessment of total water resources that treats surface and groundwater as independent reservoirs underestimates the role of subsurface–surface interactions in both high-flow and low-flow regimes [86,87].

Uncertainty in model parameters (for example soil hydraulic parameters, aquifer transmissivity and specific yield, and streambed conductance) is a principal source of simulated runoff variability and must be quantified explicitly. Bayesian calibration and Monte-Carlo frameworks (e.g., DREAM, GLUE and related MCMC approaches) are standard and effective tools to estimate posterior parameter distributions and predictive uncertainty for hydrological models; these methods also separate forcing, parameter and structural uncertainty and thereby support robust uncertainty attribution [88,89].

Global (variance-based) sensitivity analysis is required to identify the subset of parameters that dominates uncertainty in key hydrological targets (peak flow, low flow, seasonal volume). Variance-based indices (Sobol’/Saltelli family) and current hydrology-specific reviews provide a reproducible workflow for (1) defining parameter ranges based on measurement/field constraints, (2) propagating parameter uncertainty via Monte Carlo sampling, and (3) ranking parameters by first-order and total-effect sensitivity indices to prioritise calibration and field campaigns. Prior regional studies confirm that climate and parameter sensitivities can be large (for example, runoff sensitivity to temperature and precipitation in the Songhua River basin reaches on the order of ±30% under certain conditions [85,90]); thus targeted global SA is a necessary precondition for defensible model calibration and uncertainty interpretation [90,91].

Machine-learning components (here, XGBoost) require probabilistic validation and interpretability analysis in addition to point-error metrics. Robust ML validation employs ensemble strategies (k-fold cross-validation, bootstrap aggregation, stacking/ensembling) to quantify generalization uncertainty and to produce prediction intervals; post hoc explainability methods (for example SHAP values) provide consistent, quantitative feature-attribution that links predictive importance to hydrologically meaningful drivers such as precipitation, temperature and land-use fractions. Interpretable ML (SHAP/XGBoost) therefore bridges predictive skill with process-level attribution and supports physically informed scenario analysis [92,93].

Evaluation of hydrological performance must extend beyond R², RMSE and MAE. Diagnostics should include Nash–Sutcliffe efficiency (NSE) and its logarithmic variant for sensitivity to low flows; Kling–Gupta efficiency (KGE) and its component decomposition to expose bias, variability and correlation errors; percent bias (PBIAS) and volumetric error for water-balance consistency; peak-flow timing and magnitude diagnostics; flow-duration curve comparisons; extreme-event indices (e.g., R95p, FHV) for tails; and strictly proper scoring rules such as the Continuous Ranked Probability Score (CRPS) when probabilistic forecasts are produced. Using this suite of complementary metrics prevents over-reliance on a single statistic and provides transparent, process-relevant evidence for model adequacy [94,95].

Present a systematic uncertainty framework for this study. The framework comprises: (1) ensemble climate forcing from multiple CMIP6 realizations with bias-correction and model-weighting following multi-model best-practice; (2) Bayesian/Monte-Carlo parameter calibration to obtain posterior parameter distributions and predictive intervals; (3) global sensitivity analysis (Sobol’/Saltelli and related methods) to prioritize parameter calibration and field measurements; (4) ensemble ML validation and interpretable feature-attribution (SHAP) to couple ML predictive skill with hydrological drivers; and (5) an expanded set of hydrological performance diagnostics (NSE, log-NSE, KGE, PBIAS, peak/timing metrics, FDC and CRPS) for both deterministic and probabilistic outputs. Each component of this framework is referenced to established methodological literature to ensure reproducibility and to enable direct comparison with similar basin-scale assessments [85,88].

4.6. Policy and Water-Resources Management Implications

The observed and projected runoff changes in the Songhua River Basin must be interpreted in light of China’s established water laws, basin governance structures, and recent management practices. The following implications show how the quantitative results of this study align with existing policy instruments and legal frameworks.

Flood Control Law and Basin-Scale Flood Planning

The Flood Control Law of the People’s Republic of China provides the legal basis for formulating flood control plans for major river basins and coordinating flood-control infrastructure across administrative boundaries. Under Article 9–10, flood control plans must be aligned with comprehensive basin plans and involve provincial governments, river basin authorities, and the State Council [96,97].

Our finding that summer runoff and flood peaks increase markedly under higher radiative forcing (e.g., in SSP5-8.5) underscores the necessity of these legally required basin-wide planning instruments to manage infrastructure (reservoirs, detention basins, dikes) and operations in ways that both protect human settlements and maintain ecological flows. The institutional mandate in law to integrate engineering and non-engineering flood control (e.g., detention basins, flood storage) as well as upper and lower reach coordination is directly relevant [97,98].

River Chief System and Integrated Governance

The River Chief System has been empirically shown to improve water environment and governance outcomes by clarifying responsibility across administrative levels, promoting performance assessments, and transcending fragmented water administration. For instance, Evaluating the Effectiveness of the “River Chief System” shows that after implementation, there is measurable improvement in water quality and accountability [99].

In relation to our results, the increase in runoff variability and flood risk demands precisely that kind of integrated and accountable governance. The nonlinear and threshold responses (e.g., strong sensitivity beyond ~2.5 °C warming) imply that fixed or static management regimes will be insufficient; the River Chief System’s structure supports dynamic regulation, particularly with respect to flood and drought extremes, which our projections indicate will become more pronounced.

National Flood Control & Drought Relief Emergency Response Plan

China’s central government has also established Emergency Response Plans for Flood Control and Drought Relief that outline institutional response during extreme hydrological events. These plans facilitate coordinated operations among water administration, meteorological services, and local governments.

Our projection—summer increases in runoff, autumn decreases, and potential low-flow stress—aligns with conditions under which these emergency plans are activated. The projected increased coefficient of variation in runoff supports needing robust, legally backed emergency response protocols that can mobilize resources and adjust operations quickly.

Policy Implications for Water Resource Allocation and Efficiency

Existing policies relevant here include water allocation regulation, water-saving actions in agriculture, and land use conservation. While specific central legislation on water allocation under variable climate is less directly addressed, the institutional policy framework (including IWRM principles under Water Law and local water resource plans) already provides for regulating abstraction and promoting efficiency. Review of China’s River Chief System and The Effectiveness of “River Chief System” Policy studies show that governance reforms have improved pollution control and allocation oversight [99,100].

The projected decline in non-summer seasonal flows (especially autumn and spring under certain scenarios) indicates that during low-flow periods, existing allocation systems will be tested. Policies that limit water use or adjust allocations seasonally are embedded in Water Resource Plans; our quantitative estimates could inform those plans to ensure allocations are consistent with hydrologic change [97].

Soil and Land Conservation Laws, and Watershed Management

The runoff changes driven by permafrost degradation and altered snowmelt timing suggest that soil water retention and watershed land cover are critical in moderating runoff variation. China’s Law on Soil and Water Conservation and regulations for watershed protection already mandate conservation tillage, reforestation, and slope management as part of erosion control and hydrological regulation.

Our results highlight that these measures will also mitigate flood and drought extremes; existing legal frameworks authorizing watershed rehabilitation programs thus find support in the observed and projected hydro-behavior.

5. Conclusions

Based on historical data from 1980 to 2022, this paper studies the evolution of water resources in the Songhua River Basin from 2026 to 2068. It discusses strategies for the comprehensive management of water resources, with a focus on regional sustainable development in areas prone to water safety risks. The findings are as follows:

• Major shifts in land use occurred between 1990 and 2000, primarily reflected as a marked expansion of cultivated land and a considerable reduction in forest area.
• The integrated XGBoost–Savitzky–Golay hybrid algorithm achieved high-precision simulation of complex hydrological processes, demonstrating strong generalization capability under varying climate conditions.
• The WEP model, based on physical mechanisms, effectively simulated both natural and human-influenced hydrological patterns across basin subsystems, showing consistently robust performance.
• Future projections indicate rising temperatures across all scenarios, with the most pronounced warming under high-emission pathways. Precipitation changes exhibit spatial heterogeneity, alongside increased probability of extreme rainfall events under higher emissions.
• Runoff is projected to increase throughout the basin, predominantly driven by precipitation changes, with seasonal and regional variability. Southern sub-basins show heightened sensitivity to warming, approaching critical thresholds under intense warming scenarios.