Climate Warming-Induced Hydrological Regime Shifts in Cold Northeast Asia: Insights from the Heilongjiang-Amur River Basin

Abstract

Rapid climate warming and intensified human activities are causing profound alterations in terrestrial hydrological systems. Understanding shifts in hydrological regimes and the underlying mechanisms driving these changes is crucial for effective water resource management, watershed planning, and flood disaster mitigation. This study examines the hydrological regimes of the Heilongjiang-Amur River Basin, a transboundary river basin characterized by extensive permafrost distribution in northeastern Asia, by analyzing long-term daily meteorological (temperature, precipitation, evaporation) and hydrological data from the Komsomolsk, Khabarovsk, and Bogorodskoye stations. Missing daily runoff data were reconstructed using three machine learning methods: Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Convolutional Long Short-Term Memory Networks (CNN-LSTM). Trend analysis, abrupt change detection, and regression techniques revealed significant warming and increased actual evapotranspiration in the basin from 1950 to 2022, whereas precipitation and snow water equivalent showed no significant trends. Climate warming is significantly altering hydrological regimes by changing precipitation patterns and accelerating permafrost thaw. At the Komsomolsk station, an increase of 1 mm in annual precipitation resulted in a 0.48 mm rise in annual runoff depth, while a 1 °C rise in temperature led to an increase of 1.65 mm in annual runoff depth. Although annual runoff exhibited no significant long-term trend, low-flow runoff demonstrated substantial increases, primarily driven by temperature and precipitation. These findings provide critical insights into the hydrological responses of permafrost-dominated river basins to climate change, offering a scientific basis for sustainable water resource management and strategies to mitigate climate-induced hydrological risks.

1. Introduction

Global warming is driving significant environmental transformations, particularly in mid- to high-latitude regions, where temperature increases exceed the global average [1]. This rapid warming has led to profound hydrometeorological changes, including shifts in precipitation patterns [2], increased evapotranspiration [3], altered snowmelt dynamics [4], and permafrost thawing [5], with substantial implications for regional water resources and ecosystems [6,7,8]. Additionally, these regions show trends toward more intense and frequent precipitation events, leading to increased flooding [9]. As warming continues, permafrost thawing releases previously trapped water and organic carbon into the hydrological system, profoundly impacting the timing, magnitude, and quality of river runoff [10,11,12].

The Heilongjiang-Amur River Basin, located in mid- to high-latitude regions, is highly sensitive to climate change. The densely populated Chinese side of the basin faces significant water resource allocation issues, which will have widespread economic and ecological impacts [13]. Since the 1990s, ecological problems in the Heilongjiang-Amur River Basin have intensified, leading to extreme hydrological events such as floods and droughts, which may be related to climate change [14]. Between 1891 and 2004, the annual precipitation in the Heilongjiang-Amur River Basin increased by 8%, and the surface air temperature rose by 1.3 °C, significantly higher than the average warming rate in China during the same period [14]. Under the influence of climate warming, snowmelt and permafrost thawing may lead to increased runoff, especially low-flow runoff, thereby affecting the entire water cycle.

Unlike other high-latitude basins, the eastern part of the Heilongjiang-Amur River Basin has a humid temperate monsoon climate, where the monsoons reach its northernmost latitude [15]. The western part, less affected by the monsoon, is more arid. Soil freezing depths exceed 2 m in most parts of the basin, and permafrost is widespread in the most Russian parts of the basin [15]. The flood season of the Heilongjiang-Amur River is June-September, and the river is mainly fed by monsoon rains during the summer and fall [16]. In addition, China engages in more cooperation with neighboring basin countries in the northeast region on transboundary rivers, and the Heilongjiang basin currently serves as a crossroads of China’s Silk Road Economic Belt, Russia’s Far East Development Strategy, and Mongolia’s Steppe Road [17]. Political and economic cooperation in this region is expected to increase.

Stream gauging is effective for monitoring river discharge but is prone to errors from environmental factors and geographic conditions, such as out-of-bank flows, shifting channel geometry, under-ice calibration issues, infrequent site access, and ice breakup events [18,19]. Furthermore, publicly available and up-to-date gauges have declined dramatically [20,21], necessitating runoff series reconstruction and data imputation for accurate hydrological regime analysis. Current models to runoff simulation and data reconstruction can be broadly categorized into process-based models and data-driven models [22]. Process-based models are grounded in representations of hydrological processes and water balance principles, requiring detailed physical modeling, comprehensive understanding of runoff generation mechanisms, and extensive hydrometeorological data inputs [23]. Despite the development of advanced optimization techniques, parameter selection and model configuration often rely heavily on expert judgment, introducing subjectivity and potentially limiting model performance [23]. In recent years, data-driven models—particularly those based on machine learning—have gained considerable attention for their ability to capture complex, nonlinear relationships among hydrological variables without the need for explicit process representations [22]. These models often outperform traditional process-based approaches in terms of predictive accuracy and computational efficiency [24]. However, their lack of embedded physical constraints can introduce uncertainties in the simulation results. In this study, runoff data were reconstructed using three machine learning models. A comparative analysis was conducted to identify the optimal model, thereby reducing uncertainties associated with data reconstruction and improving the accuracy and reliability of the simulated runoff data.

Several studies have examined hydrological changes in the Heilongjiang-Amur River Basin. Zhang et al. [25] applied linear regression, Mann–Kendall trend analysis, and the cumulative anomaly method to assess streamflow and water level trends over the past 30 years. They found a slight decline in discharge and water level at the Luoguhe, Shangmachang, and Kalunshan stations from 1988 to 2017. Although flood events decreased, high-flow durations at Shangmachang and Kalunshan stations were prolonged. Wen et al. [26] developed the distributed hydrological model GBHM-HLJ for the basin. Model projections indicated that future precipitation is likely to decrease, while runoff at the Khabarovsk station could increase by 22.5% and 19.2% under the SSP245 and SSP585 scenarios, respectively. The frequency of extreme runoff events is also expected to rise. Zhou et al. [27] used the SWAT model to investigate the impacts of climate and land use on runoff. Their findings highlighted summer precipitation and temperature as dominant climatic controls. Zhang et al. [28] employed runoff sensitivity coefficients and the SIMHYD model to analyze runoff changes in the Taoer River Basin from 1961 to 2011. They divided the runoff series into three periods—1961–1985, 1986–2000, and 2001–2011—and attributed 41.93% and 43.14% of the runoff increase in the latter two periods to climate change. However, most existing studies about Heilongjiang-Amur River are limited to small or medium-sized basins and relatively short time spans due to the scarcity of observed data. Moreover, current research tends to focus on general runoff trends, with limited attention to low-flow regimes and the mechanisms driving their changes.

By analyzing hydrological and meteorological data of the Heilongjiang-Amur River Basin spanning 1950 to 2022, this study aims to: (1) apply machine learning methods to reconstruct missing daily runoff data of three hydrological stations; (2) examine changes in annual runoff and low flow streamflow; and (3) reveal contrasting responses of streamflow in typical rivers across Northeast Asia to climate warming.

2. Data and Methods

2.1. Study Area

The geographic extent of the Heilongjiang-Amur River Basin spans from 42°0′ N to 55°54′ N and 108°21′ E to 141°21′ E (Figure 1), covering an area of approximately 1.855 million km². This river flows through four countries: China, Russia, Mongolia, and North Korea, with a total length of approximately 4444 km [15]. The annual mean discharge is 10,900 m³/s at its outlet, predominantly sourced from monsoonal rains in summer and autumn. The eastern part of the Heilongjiang-Amur Basin experiences a humid temperate monsoon climate, while the western part is drier and less affected by monsoons. Annual precipitation ranges from 400–500 mm in the upstream and downstream areas to 500–700 mm in the middle reaches. Nearly two-thirds of the annual precipitation occur from June to August, while only 15% falls during the dry season. The basin also exhibits significant temperature variations, with intra-annual differences reaching up to 50 °C [15]. The region experiences severe winter cold, with shallow snow covers and freezing depths exceeding 2 m across much of the basin. Permafrost is a dominant feature in the Russian section of the basin, occurring as continuous, discontinuous, and sporadic permafrost [29].

2.2. Data

The daily meteorological data, including precipitation, air temperature, potential evapotranspiration, and snow water equivalent, were obtained from the ERA5 dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). It spans from 1950 to 2022 and has a spatial resolution of 0.1° × 0.1° (https://cds.climate.copernicus.eu/, accessed on 10 February 2024). We obtained daily actual evapotranspiration data from the Global Land Evaporation Amsterdam Model (GLEAM) website. The potential evapotranspiration data were calculated using the Priestley-Taylor equation and have a spatial resolution of 0.25° × 0.25° (https://www.gleam.eu/, accessed on 24 January 2024). Additionally, precipitation and temperature station data were obtained from the National Oceanic and Atmospheric Administration (NOAA) to validate the accuracy and reliability of the ERA5 dataset (https://www.ncei.noaa.gov/, accessed on 10 February 2024).

The hydrological data used in this study include daily runoff data from three hydrological stations: Komsomolsk (January 1940–January 2007, January 2012–December 2022), Khabarovsk (January 1940–January 2007, January 2012–December 2022), and Bogorodskoye (January 1940–December 1987, January 2008–December 2022). These data were obtained from the Global Runoff Data Centre (GRDC) (https://grdc.bafg.de/GRDC/, accessed on 25 November 2023) and the hydrological yearbooks. Digital elevation model (DEM) data were obtained from the global 30 arc-second elevation dataset (GTOPO30) by the United States Geological Survey (USGS, 1996) (https://www.usgs.gov/, accessed on 24 November 2023).

To examine the relationship between the ERA5 and NOAA data, a scatterplot is drawn, as shown in Figure 2. The correlation coefficient (R) for precipitation data from NOAA and ERA5 is 0.863, while for temperature data it is 0.999, indicating a strong correlation. The linear regression coefficients are 0.97 and 0.96, respectively, suggesting minimal deviation between the two datasets. Since the available stations from 1950 to 1974 are primarily concentrated in Russia and lack representativeness, we only use the basin-averaged data from 1975 to 2022.

2.3. Methods

2.3.1. The Machine-Learning Models for Imputation of the Missing Data

The missing daily runoff data from the KOMSOMOLSK, KHABAROVSK, and BOGORODSKOYE hydrological stations were reconstructed with three machine learning models and their architectures are shown in Figure 3. The input variables for the models include temperature, precipitation, potential evapotranspiration, and snow water equivalent, while the output variable is runoff. The dataset was split into training, validation, and test sets in a 7:1.5:1.5 ratio. Bayesian optimization (BO) was used to optimize the hyperparameters of the three machine learning models during training [30].

The convolutional neural network is a deep learning model that automatically learns features from raw input data, avoiding the process of manual extraction of data features which is usually required for traditional machine learning models [31]. One-dimensional convolutional layers are used to extract features from one-dimensional sequences. Input variables pass from the input layer to the convolutional layer, where convolutional kernels perform local computations on the feature maps from the previous layer [32].

Long Short-Term Neural Network (LSTM) is a variant of the traditional Recurrent Neural Network (RNN) and a special RNN structure with a deeper hierarchy along the temporal dimension [33]. A distinctive feature of the Long Short-Term Memory (LSTM) model is its gating mechanism, which includes three types of gates: input gate, output gate, and forget gate [31]. With forget gates and memory cells, LSTMs can determine the duration over which state information is forgotten and retained, effectively addressing the issue of vanishing and exploding gradients [34].

The CNN-LSTM model, which integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks, is effective for runoff prediction [35]. In this model, the CNN consists of alternating convolutional and pooling layers. Each layer uses the ReLU activation function to accelerate convergence and employs convolutional kernels to extract latent features from the samples. The LSTM then processes the features provided by the CNN, capturing its long-term features to perform temporal predictions.

In this study, model accuracy was evaluated with Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Nash–Sutcliffe Efficiency (NSE), and the Pearson correlation coefficient (R), which were calculated as follows:

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

(1)

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

(2)

$$NSE=1-{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^N{\left(y_i-{\widehat{y}}_i\right)}^2}}{{\displaystyle {\sum }_{i=1}^N{\left(y_i-\overline{y}\right)}^2}}}$$

(3)

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

(4)

where $N$ represents the number of data points, $y_i$ and ${\widehat{y}}_i$ represent the observed and simulated runoff, and $\overline{y}$ and ${\widehat{y}}_i$ represent the mean of the observed and simulated runoff.

2.3.2. Time-Series Analysis and Runoff Indicator

The Mann–Kendall nonparametric test is a statistical method proposed by Mann [36] and Kendall [37] for detecting trends in time-series data. It has a great advantage in identifying both linear and non-linear trends, as well as detecting abrupt change points [38]. In the Mann–Kendall method, there are two statistical indicators: the statistic sequence curve ($U{F}_k$) and the statistic sequence reverse curve ($U{B}_k$). When $U{F}_k$ is greater than 0, it indicates an upward trend in the time series; conversely, when $U{F}_k$ is less than 0, it indicates a downward trend. If the $U{F}_k$ and $U{B}_k$ have intersections within the significant level range, the intersection is the change point. Additionally, this study also employs linear regression and the change rate of climate element tendency to analyze the long-term trends of runoff and meteorological factors. The change rate of climate element tendency is generally expressed quantitatively using a single linear equation:

$$Y=a_0+a_{1}t$$

(5)

where $a_0$ is constant term, $a_1$ is linear trend term, Y is a meteorological factor, t represents time, and $a_1\times 10$ represents the 10-year change rate of climate element tendency.

In this study, the coefficient of variation ($C_v$) [8] and concentration ratio ($C_n$) are used to assess the intra-annual and inter-annual variability of runoff. $C_v$ can be used to measure the magnitude of inter-annual runoff variability [39]. $C_n$ is used to assess the intra-annual distribution of runoff, with higher $C_n$ values indicating a more concentrated distribution within the year.

Calculating $C_n$ requires considering the monthly runoff for each year as a vector, with the runoff volume as the vector length and the month as the vector direction. The azimuths from January to December are 0°, 30°, 60°, …, and 330°, respectively. Then, we can calculate the vector decomposition of monthly runoff in the x and y directions [40]:

$$R_x={\displaystyle {\sum }_{i=1}^{12}{R}_i}\cos {\theta }_i$$

(6)

$$R_y={\displaystyle {\sum }_{i=1}^{12}{R}_i}\sin {\theta }_i$$

(7)

The calculation formula for $C_n$ is as follows:

$$C_n={\displaystyle \frac{\sqrt{R_x^2+R_y^2}}{{\displaystyle {\sum }_{i=1}^{12}{R}_i}}}$$

(8)

where $R_i$ represents the monthly runoff.

2.3.3. Principal Component Regression (PCR)

Principal Component Regression (PCR) is a method of multiple linear regression which employs the principal components of the independent variables as regression variables. This is particularly useful for datasets exhibiting multicollinearity or where the number of variables greatly exceeds the number of observations [41]. The principal component regression equation can be expressed as:

$$\mathrm{y}={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_k{x}_k+\epsilon$$

(9)

where y is the dependent variable, $x_1$, $x_2$, …, $x_k$ are the independent variables, ${\beta }_0$, ${\beta }_1$, ${\beta }_2$, …, ${\beta }_k$ are the model parameters, ${\beta }_0$ is the intercept, ${\beta }_0$ (i = 1, 2, …, k) are the coefficients corresponding to each variable, and ε is the error variable.

3. Results

3.1. Evaluation of the Performance of Three Deep Learning Models in Daily Runoff Reconstruction

To evaluate the applicability of three models, the runoff is simulated using the test dataset, and the RMSE, MAE, NSE and R are presented in

Table 1. Simulation accuracy evaluation of machine learning models.

Model	Hydrological Station	RMSE (mm)	MAE (mm)	NSE	R
	Komsomolsk	0.16	0.12	0.84	0.94
CNN	Khabarovsk	0.16	0.12	0.80	0.93
	Bogorodskoye	0.18	0.13	0.83	0.94
	Komsomolsk	0.16	0.11	0.79	0.94
LSTM	Khabarovsk	0.15	0.11	0.80	0.94
	Bogorodskoye	0.19	0.13	0.78	0.93
	Komsomolsk	0.17	0.12	0.82	0.93
CNN-LSTM	Khabarovsk	0.18	0.13	0.76	0.91
	Bogorodskoye	0.18	0.13	0.84	0.93

. The average RMSEs of CNN, LSTM, and CNN-LSTM for the whole interval are 0.17 mm, and the average MAE is 0.12, 0.12, and 0.13, respectively, indicating that the errors of these three models are comparable. The average NSE is 0.82, 0.79, and 0.81, respectively, showing that NSE of CNN is slightly higher than the NSE of LSTM and CNN-LSTM. Evidently, these models are capable of acquiring precise data on changes in runoff.

Figure 4(a1,b1,c1) shows the time-series changes of the measured and predicted values. The time-series change of the daily runoff predicted by the neural network models follows the trend of the observed values. Comparative analysis shows that all three models tend to underestimate flood peaks during the flood season. Compared to LSTM and CNN-LSTM, the CNN model captures the position of the peaks more accurately. During the testing period, the CNN and CNN-LSTM models generally underestimate the low runoff values, whereas the LSTM model provides more accurate low runoff value simulations. In terms of the overall trend of runoff simulation, the LSTM model’s simulated values closely align with the actual daily runoff values. At the Bogorodskoye station, which has the least recorded observed runoff data and the smallest dataset, all three models perform poorly in simulating low runoff values, failing to capture the finer details of the variation process.

Figure 4(a2,b2,c2) illustrates the linearity between the observed values and the daily runoff simulated by CNN, LSTM, and CNN-LSTM. The values predicted by CNN, LSTM, and CNN-LSTM are shown to have linearity with the observed values. The lower the runoff, the stronger the linearity, and the model tends to underestimate when the runoff is relatively high. The linearity of the CNN and LSTM is similar, but CNN-LSTM shows weaker linearity compared with CNN and LSTM in the Khabarovsk basin.

To evaluate the impact of forecast periods on model accuracy, we calculate the RMSE and NSE of the three models for different forecast periods at the Komsomolsk station. As shown in Figure 5, the forecast periods with the highest Nash efficiency coefficients for CNN, LSTM, and CNN-LSTM are 20 days, 25 days, and 10 days, respectively. The forecast periods with the lowest root mean square errors for CNN, LSTM, and CNN-LSTM are 25 days, 25 days, and 20 days, respectively. Across different forecast periods, the LSTM model exhibits relatively stable accuracy and error changes, whereas the CNN model shows relatively larger variations in accuracy and error.

After a comprehensive analysis of model accuracy evaluation indexes and hydrological process simulation results, CNN, LSTM, and CNN-LSTM were selected as models for reconstructing runoff data at Komsomolsk, Khabarovsk, and Bogorodskoye stations, respectively. For convenience and consistency, the forecast period for the models was uniformly set to 20 days.

3.2. Climate Change Evidenced by Rising Air Temperatures and Increasing Precipitation and Evapotranspiration

Figure 6 and

Table 2. Seasonal trend analyses by Mann–Kendall test. P is precipitation. T is temperature. ET is actual evapotranspiration. SWE is snow water equivalent.

	P			T			ET			SWE
	Trend	Z	Slope (mm/a)	Trend	Z	Slope (°C/a)	Trend	Z	Slope (mm/a)	Trend	Z	Slope (mm/a)
Spring	no	0.615	0.003	upward	5.291	0.052	upward	3.834	0.035	no	−0.824	−0.035
Summer	no	1.390	0.123	upward	5.929	0.022	no	1.824	0.031	no	−1.548	−0.014
Autumn	no	−0.655	−0.021	upward	3.700	0.016	no	0.929	0.010	no	−1.538	−0.005
Winter	upward	2.462	0.037	upward	4.786	0.046	upward	3.672	0.002	no	0.052	0.007

show the change rate of climate element tendency and Mann–Kendall trend analysis results. The analysis of interdecadal variation shows that precipitation followed a slow increasing trend before 2000 and that the growth rate has increased to 9.08 mm per decade since the 21st century. The maximum value occurred in 2020, significantly higher than the average level. Among the four seasons, only winter precipitation exhibits an evident increasing trend, with a rate of 0.037 mm/year. The multi-year average precipitation in July is the highest, reaching 121 mm. The annual average temperature shows an upward trend, with a change rate of 0.341 °C/decade. The Mann–Kendall test results reveal an increasing trend in temperature throughout all seasons. However, an exceptionally low value of −3.33 °C was observed in 1969, with the average winter temperature reaching −23 °C. Similar to temperature, evapotranspiration also shows a significant increasing trend, with a change rate of 2.34 mm/decade, probably driven by rising temperatures in spring and winter. The snow water equivalent has no obvious variation tendency. The change process can be divided into two phases: an upward trend from 1950 to 1980, followed by a downward trend from 1980 to 2022, with larger fluctuations after 2008.

The results of trend tests of 73 years of meteorological elements are presented in Figure 6(b1, b2, b3, b4). These results show that precipitation mutations occurred in 1963 and 2019. The UF line and UB line of the temperature intersected in 1987, although the intersection point in 1987 is slightly above the critical value. This is close to the assessment of the change points for annual mean temperature in 30 among 44 weather station [42]. The multi-year average temperature from 1987 to 2022 is also notably higher than that from 1950 to 1987. The abrupt changes in actual evapotranspiration and snow water equivalent occurred in 1964 and in 1952 and 2016, respectively.

3.3. Annual Runoff and Low-Flow Runoff Exhibited Interannual Variability

Considering the similarity of the runoff at three stations, only the results at Komsomolsk station are shown in the following sections (Figure 7 and Figure 8). The results at Khabarovsk and Bogorodskoye stations are presented in Figure A1 (Appendix A). According to the linear analysis, there is no trend in the annual runoff but it exhibits year-to-year variability, following a cyclical pattern of change. It can be seen that the trends in annual runoff and precipitation generally align, suggesting that runoff is primarily influenced by precipitation. In contrast, the low-flow runoff shows a significant increasing trend, with a change rate of 0.032 mm/year. The abrupt change point occurs in 1980, after which the multi-year average low-flow runoff is higher than before. This may be related to the increase in temperature and the decrease in snow depth. The highest average annual low-flow runoff of 0.869 mm was observed in 1951, while the lowest average annual low-flow runoff of 5.052 mm occurred in 2013.

Figure 9 shows the trends of seasonal flow at three stations from 1950 to 2022. It can be seen that there is a decreasing trend in the basin with higher flows in summer (June–August) and autumn (September–October), and an increasing trend with lower flows in spring (March–May) and winter (December–February).

Figure 10a depicts the intra-annual variation of monthly runoff depth and coefficient of variation in the Heilongjiang-Amur River Basin. The results indicate that the runoff is primarily concentrated between May and October. The intra-annual variation trends of runoff differ among the three hydrological stations. The Komsomolsk and Bogorodskoye stations show two peak values in the multi-year average monthly runoff, occurring in June and September, while the Khabarovsk station shows a single peak in August. At the three stations, the coefficient of variation of simulated annual runoff over the 1950–2022 baseline period ranges from 0.177 to 0.516. The coefficient of variation for interannual runoff is higher in spring and winter, indicating strong fluctuations in total runoff during these seasons. Figure 10b indicates the interannual variation trend of concentration ratio. It can be seen that from 1950 to 2022 the annual runoff concentration index mainly ranged between 0.37 and 0.6, with the lowest value (0.33) occurring in 2008 at the Bogorodskoye station. During this period, $C_n$ exhibits a significant decreasing trend, suggesting that the intra-annual distribution of runoff is becoming increasingly uniform.

3.4. Runoff Response to Climate Change

For principal component regression, we used precipitation, temperature, potential evapotranspiration, and snow water equivalent as independent variables, with annual runoff and low-flow runoff as dependent variables. The results of the collinearity diagnostics are shown in

Table 3. Covariance diagnostic condition index.

Index (k)	1	2	3	4	5
Komsomolsk	1.000	2.230	11.592	20.018	82.948
Khabarovsk	1.000	2.511	11.550	19.393	80.445
Bogorodskoye	1.000	2.288	11.644	20.468	84.052

. Typically, a condition index of 10 ≤ k ≤ 100 indicates strong collinearity among the factors, while an index of k ≥ 100 indicates severe collinearity. The indexes 3, 4, and 5 for the three hydrological stations are all greater than 10 but less than 100, suggesting that collinearity between variables should be eliminated before running multiple regression to ensure more accurate results.

With an average coefficient of determination of 0.829, the principal component regression model effectively represents the quantitative relationship between runoff and meteorological factors (

Table 4. Summary of principal component regression model parameters.

Hydrological Station	${\mathit{\beta }}_0$	${\mathit{\beta }}_1$	${\mathit{\beta }}_2$	${\mathit{\beta }}_3$	${\mathit{\beta }}_4$	$\mathit{R}$
Komsomolsk	$41.143$	0.484	1.647	−0.058	1.437	0.838
Khabarovsk	18.107	0.457	3.361	−0.042	1.768	0.839
Bogorodskoye	55.848	0.415	3.156	−0.037	0.561	0.811

). In the Komsomolsk basin, an increase of 1 mm in annual precipitation results in a 0.484 mm increase in annual runoff. An increase of 1 °C in temperature corresponds to a 1.647 mm increase in annual runoff. Runoff will increase 0.058 mm per 1 mm decrease in potential evapotranspiration, while it will increase 1.437 mm per 1 mm increase in snow water equivalent.

To assess the importance of meteorological factors on runoff, we performed principal component regression with standardized variables.

Table 5. Summary of parameters of the standardized principal component regression model (annual runoff).

Hydrological Station	${\mathit{\beta }}_1$	${\mathit{\beta }}_2$	${\mathit{\beta }}_3$	${\mathit{\beta }}_4$	$\mathit{R}$
Komsomolsk	0.736	0.04	−0.244	0.122	0.838
Khabarovsk	0.776	−0.045	−0.337	0.139	0.809
Bogorodskoye	0.710	0.088	−0.178	0.058	0.769

shows that annual runoff is positively correlated with precipitation and snow water equivalent, and negatively correlated with potential evapotranspiration. Precipitation remains the dominant factor influencing runoff variability, followed by potential evapotranspiration, while snow water equivalent and temperature have relatively minor effects on annual runoff. As shown in Figure 1, the Khabarovsk, Komsomolsk, and Bogorodskoye stations are situated at latitudes of approximately 135° N, 137° N, and 140° N, respectively, indicating a northward expansion of the watershed areas.

Principal component regression is also performed on the standardized low-flow runoff to assess the relative importance of climate factors. The low-flow runoff at the Komsomolsk station can be expressed as:

$$y=0.348x_1+0.420x_2+0.228x_3+0.225x_4$$

(10)

where y is normalized low-flow runoff, and $x_1$, $x_2$, $x_3$, $x_4$ are normalized precipitation, air temperature, potential evapotranspiration, and snow water equivalent, respectively.

In the regression equation, the regression coefficients for precipitation and temperature are relatively large, at 0.348 and 0.42, respectively. Therefore, temperature dominated low-flow runoff variations in the Heilongjiang-Amur River Basin. This is likely due to the extensive distribution of permafrost in the basin, most of which is seasonal [43]. Rising temperatures accelerate permafrost thawing, increasing groundwater storage and aquifer permeability [44]. This enhances the connectivity and exchange intensity between surface water and groundwater [45]. This influx of thawed water into river channels increases baseflow [46], affecting runoff dynamics. Studies indicate that the phenomenon of underground supply supplementing surface runoff due to permafrost degradation is particularly pronounced in spring and winter [47], corroborating the significant rise in low-flow runoff observed in the Heilongjiang-Amur River Basin. Furthermore, the increasing temperatures during spring and winter seasons in the basin also facilitate snowmelt, further impacting low-flow runoff.

To further investigate the impact of temperature and precipitation on annual runoff and low-flow runoff in the Heilongjiang-Amur River Basin, this study adopted the methods of Risbey and Entekhabi [48] and Wang [49]. The deviation percentages for annual runoff ($\Delta Q={\displaystyle \frac{Q-\overline{Q}}{\overline{Q}}}\times 100\%$), low-flow runoff ($\Delta Q_s={\displaystyle \frac{Q_s-{\overline{Q}}_s}{{\overline{Q}}_s}}\times 100\%$), precipitation ($\Delta P={\displaystyle \frac{P-\overline{P}}{\overline{P}}}\times 100\%$), and temperature ($\Delta T={\displaystyle \frac{T-\overline{T}}{\overline{T}}}\times 100\%$) were calculated and represented as contour maps (Figure 8). In order to better compare the relationships between annual runoff and low-flow runoff with temperature and precipitation, the deviation percentages for temperature and precipitation were divided by their respective standard deviations.

As shown in Figure 11, annual runoff is highly sensitive to changes in precipitation and exhibits very low sensitivity to changes in temperature. In contrast, low-flow runoff is highly sensitive to both temperature and precipitation, with a greater temperature sensitivity. Over the 70 years, the variability in low-flow runoff is greater than that of annual runoff, with changes reaching up to 112%. The contour plots at Khabarovsk and Bogorodskoye station are presented in Figure A2.

4. Discussion

The hydrological regime shifts observed in the Heilongjiang-Amur River Basin over recent decades reflect a broader global pattern of climate-induced hydrological changes. From 1950 to 2022, the basin has undergone substantial climate-driven hydrological shifts, likely caused by a 0.3 °C per decade rise in mean annual temperature and a 6.8 mm per decade increase in evapotranspiration. Although annual precipitation has exhibited high interannual variability, a pronounced upward trend was identified between 2000 and 2020, peaking at 730 mm in 2020. Concurrently, the SWE showed a slight decline with significant fluctuations. These changes align with global trends, particularly in cold regions [3,50,51,52,53,54], where climate warming increasingly altering hydrological regimes.

At a global scale, the hydrological shifts in the Heilongjiang-Amur River Basin resemble those observed in other northern cold river basins. The upward trend in low-flow runoff, primarily driven by enhanced groundwater contributions from permafrost thaw, is consistent with findings from other major Eurasian rivers [55,56]. However, the relatively stable or declining annual runoff in this basin contrasts with the rising trends typically observed in Siberian rivers, such as the Lena and Yenisei [49]. This divergence highlights the unique climatic characteristics of the basin, where the humid, monsoonal environment renders annual runoff more sensitive to variations in evapotranspiration than to permafrost degradation. The interplay of these factors underscores the importance of regional climatic characteristics in modulating hydrological responses to global warming [57].

As shown in Figure 10, the intra-annual variability of runoff in the Heilongjiang-Amur River Basin exhibits a clear decreasing trend. This pattern is consistent with the seasonal runoff dynamics illustrated in Figure 9, characterized by increasing runoff in spring and winter. Since no significant trend is observed in precipitation seasonality during this period, the reduction in runoff variability is primarily attributed to rising air temperatures. In spring, elevated temperatures lead to earlier snowmelt, advancing the onset of runoff prior to the flood season and narrowing the range of intra-annual runoff [58]. When temperatures approach or exceed the freezing point, a larger proportion of precipitation occurs as rain [59]. Simultaneously, warmer conditions reduce soil frost depth, enhancing infiltration and increasing soil moisture [59]. These combined effects result in greater spring runoff, which weakens the intra-annual runoff variability [60]. Additionally, warming-induced permafrost degradation plays a crucial role in increasing winter runoff.

Permafrost thaw is critical in explaining the observed increase in low-flow runoff [11,61], which corresponds with a global trend of increased groundwater discharge in permafrost-affected regions [62,63]. In winter, when the upper soil and active layer are frozen, streamflow is sustained mainly by deep groundwater from below the permafrost [64]. This represents the annual minimum flow, reflecting subsurface hydrological connectivity [65]. As ground temperatures rise and ice-rich permafrost thaws, soil permeability improves and previously restricted vertical and lateral flow paths are reactivated [66]. This leads to the release of stored water into the subsurface system, thus increasing winter baseflow—a mechanism well documented in high-latitude basins [49]. For example, Ala-aho et al. [67] conducted spatial and temporal sampling of stable water isotopes across soils, lakes, and rivers in Western Siberia, revealing that snowmelt mixes with substantial volumes of water stored in organic soils and lakes, subsequently contributing to runoff generation during the thaw season. Their findings suggest that hydrological connectivity among landforms and water bodies persists throughout the permafrost region and becomes increasingly pronounced under warming conditions. Park et al. [68] utilized satellite microwave observations of freeze–thaw dynamics to monitor permafrost extent and condition, demonstrating that the observed deepening of the active layer is consistent with widespread permafrost degradation in response to recent climate change. Shan et al. [69] applied an enhanced ground freezing index model to simulate permafrost extent and condition between 2000 and 2020, and reported a consistent deepening trend indicative of ongoing permafrost degradation driven by recent warming. Permafrost influences regional hydrological processes by altering evapotranspiration patterns, reducing soil infiltration rates and groundwater recharge, and ultimately leading to intra-basin runoff changes. Moreover, permafrost thaw may facilitate additional subsurface hydrological processes, such as perennial subzero water flow, soil water recharge, and enhanced surface-groundwater interactions [66]. While direct observational evidence remains limited, many studies have provided related support for these mechanisms [66]. The degradation of permafrost can also trigger geomorphic changes, such as thermokarsting and plateau subsidence, further altering watershed-scale hydrological processes [70]. In the Heilongjiang-Amur River Basin, the transition zone between sporadic and isolated permafrost lies near key stations such as Komsomolsk and Bogorodskoye [71], indicating the region’s sensitivity to subsurface thermal conditions. Despite differences in climate conditions among Arctic basins, these underlying processes are fundamental to understanding the ongoing increase in low-flow runoff and the overall attenuation of runoff seasonality in this region [72].

The distinctive hydrological response of the Heilongjiang-Amur River Basin provides valuable insights into the complexities of climate change impacts in transitional regions. Here, the accelerating thaw of permafrost and shifts in monsoonal precipitation patterns interact to create a complex balance between atmospheric and subsurface dynamics. As global warming intensifies the hydrological cycle, regions like the Heilongjiang-Amur Basin, which are sensitive to both warming and seasonal precipitation shifts, are likely to experience further hydrological variability. This increasing unpredictability underscores the necessity for adaptable and forward-looking water resource management strategies, which are crucial for mitigating the risks associated with these evolving hydrological patterns.

5. Conclusions

This study investigates the hydrological regimes of the Heilongjiang-Amur River Basin, a region highly sensitive to climate change. We demonstrate that CNN, LSTM, and CNN-LSTM models are effective for reconstructing missing daily discharge data. These models achieved average NSE values of approximately 0.8 and correlation coefficients exceeding 0.9 across three hydrological stations. Among them, the CNN model exhibited the highest performance, followed by the CNN-LSTM and LSTM models.

While no statistically significant trends were detected in precipitation, snow water equivalent and annual runoff, a pronounced upward trend was observed in air temperature and actual evapotranspiration. All three hydrological stations recorded significant increases in low-flow runoff, without evidence of abrupt regime shifts.

Our analysis identifies precipitation as the principal driver of interannual runoff variability, with temperature-driven changes in potential evapotranspiration and SWE playing secondary roles. At the Komsomolsk station, regression analysis indicates that a 1 mm increase in annual precipitation results in a 0.48 mm increase in annual runoff depth, while a 1 °C increase in temperature results in a 1.65 mm increase in annual runoff depth. Furthermore, variability in low-flow runoff at Komsomolsk is primarily governed by temperature, suggesting a strong link to ongoing permafrost degradation.

Understanding the changes in hydrological conditions in the Heilongjiang-Amur River Basin and their response to climate warming enhances our understanding of evolving hydrological processes in the region. It also provides a scientific basis for managing disasters such as floods and droughts under changing environmental conditions. However, this study utilized data from only three outlet hydrological stations, which may not fully capture spatial heterogeneity in hydrological responses across the entire basin. Future research could build upon this work by examining regional heterogeneity in greater detail.