Multi-Model Analyses of Spatiotemporal Variations of Water Resources in Central Asia

Abstract

Over the past 70 years, Central Asia has emerged as a globally recognized water security hotspot due to its unique geographic location and uneven distribution of water resources. In arid and semi-arid regions, understanding runoff dynamics under climate change is essential for ensuring regional water security. This study addresses the data-sparse Central Asian region by applying the ISIMIP3b multi-scenario analysis framework, selecting three representative global hydrological models. Using model intercomparison, trend analysis, and geographically weighted regression, we assess the spatiotemporal evolution of runoff from 1950 to 2080 and investigate the spatial heterogeneity of runoff responses to precipitation and temperature. The results show that under the historical scenario, all models consistently identify similar spatial pattern of runoff, with higher values in southeastern mountainous regions and lower values in western and central regions. However, substantial differences exist in runoff magnitude, with regional annual means of 10, 26, and 68 mm across the three models, respectively. The spatial disparity of runoff distribution is projected to increase under higher SSP scenarios. During the historical period, most of Central Asia experienced a slight decreasing trend in runoff, but the overall trends were −0.022, 0.1, and 0.065 mm/year, respectively. In contrast, future projections indicate a transition to increasing trends, particularly in eastern regions, where trend magnitudes and statistical significance are notably greater than in the west. Meanwhile, the spatial extent of significant trends expands under high-emission scenarios. Precipitation exerts a positive influence on runoff in over 80% of the region, while temperature impacts exhibit strong spatial variability. In the WaterGAP2-2e and MIROC-INTEG-LAND models, temperature has a positive effect on runoff in glaciated plateau regions, likely due to enhanced snow and glacier melt under warming conditions. This study presents a multi-model framework for characterizing climate–runoff interactions in data-scarce and environmentally sensitive regions, offering insights for water resource management in Central Asia.

1. Introduction

Runoff, a fundamental component of the hydrological cycle, serves as a critical freshwater source, especially for arid regions [1], directly influencing agricultural irrigation, land use, ecosystem stability, and water security [2]. In water-limited areas, runoff variability can exacerbate ecological and hydrological stresses [3], particularly in Central Asia. The five Central Asia countries, located in the core of Eurasia, constitute the largest arid and semi-arid area in the Northern Hemisphere, making their water security globally significant [4]. However, Central Asia faces escalating water scarcity due to climatic and anthropogenic pressures. The region has warmed at 0.4 °C per decade, exceeding global averages, and lost 30% of its glacial mass over the past 50 years [5,6]. Inefficient irrigation practices [7] and fragmented governance [8] further strain water resources, while projected climate change, urbanization, and growing food demand [9,10] threaten to intensify these challenges. Hence, a comprehensive assessment of current and future dynamics of water resources is thus imperative to inform adaptive strategies and ensure regional water sustainability.

Central Asia’s water resources exhibit extreme spatial heterogeneity, driven by its complex topography. The Tian Shan and Pamir Mountains, known as the region’s “water towers” [11,12], generate 80% of total runoff but consume only 16% [13]. Upstream countries (Tajikistan and Kyrgyzstan) struggle to utilize glaciers and groundwater reserves [14], while downstream nations (Kazakhstan, Uzbekistan, and Turkmenistan) account for 83% of water use despite receiving merely 14% of runoff [13]. In this context, Kazakhstan’s annual surface runoff declined by 39.4 km³ over the period 1965–2016 [15], and Turkmenistan holds just 0.5% of regional surface water [16]. Although transboundary rivers like the Amu Darya and Syr Darya have been studied extensively [17,18], systematic comparisons of runoff–climate responses across all five countries remain scarce, particularly regarding regional differences in climate–runoff responses under extreme spatial heterogeneity.

Historically, hydro-meteorological station data and statistical records have supported Central Asia water research, but their temporal gaps and cross-border inconsistencies limit large-scale analyses [19]. The development of large-scale hydrological models has introduced a new paradigm for water resources assessment through physically based simulations. Early studies, such as WaterGAP3 applications under the SRES scenarios, revealed divergent projections of water availability due to climate forcing uncertainties [20], while SWAT models in the Vakhsh Basin linked rising runoff to glacier melt [21]. However, these works were constrained by limited GCM ensembles and outdated scenarios, necessitating reassessment with IPCC AR6 multi-model frameworks. Recent advances in high-resolution datasets, including ISIMIP [22], GLADS [23], and CMIP [24], which have enabled robust hydrological evaluations. Notably, the ISIMIP project has performed bias correction on GCM outputs, enabling numerous climate change impact assessments and future projections. Davie et al. evaluated ISIMIP datasets using runoff trend indicators, confirming their effectiveness in reproducing observed runoff patterns across most global regions [25].

This study provides a systematic assessment of climate-driven runoff changes across Central Asia by analyzing ISIMIP3b multi-model ensemble data (1950–2080) under SSP-RCP scenarios. We address two critical questions: (1) How will spatial heterogeneity in runoff trends evolve under different climate projections? (2) Which subregions exhibit the highest sensitivity to precipitation and temperature changes? By employing geographically weighted regression to quantify localized climate-runoff relationships, our results offer spatially explicit insights for adaptive water management and support the development of differentiated adaptation strategies, ultimately contributing to water security cooperation along the belt and road.

2. Materials and Methods

2.1. Study Region

This study focuses on the core Central Asia region, encompassing five countries: Kazakhstan, Tajikistan, Kyrgyzstan, Uzbekistan, and Turkmenistan (Figure 1). Encompassing an area of approximately 4 million km², the region lies at the heart of the Eurasian continent. It is bordered by the Caspian Sea to the west, China’s Xinjiang Province to the east, Russia to the north, and Iran and Afghanistan to the south, serving as a critical geopolitical and ecological transition zone. The topography exhibits complex and pronounced elevation gradients. Tajikistan and Kyrgyzstan, located in the mountainous southeast, are rich in glacial and snowmelt runoff, acting as upstream water suppliers [26]. These regions are particularly vulnerable to glacier retreat and face challenges related to sustainable hydropower development. In contrast, Uzbekistan and Turkmenistan are largely downstream countries situated in lowland arid zones, heavily dependent on transboundary river inflows. These countries often grapple with irrigation inefficiencies, salinization, and heightened risks of interannual water scarcity. Kazakhstan spans diverse ecological zones from arid deserts in the south to more humid steppe regions in the north, with varying runoff conditions [27]. Several transboundary rivers traverse Central Asia, among which the Amu Darya and Syr Darya serving as the main sources for agricultural irrigation. Irrigation constitutes the dominant water use in the region, despite its low efficiency [4]. The continental interior location and enclosed topography, particularly the southeastern mountain barriers that block moisture transport from the Indian and Pacific Oceans, create a typical temperate continental arid climate. Mean annual temperatures range from 4 to 8 °C, with strong seasonal and diurnal variability. The region experiences scarce precipitation (annual average < 300 mm) but intense evaporation (900–1500 mm/year) [28,29]. These arid and enclosed hydroclimatic conditions make the regional runoff highly sensitive to climate change.

2.2. Data and Processing

This study utilizes data from the Inter-Sectoral Impact Model Intercomparison Project Phase3b (ISIMIP3b; https://www.isimip.org/), which provides a framework for assessing the impacts of climate change across multiple sectors. The ISIMIP3b protocol is specifically designed to quantify climate-related risks under varying radiative forcing and socioeconomic conditions. Its multi-model ensemble approach facilitates a comprehensive assessment of uncertainty in climate impact projections [30]. It is worth noting that ISIMIP 3b uses bias-corrected climate data from GCM as input.

Simulation data were obtained for two time periods, including the historical baseline (1950–2014) and the future projection period (2015–2080), driven by different climate and socioeconomic scenarios (

Table 1. Explanation of ISIMIP 3b data used in this study.

Classification	Secondary Classification	Description
Model	H08	Two representative global hydrological models and one representative land surface model.
	WaterGAP2-2e
	MIROC-INTEG-LAND
GCMs	GFDL-ESM4	Using multi-model ensemble of five CMIP6 climate models in ISIMIP3b.
	IPSL-CM6A-LR
	MPI-ESM1-2-HR
	MRI-ESM2-0
	UKESM1-0-LL
Scenarios	historical + histsoc	histsoc: Time-varying, historical socio-economic scenarios. 2015 soc: Fixed year-2015 direct human influences.
	SSP126 + 2015 soc
	SSP370 + 2015 soc
	SSP585 +2015 soc

). Three GHMs (global hydrological models)/LSMs (land surface models) were selected: H08 [31,32], WaterGAP2-2e [33], and MIROC-INTEG-LAND [34] (

Table 2. Details of GHMs/LSMs used in this study.

Model	Evapotranspiration Scheme	Snow Scheme	Groundwater Scheme	Runoff Scheme	River Routing Scheme	Human Water Use	CO2 Effect	Reference
H08	Bulk formulation	Energy balance method	Explicit (renewable and nonrenewable reservoirs)	Saturation excess, baseflow	Linear reservoir model	Irrigation	No	[31,32]
WaterGAP 2-2e	Priestley–Taylor with varying alpha values for arid and humid areas	Degree-day method	Explicit (single reservoir)	Saturation excess, beta function	Linear reservoir cascade	Irrigation, domestic electricity, manufacturing, livestock	No	[33]
MIROC-INTEG-LAND	Penman–Monteith formulation	Energy balance method	Implicit (assumed to be handled within the soil moisture)	Saturation excess, baseflow	Kinematic wave model	Irrigation, domestic, industry	Yes	[34]

). Each model was forced by climate data from five global circulation models (GCMs), including GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL, resulting in 15 model combinations (5 GCMs × 3 models). To minimize individual model biases, ensemble means were calculated for the five GCMs under each model configuration [35]. Despite the robustness of using ensemble means, some uncertainties in the simulations remain due to the model structures, spatial resolution, and so on. The historical simulations (1950–2014) were based on historical radiative forcing (HIST) and “histsoc” socioeconomic scenarios, while future simulations (2015–2080) incorporated three shared socioeconomic pathways (SSP126, SSP370, SSP585, representing low, moderate, and high forcing scenarios respectively) under the “2015soc” socioeconomic framework. Temperature and precipitation inputs were also obtained from ISIMIP3b, using the five GCM ensemble means. Historical runoff and meteorological data were specifically employed for analyzing regional variations in climate change impacts on Central Asia runoff. All datasets share a uniform spatial resolution of 0.5° × 0.5°.

The complete dataset includes 12 runoff datasets providing monthly runoff (kg m⁻² s⁻¹) and 4 meteorological datasets containing daily temperature (K) and precipitation (kg m⁻² s⁻¹). Prior to analysis, all variables were converted to common hydrological units (runoff and precipitation to mm; temperature to °C). The data were then temporally aggregated to the annual scale and spatially cropped to the Central Asia domain, resulting in annual runoff and meteorological datasets for subsequent analysis.

2.3. Methodology

2.3.1. Relative Change Analysis

Due to the strong skewness and presence of extreme values in Central Asian runoff data, a relative change analysis was adopted to assess the variations between future runoff projections and reference scenarios under each model. The relative percentage change (RPC) [36] was calculated as follows:

$$\begin{array}{c}RPC={\displaystyle \frac{Q^{\prime }-Q}{Q}}\times 100\%\end{array}$$

(1)

where Q′ represents the future runoff projection, and Q is the historical runoff simulation. A larger absolute RPC value indicates a more substantial change in future runoff relative to the historical baseline.

2.3.2. Trend Analysis

Sen’s slope estimator The Sen’s slope estimator [37] is a robust nonparametric statistical method for trend analysis. It is computationally efficient and does not require assumptions of data normality or serial autocorrelation, making it particularly suitable for datasets with outliers or minor gaps [38]. The World Meteorological Organization (WMO) recommends this method for trend detection in hydro-meteorological data. The method calculates the slopes between all possible pairs of data points within the time series, and the median of these slopes is taken as the estimate of the overall trend. The Sen’s slope (β) is computed as follows:

$$\begin{array}{c}\beta =Median\left({\displaystyle \frac{X_j-X_i}{j-i}}\right),\forall j>i\end{array}$$

(2)

where X_j and X_i represent data points at times j and i in the runoff time series, respectively. The median slope β serves as the trend indicator: a positive β signifies an upward trend, while a negative β indicates a downward trend in the time series.

Mann–Kendall test The significance of trend variations was evaluated using the Mann–Kendall statistical test [39,40]. As a widely used nonparametric method, the MK test is well-suited for hydrological and climatic data, as it does not assume normality and is capable of distinguishing statistically significant trends from natural variability. The test determines trends based on the sign relationships between data pairs in a time series. For a time series of length n, the test statistic S, variance δ², and standardized statistic Z are sequentially calculated as follows:

$$\begin{array}{c}S={\displaystyle \sum _{k=1}^{n-1}}{\displaystyle \sum _{j=k+1}^n}Sgn\left(X_j-X_k\right)\end{array}$$

(3)

$$\begin{array}{c}{\delta }^2=Var\left(S\right)={\displaystyle \frac{n\left(n-1\right)\left(2n+5\right)}{18}}\end{array}$$

(4)

$$\begin{array}{c}Z={\displaystyle \frac{S}{\sqrt{{\delta }^2}}}\end{array}$$

(5)

where X_j and X_k represent runoff values at times j and k, respectively (k < j); Sgn() denotes the sign function.

A significance level of α = 0.05 was selected for this study. The null hypothesis of “no trend” is rejected if the absolute value of the test statistic |Z| > Z_1−α/2, indicating a statistically significant increasing or decreasing trend in the runoff series. Specifically, Z > 0 suggests a significant increasing trend, and Z < 0 indicates a significant decreasing trend. Otherwise, the series shows no significant trend.

2.3.3. Geographically Weighted Regression

The geographically weighted regression (GWR) model [41] is widely applied to analyze spatial heterogeneity and driving factor. Unlike the global regression model that assumes a uniform relationships across space, GWR allows regression coefficients to vary dynamically with geographic coordinates, effectively capturing spatial non-stationarity [42]. Given the pronounced spatial heterogeneity of runoff distribution in Central Asia and the varying responses of runoff to precipitation and temperature across different regions, this study applies GWR using the “spgwr” package (version 0.6-35) in R version 4.1.1 to explore the spatial patterns of interactions between runoff and climatic factors (temperature and precipitation). The GWR model is expressed as follows:

$$\begin{array}{c}{Y}_i={\beta }_0\left({\mu }_i,v_i|b\right)+{\beta }_P\left({\mu }_i,v_i|b\right)x_{iP}+{\beta }_T\left({\mu }_i,v_i|b\right)x_{iT}+{\epsilon }_i\end{array}$$

(6)

where Y_i, x_iP, x_iT, and ${\epsilon }_i$ represent the dependent variable (runoff), independent variables (precipitation /temperature), and random error at spatial location i, respectively; $\left({\mu }_i,v_i\right)$ represents the geographic coordinates (longitude, latitude); ${\beta }_P\left({\mu }_i,v_i|b\right)$ and ${\beta }_T\left({\mu }_i,v_i|b\right)$ are spatially varying coefficients for precipitation and temperature, conditioned on bandwidth b; b is the kernel bandwidth controlling spatial smoothing; and ${\beta }_0\left({\mu }_i,v_i|b\right)$ is the local intercept.

The local coefficients ${\beta }_{P/T}\left({\mu }_i,v_i|b\right)$ are estimated using weighted least squares, with weights (w_ij(b)) assigned by a Gaussian kernel as follows:

$$\begin{array}{c}{w}_{ij}\left(b\right)=exp(-{\displaystyle \frac{d_{ij}^2}{b^2}})\end{array}$$

(7)

where d_ij is the great circle distance between location i and j.

The optimal bandwidth was determined by minimizing the Akaike information criterion (AIC) to balance model fitting accuracy and spatial smoothness, ensuring robust local parameter estimation. The optimal bandwidths for the three sets of data output by the H08, WaterGAP2-2e, and MIROC-INTEG-LAND models are 0.53° (≈50 km), 0.7079° (≈70 km), and 0.5516° (≈54 km), respectively. The derived bandwidths (50–70 km) correspond to 5–7 grid cells in our 5-min spatial resolution dataset, capturing subregional hydroclimatic patterns.

3. Results

3.1. Long-Term Mean Runoff in Central Asia

Figure 2 presents the spatial distribution of multi-year mean runoff (1950–2014) across Central Asia simulated by three models. The H08 model (Figure 2a) indicates that areas of high runoff are primarily concentrated in the southeastern regions of Central Asia (Kyrgyzstan and Tajikistan). Runoff values exhibit a declining gradient from the southeast and northern regions toward the central areas, with approximately 85% of the region showing annual runoff below 40 mm. The WaterGAP2-2e model (Figure 2b) presents a more spatially heterogeneous runoff pattern. In addition to elevated values in the southeastern region, high runoff is also simulated along major river corridors, with values ranging widely from 0 to 640 mm. By contrast, the MIROC-INTEG-LAND model (Figure 2c) depicts generally low runoff across the entire domain. Approximately 90% of the region records annual runoff below 40 mm, with only localized increases near the tri-border area of Uzbekistan, Kyrgyzstan, and Tajikistan. High-runoff zones are significantly smaller in extent than in the other two models. Despite differences in magnitude and spatial coverage, all three models consistently identify the mountainous headwaters of the Amu Darya and Syr Darya as key high-runoff regions.

Figure 3 depicts the spatial distribution of relative percentage change (RPC) in annual runoff across Central Asia for the period 2015–2080, relative to the historical baseline (1950–2014), as simulated by the three hydrological models under SSP126, SSP370, and SSP585 scenarios. The H08 model projects predominantly positive changes across all scenarios, with RPC values mainly ranging from 0% to 40%. Under higher-emission pathways, increases in runoff become more pronounced, particularly in central Kazakhstan. Negative changes are localized, mainly along the Turkmenistan–Uzbekistan border, especially in segments of the transboundary Amu Darya Basin. The WaterGAP2-2e model also projects widespread positive changes under SSP126 and SSP370, with stronger increases in the north than in the south. The extent and intensity of positive change expand under SSP3-7.0 relative to SSP1-2.6. However, under SSP5-8.5, spatial disparities become more evident: positive signals weaken in northern regions, while negative changes emerge and intensify, particularly over the Tianshan Mountains and Pamir Plateau. Notably, this model more explicitly captures runoff variability in riverine corridors. In these riverine zones, including the Irtysh River (northeast of Central Asia), Ural River (northwest of Central Asia), and the Amu Darya and Syr Darya basins, positive runoff changes are generally less pronounced than in adjacent areas, and even shift to negative under SSP585. The MIROC-INTEG-LAND model projects relatively modest changes across all scenarios. Slight positive changes dominate the region, interspersed with scattered areas of negative change. The overall spatial patterns show limited sensitivity to emission intensity, reflecting a more conservative hydrological response in this model.

3.2. Trend Analysis of Runoff in Central Asia

Under the historical scenario, the H08 model simulates predominantly decreasing runoff trends across Central Asia, with particularly strong negative signals in northern Kazakhstan, where Sen’s slope values range from −0.4 to −0.1 mm/year (Figure 4). In contrast, projections under all SSP scenarios reveal a spatial transition toward increasing trends, and the spatial extent of statistically significant trends expands. These increasing trends are primarily concentrated in high-altitude zones, extending from the eastern margins of Central Asia to the southeastern plateau. Under the high-emission scenario, the contrast between regions of increasing and decreasing trends becomes more pronounced. The WaterGAP2-2e model shows marked spatial heterogeneity in runoff trends during the historical period, although the extent of statistically significant trends is limited. Under future scenarios, both the magnitude and significance of increasing trends become more evident, particularly across central and eastern regions, with stronger responses under higher forcing scenarios. The MIROC-INTEG-LAND model simulates widespread decreasing runoff trends during the historical period, with localized positive signals in northern Kazakhstan and near the confluence of the five countries. Under future scenarios, increasing trends gradually become apparent, both in intensity and statistical significance. Compared to the other two models, this model presents a more conservative projection of future runoff increases, with generally weaker trend magnitudes.

Overall, all three models demonstrate a consistent shift from decreasing to increasing runoff trends under SSP scenarios relative to the historical period. Eastern regions show a higher proportion of statistically significant trends than the west, and the SSP585 induces greater increase trends in the north. However, inter-model differences in trend magnitude, spatial extent, and statistical significance underscore the structural uncertainty in hydrological modeling of climate change impacts in Central Asia.

The overall trend of runoff in the five Central Asian countries is shown in Figure 5. During the historical period, the H08 model simulates a slight declining trend in regional runoff, with a Sen’s slope of −0.022 mm/year. This trend reverses under all SSP scenarios, showing an overall increase in runoff that intensifies with scenario. The most pronounced increase occurs under the high-emission scenario (SSP585), indicating a pronounced response to enhanced warming and precipitation patterns. For the WaterGAP2-2e model, runoff exhibits a fluctuating upward trend in the historical period (slope: 0.1 mm/year), larger than that simulated by the MIROC-INTEG-LAND model. Under SSP scenarios, it shows a pattern similar to H08, but with greater sensitivity to high-emission forcing. The projected slopes under SSP370 and SSP585 reach 0.232 mm/year and 0.255 mm/year, respectively, representing an increase of approximately 40% compared to SSP126. The MIROC-INTEG-LAND model also indicates an upward runoff trend during the historical period, characterized by a gradual increase with interannual variability. Although runoff continues to increase under all SSP scenarios, the rates of increase remain moderate, with trend slopes consistently below 0.05 mm/year. Scenario differences are also less pronounced in this model compared to H08 and WaterGAP2-2e. Overall, all three models consistently project an upward trend in future runoff across Central Asia. The H08 and WaterGAP2-2e models exhibit stronger sensitivity to high-emission scenarios, while the MIROC-INTEG-LAND model suggests a more gradual and stable increase in runoff.

3.3. Influencing Factors

Based on the skewed distributions and diverse trends of annual mean runoff under different scenarios, and considering the complex topographic features of Central Asia, ranging from arid lowlands to glaciated highlands, we applied geographically weighted regression to investigate the spatial heterogeneity in runoff response to temperature and precipitation during the historical period.

Table 3. Comparison of OLS and GWR model results.

Indicators	H08		WaterGAP2-2e		MIROC-INTEG-LAND
	OLS	GWR	OLS	GWR	OLS	GWR
Sigma	37.40	13.91	102.60	77.55	19.89	11.37
R2	0.6142	0.9649	0.3531	0.7075	0.4921	0.8870

compares the fit goodness of the global OLS (ordinary least squares) model and the GWRF model. Compared to the OLS model, the GWR model yields a substantially lower sigma and a markedly higher R². Here, sigma refers to the standard deviation of residuals, indicating the average deviation between simulated and observed values; R² denotes the coefficient of determination, with values closer to 1 reflecting stronger and more robust explanatory power. These results indicate that accounting for spatial non-stationarity significantly improves the model’s ability to explain runoff variability.

In terms of the precipitation–runoff relationship (Figure 6), the H08 model shows a positive regression coefficient of 89.24%, mostly within the range of 0 to 0.8. This implies that a 1 mm increase in precipitation leads to a 0–0.8 mm increase in runoff, with the strongest sensitivities observed in high-altitude mountainous regions. Negative coefficients are limited to localized areas in the south of Central Asia, primarily in Turkmenistan. The WaterGAP2-2e model displays more spatial variability in coefficients, with a mix of positive and negative coefficients distributed throughout the region. Despite this heterogeneity, positive impacts remain dominate (86.56%), particularly across the Syr Darya and Aral Sea regions. The histogram of coefficients peaks around 0.3, with a notable proportion of negative values. The MIROC-INTEG-LAND model shows generally positive coefficients concentrated in the 0–0.6 range (82.92%), similar to the H08 model. The highest frequency of coefficients appears in this interval, while negative correlations are sparse and mainly confined to parts of northern Kazakhstan. Collectively, all three models suggest a robust and regionally consistent positive influence of precipitation on runoff generation.

The temperature–runoff relationships exhibit greater model-dependent variability. For the H08 model, most coefficients are clustered near zero, indicating a weak direct influence of temperature on runoff (Figure 7). Nonetheless, widespread slightly negative coefficients (ranging from −25 to 0) suggest that rising temperatures may suppress runoff, likely through enhanced evapotranspiration. The WaterGAP2-2e model exhibits the most pronounced spatial heterogeneity, with coefficients ranging from −100 to 100, suggesting substantial regional variation in hydrological response to warming. In contrast, the MIROC-INTEG-LAND model presents a distinct pattern, with more frequent positive coefficients than negative ones. Notably, regression coefficients around +20 are observed over the Pamir Plateau, likely reflecting the strong impact of warming on snow and glacier melt, characteristic of glacial hydrological system. Downstream regions, such as the Syr Darya Basin, also exhibit positive coefficients, further supporting the contribution of glacier meltwater from high-altitude sources to regional runoff.

4. Discussion

4.1. Model Selection and Uncertainty

Considering model complexity, global applicability, and compatibility with the 2015soc socio-economic scenario, this study selected three representative global hydrological models: H08, WaterGAP2-2e, and MIROC-INTEG-LAND. Their suitability for hydrological analysis in Central Asia has been previously validated [7]. Within the ISIMIP framework, these GHMs are driven by bias-corrected outputs from five CMIP6 GCMs. Therefore, our focus lies in inter-model comparisons to assess structural uncertainty, while uncertainty from climate forcing is mitigated by using multi-GCM ensemble means.

Global hydrological models simulate both natural hydrological processes and human interventions (such as irrigation, reservoir operations, and groundwater abstraction), thereby enabling integrated runoff modeling [43]. In this study, the three selected models produced distinct patterns in runoff magnitude and temporal response. The H08 model, for example, projects a transition from wet to drier conditions (“wet-to-dry”), as shown in Figure 2 and Figure 3, whereas WaterGAP2-2e and MIROC-INTEG-LAND simulate a “wet-to-wetter” trend under similar scenarios. These inter-model discrepancies in runoff estimates primarily stem from differences in the models’ core hypotheses and the cumulative differences in parameterizations of hydrological processes. Among them, the H08 model is process based and assumes that soil hydrology is the dominant control on runoff dynamics. WaterGAP2-2e emphasizes anthropogenic regulation of water resources, incorporating all five sectors of human water use [33], potentially increasing its sensitivity to anthropogenic influences. MIROC-INTEG-LAND places greater emphasis on the integrated feedbacks among climate, vegetation, and water resources. Compared to the other two models, it explicitly accounts for processes related to CO₂ impacts and their interactions with the land surface system [34,44]. In addition, the long simulation period from 1950 to 2080 may amplify inter-model differences due to the accumulation of uncertainty over time. The large spatial extent of Central Asia, which encompasses diverse hydroclimatic zones such as glacier mountains, semi-arid steppes, and arid lowlands, also contributes to greater variability in climate–runoff responses across models. Under high-emission scenarios like SSP585, factors such as increased evapotranspiration, meltwater dynamics, and human water demand become more influential. Despite these differences, all three models show broadly consistent spatial patterns of runoff distribution. In regions with sparse hydro-meteorological observations, model validation remains a challenge. However, due to the caution and confidentiality requirements of cross-border river data, as well as the lack or even suspension of many hydrological and meteorological stations since the early 1990s, the data that can be collected have stagnated [19]. Therefore, in data-scarce areas, multi-model ensembles provide a valuable probabilistic framework to fill data gaps and support water resource management under uncertainty [45].

4.2. Spatial Heterogeneity of Influencing Factors

Among the primary contributors to runoff, precipitation and snowmelt generate surface flow directly, whereas evapotranspiration, which including both soil evaporation and plant transpiration, reduces the water available for runoff—a process typically intensified by rising temperatures [46]. Previous modeling studies have shown that the effects of glacier mass and precipitation variability on runoff are both nonlinear and complex, particularly in snow-dominated basins such as those in Central Asia [47,48]. The GWR results in this study similarly reveal a generally positive relationship between precipitation and runoff, with regression coefficients predominantly ranging from 0 to 0.8. This influence is especially pronounced in high-altitude regions, reflecting the impact of orographic precipitation. In contrast, the temperature–runoff relationship is more complex. Negative correlations dominate across most regions, consistent with increased evapotranspiration under warming. However, positive temperature effects are observed in glacierized highlands and downstream river systems, particularly in the WaterGAP2-2e and MIROC-INTEG-LAND simulations, highlighting the runoff contribution from glacial and snowmelt under warming conditions. This response is particularly evident at seasonal timescales [49]. Despite the current buffering role of glacier melt, long-term trends indicate a significant decline in glacial water storage. Along the Tianshan Mountains, glacier mass has decreased by approximately 27% and area by 18% over the past five decades [50]. Although glacial meltwater currently contributes significantly to runoff at seasonal or short-term scales, continued retreat under persistent warming is expected to reduce future runoff volumes [51]. This underscores the need for adaptive strategies to mitigate future imbalances in water availability.

4.3. Implications for Water Resources Management and Outlook

In this study, geographically weighted regression was applied to annual mean values of variables under the historical scenario (1965–2014), aiming to minimize the influence of interannual climate variability and to highlight the long-term, stable climate–runoff relationships. Future research could incorporate time-series segmentation or employ spatiotemporally weighted regression approaches, which allow for joint assessment of both spatial and temporal non-stationarity [52]. Given the pronounced geographic heterogeneity in Central Asia, this study primarily focused on the climatic drivers of runoff in arid regions, particularly precipitation and temperature. However, the hydrological system in Central Asia is also influenced by non-climatic factors such as land use change and water infrastructure development. For instance, widespread irrigated agriculture substantially alters both surface runoff and evapotranspiration regimes. Incorporating these anthropogenic drivers in future analyses could help disentangle the complex mechanisms governing regional water availability.

The impacts of climate change in Central Asia are often expressed through escalating water resource crises. A prominent example is the desiccation of the Aral Sea, which has shrunk to less than 10% of its 1960 size, largely due to inefficient irrigation. Currently, irrigated agriculture accounts for approximately 70% of land use across the five Central Asian countries [53], yet more than half of the irrigation water is lost via leakage and evaporation. Based on GWR results, targeted water management strategies could be prioritized in regions with strong negative runoff–temperature correlations. Improving irrigation efficiency (by replacing flood irrigation with drip systems, expanding agroforestry along canals to reduce evaporation, and promoting drought-resistant crops) in arid lowlands is imperative. Meanwhile, with projected runoff increases under high-emission scenarios, proactive infrastructure measures such as expanding reservoir capacity (e.g., Rogun Dam) to store excess meltwater during wet seasons could play a key role in alleviating downstream water scarcity during dry periods.

Moreover, the runoff trends observed in this study are consistent with findings from Huang et al., which revealed substantial surface water loss across northern and western Kazakhstan [54]. Permanent surface water bodies have markedly declined in the Amu Darya Basin (where the Aral Sea is located), the Irtysh River Basin (northern Kazakhstan), and the Ural River Basin (western Kazakhstan). The mismatch between water supply and demand, combined with the uneven spatial distribution of water resources [55], underscores the urgent need for efficient water-saving technologies and coordinated transboundary water governance. Notably, these challenges resonate with those in geographically contiguous arid regions. As part of the contiguous arid/semi-arid belt spanning China’s Northwest region and Central Asia, Xinjiang Province shares key hydroclimatic and socioeconomic similarities with its neighboring Central Asian countries. Both regions exhibit glacier-fed river systems originating from the Tien Shan Mountains. Runoff observations from several major rivers in China’s Xinjiang region between 1957 and 2009 reveal a marked increase in natural runoff after the mid-1980s [19]. This change has been linked to a “warmer and wetter” climate trend. Our analysis of the 1965–2014 shows increasing runoff in the major rivers and headwater regions of Central Asia, while most arid lowland areas exhibit declining runoff. In both Xinjiang Province and Central Asia, water resource development in inland river basins must account for ecological water demands and balance upstream and downstream as well as cross-regional interests. Although both regions face water scarcity due to climate change and rising irrigation demand, the situation in Central Asia is further complicated by poorly maintained water infrastructure, extensive transboundary river systems, and limited regional cooperation. These factors amplify the urgency of integrating scientific projections into water governance frameworks, help mitigate the risk of resource-based conflicts.

5. Conclusions

This study leverages the ISIMIP3b multi-model framework to characterize spatiotemporal runoff variations and its climate drivers in data-scarce Central Asia under historical and SSP scenarios, using H08, WaterGAP2-2e, and MIROC-INTEG-LAND. Key findings from model intercomparison, trend analysis, and GWR are as follows:

All models consistently identify similar spatial patterns of mean runoff, with high values concentrated in the mountainous regions of Kyrgyzstan, Tajikistan, and southeastern Central Asia, and low values across western arid zones. Notable discrepancies are observed in runoff magnitude: WaterGAP2-2e simulates the highest historical annual mean runoff (60–80 mm), while MIROC-INTEG-LAND produces the lowest values (<15 mm). Under SSP scenarios, most regions are projected to experience increased runoff relative to historical baseline, except for WaterGAP2-2e in riverine areas under SSP585, which shows a 15% reduction.

Trend analysis reveals divergent model responses: H08 indicates a “wet-to-dry” transition, whereas the other models project a “wet-to-wetter” pattern. SSP scenarios generally intensify increasing runoff trends, with eastern Central Asia exhibiting higher trend magnitudes and statistical significance than western areas. H08 and WaterGAP2-2e demonstrate greater sensitivity to mid- and high-emission scenarios, showing trend slopes more than twice those under low-emission scenarios.

GWR results highlight precipitation as the dominant positive driver of runoff across most regions, particularly in high-altitude terrains. Temperature influence is spatially heterogeneous. While generally negative linked to increased evapotranspiration, WaterGAP2-2e and MIROC-INTEG-LAND show positive effects in the Pamir Plateau and downstream basins, attributed to climate warming-induced glacier melt. These findings underscore the role of model structural uncertainty and emphasize the need for integrated assessments in data-scarce regions. This research provides a multi-model perspective for Central Asian water resource management.