The mainstream region of the Tarim River Basin is situated in northern Xinjiang, China, between the southern alluvial plain of the Tianshan Mountains and the northern margin of the Taklimakan Desert. The study region stretches westward from the Alar region, where the Aksu, Yarkant, and Hotan rivers merge, eastward to the terminal Taitema Lake (Fig. 1). The study area covers approximately 40 956 km², featuring a gently eastward-sloping topography with elevations ranging from 805 to 1023 m.

The region exhibits a temperate continental arid climate, characterized by an annual mean temperature of 10.6 °C and total precipitation below 80 mm, mainly occurring from May to August, while the annual potential evapotranspiration is over 2500 mm (Tuoliewubieke et al., 2025; Cui et al., 2024; Ling et al., 2014). Water resources primarily originate from mountain glacier/snowmelt and upstream inflow diversions, exhibiting the arid-region characteristics of source-dominated supply with high spatiotemporal variability (Grashey-Jansen et al., 2014).

The Tarim River mainstream stretches approximately 1321 km in a west-east direction and is typically categorized into upstream, midstream, and downstream sections (Cui et al., 2024; Chen et al., 2006). The river exhibits a strongly seasonal flow pattern, with 70 %–85 % of the annual runoff occurring from June and September. Peak flows usually occur in July or August, whereas minimum flows are typically recorded in January and February. Surface water and groundwater systems within the region are closely connected, with frequent interactions influencing regional water availability.

Agriculture is the dominant economic activity in the Tarim River mainstream region. By 2015, approximately 1307 km² of land were under irrigation, of which about 1073 km² is cultivated land (Xinjiang Uygur Autonomous Region Bureau of Statistics, 2016). Major crops include cotton, maize, vegetables, and fruits. Since 1990, the cultivated land area has expanded by over 60 %, with oasis farmland continually encroaching into desert margins.

The area contains several large irrigation districts, where water for irrigation is primarily drawn from the Tarim River mainstream, supplemented by groundwater (Hao et al., 2015). Considerable differences in management and irrigation efficiency among these districts have placed significant pressure on the distribution of mainstream water resources, complicating regional water management.

The downstream ecosystem is extremely fragile. Ecological degradation has intensified since the 1950s with the area of Populus euphratica forests decreasing by approximately 86 % compared to historical conditions (Jiang et al., 2005). The river reaches between Daxihaizi Reservoir and Taitema Lake experienced prolonged channel desiccation, including a continuous dry-up of nearly 400 km at its most severe stage. As a result, Taitema Lake remained dry for extended periods, and the margins of the Taklimakan and Kuruk deserts approached near connection.

To mitigate ecological degradation in downstream reaches, China launched the Ecological Water Conveyance Project (EWCP) along the Tarim River mainstream in 2000, aiming to restore river connectivity, groundwater levels, and riparian vegetation in downstream areas (Gao et al., 2026; Wang et al., 2025; Jiao et al., 2022; Dou et al., 2022; Xu et al., 2007). According to recent monitoring and official reports, 26 ecological water conveyance events had been conducted by 2025, delivering a cumulative volume of approximately 10.2 billion m³.

The ecological water conveyance has substantially improved riparian vegetation coverage, wetland area, and groundwater levels along the downstream river corridor, playing a critical role in the recovery of ecosystem functions (Liao et al., 2020).

Despite these improvements, continued agricultural expansion upstream has increased diversion demands, while interannual inflow variability further strains ecological water regulation. Consequently, there remains a need to further optimize mainstream water resources allocation, improve irrigation water use efficiency, and achieve a dynamic balance between ecological water conveyance and agricultural development, in order to safeguard basin-wide ecological security and socio-economic sustainability.

The datasets used in this research encompass topography, land use, climate, hydrogeology, and socio-economic information, with a primary temporal coverage from 2002 to 2021.

Snow cover information was obtained from the High Asia MODIS daily snow cover fraction product developed by Qiu and Wang (2020). This dataset, derived using the MODIS Normalized Difference Snow Index (NDSI), offers daily data at a spatial resolution of 500 m and serves as a critical input for the snowmelt runoff model (SRM). Terrain-related information was sourced from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM), which has a spatial resolution of 30 m (NASA/METI/AIST/Japan Spacesystems and U.S./Japan ASTER Science Team, 2019). These elevation data were used to extract terrain attributes, including slope, aspect, and river networks.

Land use, vegetation, and soil property data were obtained from multiple sources. Land cover information was obtained from the FROM-GLC dataset (Gong et al., 2013), soil properties were derived from the Harmonized World Soil Database (Wieder et al., 2014), and vegetation attributes were derived from the Vegetation Map of China (Hou et al., 2019). Collectively, these datasets provided key inputs for simulating land surface processes and parameterizing soil-related processes in the coupled model.

Hydrogeological conditions were primarily derived from borehole data and regional hydrogeological maps reported by Li et al. (2003). Key parameters, including hydraulic conductivity and specific yield, were assigned based on dominant lithological units.

Meteorological forcing for the coupled surface water-groundwater model consisted of precipitation and temperature (maximum and minimum), which were obtained from the ERA5 reanalysis dataset (Hersbach et al., 2020). Model calibration and validation for the groundwater component relied on monthly groundwater level measurements from 139 monitoring wells distributed throughout the study area (Xue et al., 2024).

Daily streamflow records from five representative hydrological stations within the Tarim River Basin (2002–2021) were used to calibrate surface runoff simulations and model parameters.

To further evaluate the groundwater simulations of the GSFLOW model, terrestrial water storage anomaly (TWSA) data from the GRACE and GRACE-FO satellite missions (Save et al., 2016; Save, 2020) were used for model validation. In addition, the coupled model was cross-validated using evapotranspiration (GLEAM) (Martens et al., 2017) and soil moisture (Chinese Soil Moisture Dataset) (Meng et al., 2021a, b). A comprehensive summary of all datasets employed in this study is provided in Table S1.

Socio-economic data were primarily collected from the China Statistical Yearbook (National Bureau of Statistics of China, 2021), Xinjiang Water Resources Bulletin (Xinjiang Uygur Autonomous Region Water Resources Department, 2021), and the National Agricultural Product Cost-Benefit Compilation together with annual crop price statistics (Ministry of Agriculture and Rural Affairs of Xinjiang, 2025). These datasets provide information on crop planting areas, yields, and fertilizer and pesticide application rates.

Although the primary focus of this study is the Tarim River mainstream region, accurate representation of surface water-groundwater exchanges necessitates hydrological simulations at the scale of the entire Tarim River Basin. The hydrological simulation framework consists of the SRM and the GSFLOW model. The SRM is applied to simulate snow and glacier melt runoff at the mountain outlets, and the simulated results are subsequently used as the upstream inflow boundary for the GSFLOW model. GSFLOW model is then employed to simulate hydrological processes across the Tarim River plain and downstream desert areas. The basin-scale simulation outputs are then used to extract boundary conditions and fluxes for the mainstream region, which serves as the domain for the subsequent multi-objective WEA optimization.

To promote the coordinated development of agricultural production, ecosystem protection, and sustainable water resources use, a multi-objective water-ecosystem-agriculture (WEA) optimization model was developed for the Tarim River Basin mainstream region. Based on regional geomorphological features and hydrological conditions, the mainstream area was initially categorized into three major sections: upstream, middle, and downstream reaches. Each section was further partitioned into seven agricultural management sub-units. This zoning scheme primarily follows administrative divisions in combination with the spatial distribution of major irrigation districts. The seven agricultural sub-units include the Shaya Irrigation District (Zone 1), Shaya Prison Irrigation District (Zone 2), Kuqa Irrigation District (Zone 3), Kuqa Sheep Breeding Farm Irrigation District (Zone 4), Luntai Irrigation District (Zone 5), Yuli County Irrigation District (Zone 6), and the Second Agricultural Division Tarim Irrigation District (Zone 7). The geographical arrangement of these sub-units is illustrated in Fig. 1b.

Each GSFLOW simulation takes approximately 18 h, presenting a prohibitive computational cost for the simulation-optimization framework. To overcome this, a surrogate model was constructed using a radial basis function neural network (RBF-NN) to represent the relationship between key GSFLOW parameters and the four objective functions in the multi-objective optimization model. The RBF-NN is a multilayer feedforward network with strong nonlinear approximation capability and high computational efficiency. Training samples are generated using Latin hypercube sampling (LHS) to ensure adequate coverage of the parameter space. Following accuracy validation, this surrogate model replaces the original GSFLOW model within the simulation-optimization loop.

To elaborate further, the input variables of the surrogate model are the decision variables to be optimized. Specifically, they include the planting areas of six major crops (cotton, maize, oil crops, vegetables, melons, and fruits) across seven irrigation zones, as well as the allocation coefficients of ecological water use during key months (May–September), resulting in a total of 47 decision variables. The output variables of the surrogate model consist of two hydrological variables simulated by the coupled hydrological model (GSFLOW), namely the change in the average groundwater depth in the study area and the area of the terminal lake. For surrogate construction, Latin hypercube sampling was first applied to generate samples within the input variable space. Each sample set was then used as input to the GSFLOW model to obtain the corresponding outputs, thereby forming an input-output dataset. The dataset was subsequently randomly divided into a training set (70 %) and a test set (30 %). The training set was used to train the radial basis function neural network (RBF-NN), while the testing set was used to evaluate the predictive accuracy and generalization capability of the surrogate model.

Simulated daily runoff from the SRM model agrees well with observations, particularly during spring snowmelt, where both the timing and magnitude of peak flows are accurately captured. The SRM achieves Nash-Sutcliffe efficiency (NSE) values between 0.72 and 0.96, demonstrating its strong capability to accurately simulate seasonal snowmelt-driven runoff. These results demonstrate the SRM's capability in capturing the temporal variability of mountain inflows, making it suitable for providing reliable upstream boundary conditions to the GSFLOW model. The spatial distribution and numbering of the 22 sub-basins with available observations, as well as the corresponding SRM simulation results and comparisons with observed runoff, are shown in Figs. S1–S3 in the Supplement.

As shown in Fig. 3, GSFLOW simulations closely match observed monthly streamflow at Alar, Xinqiman, Yingbaza, Wusiman, and Qiala stations on the Tarim River main stem, with NSE values ranging from 0.6 to 0.7. This demonstrates that the model adequately represents runoff dynamics, including both high- and low-flow periods.

The agreement between observed and modeled groundwater levels at 139 monitoring wells throughout the validation period is shown in Fig. 4. The NSE of the groundwater simulation reached 0.96, demonstrating that the GSFLOW model can accurately capture groundwater dynamics at the observation points. It should be noted that the monitoring wells are unevenly distributed and limited in number, being mainly concentrated in the main stem of the Tarim River and the Hetian region.

The reliability of the GSFLOW model was further evaluated by comprehensive validation using multi-source remote sensing data, including evapotranspiration (ET), soil moisture (SM), and terrestrial water storage change (TWSC). The results of the comparison indicate that good overall consistency between model simulations and independent observations. Specifically, the simulated TWSC achieved an NSE of 0.66 against GRACE data, soil moisture an NSE of 0.61 against SM products, and evapotranspiration an NSE of 0.71 against the GLEAM dataset (see Figs. S6–S8). Collectively, these results indicate that the coupled SRM-GSFLOW model reliably simulates surface water-groundwater interactions in the Tarim River main stem, providing a robust hydrological foundation for optimizing agricultural and ecological water use.

The variation of the surrogate model's NRMSE and R2 with the number of training samples is shown in Fig. S9. As the number of samples increases, the surrogate accuracy steadily improves. When 2842 samples are used, the coefficient of determination (R2) reaches 0.98, corresponding to an NRMSE of 0.016. Additionally, a five-fold cross-validation was conducted to assess overfitting risk; the results are summarized in Table S2 in the Supplement, which shows consistent RMSE across folds.

This study develops a multi-objective water-ecosystem-agriculture (WEA) optimization framework to systematically elucidate the coordination between cropping structure adjustment and water allocation in the Tarim River mainstream. The results reveal pronounced trade-offs between agricultural economic benefits (fAB) and ecological objectives (fLA and fGL), as well as environmental objectives (fTN), confirming the widely recognized “water-food-ecosystem” dilemma in arid inland river basins (Li et al., 2024). Compared with the ecological crises driven by upstream-downstream water competition in the Aral Sea Basin (Azimov and Avezova, 2022), the multi-objective optimization approach adopted here provides a quantitative tool to explicitly characterize and manage such conflicts.

Notably, conflicts between fAB and ecological objectives (fGL, fLA) can be mitigated through scientifically designed adjustments to cropping structure and irrigation practices. This suggests that sustaining agricultural output can be compatible with ecological preservation, providing fresh perspectives for managing water resources in arid regions.

Hydrological conditions strongly regulate system-level trade-offs. During dry years, system resilience declines markedly and multi-objective conflicts intensify, consistent with observations from the Aral Sea Basin where upstream-downstream tensions are exacerbated under drought conditions (Saidmamatov et al., 2020). In contrast, wet years provide greater management flexibility and create critical windows of opportunity for achieving multi-objective coordination through structural adjustment rather than rigid water allocation constraints.

The compromise solutions identified in this study exhibit robust overall performance across different hydrological scenarios. Their strength lies in shifting the paradigm from a “win-lose” trade-off toward a more balanced “shared-benefit” outcome. For example, under dry-year conditions, solution S5 maintains moderate agricultural economic returns while effectively safeguarding ecological water requirements and controlling nitrogen pollution risks. This structurally optimized pathway aligns with recent practical efforts in the adjacent Yanqi Basin, where cropping structure adjustment has been used to achieve simultaneous economic and ecological gains (Tang et al., 2024). The present study further strengthens such empirical findings by providing quantitative, system-level evidence.

Spatial heterogeneity analysis underscores the necessity of differentiated management strategies. The midstream area exhibits the highest sensitivity to multi-objective trade-offs and thus requires more refined water-use regulation. Although the downstream area shows expansion potential during wet years, strict adherence to ecological thresholds remains essential. These findings are highly consistent with the differentiated management principles embodied in Xinjiang's “Three Red Lines” water policy (Tao et al., 2012).

Dynamically optimizing cropping structure is a key pathway to multi-objective coordination. As hydrological conditions improve, the proportion of high-benefit crops can be moderately increased, but such adjustments must remain firmly constrained by ecological safety considerations. In downstream ecologically sensitive zones, even under relatively abundant water availability, the expansion of high water-consuming crops should be strictly limited. This “ecology-first” optimization principle is critical for maintaining long-term ecological security in arid river basins.

Compared to previous studies in the Tarim River Basin that primarily focused on single or dual objectives, the multi-objective optimization framework developed in this study provides a more comprehensive analytical approach for coordinating WEA conflicts. By integrating a coupled surface water-groundwater model with multi-objective decision-making methods, this framework transforms ecological responses from static constraints into optimization objectives that can be dynamically adjusted with management strategies. This approach enhances the ability to characterize the relationship between agricultural water use and ecological feedback. Although simplifications in the coupled hydrological model are inevitable, it still provides a more structured cognitive perspective for analyzing trade-offs within the WEA system.

When this method is to be extended to other basins, its applicability does not depend on the simple transplantation of models or data, but on the ability to implement spatially differentiated management according to local conditions. For example, in regions lacking hydrogeological data, a more generalized hydrological model could be used to represent water cycle processes. When the accuracy of crop spatial distribution data is limited, statistical data can be used to impose total area constraints at the regional scale. Furthermore, attention should be paid to uncertainty factors that significantly influence the optimization results (such as key hydrogeological parameters and agricultural management parameters), and their impact on decision-making outputs should be quantified through sensitivity analysis to provide necessary risk warnings for practical applications. The research framework presented in this study may provide useful insights for other basins facing similar water resource challenges, although it should be further adapted and validated according to local conditions.

The practical feasibility of the optimized solutions requires further field validation. Owing to the lack of long-term, large-scale monitoring data that simultaneously capture cropping structure adjustment and ecological water conveyance processes, the real-world applicability of the identified Pareto-optimal solutions remains uncertain. Future research should integrate data from ecological water regulation projects and water-saving agriculture demonstration programs to further calibrate and validate model outputs.

Multiple sources of uncertainty have not yet been systematically quantified or propagated through the optimization process. Key inputs, including climate forcing data, hydrogeological parameters, crop water requirement coefficients, and fertilizer-related parameters, all contain uncertainties that may accumulate and propagate step by step, affecting the shape and range of the Pareto front as well as the stability of compromise solutions. In addition, scale transformation during model construction – coupling basin-scale hydrological simulations with grid-based land use and crop distribution data – introduces errors from spatial scaling and parameter aggregation that are difficult to fully quantify. The influence of future climate change on runoff patterns and crop water requirements has also not been explicitly addressed. Additionally, the proportional adjustment method used for groundwater abstraction represents a simplified treatment. When the intensity or spatial distribution of groundwater abstraction changes significantly, the model should be updated and recalibrated using the latest statistical data to ensure the reliability of simulation and prediction results. Future research may introduce a Hierarchical Bayesian Model (Wu et al., 2010) to quantify and control the impacts of different uncertainty sources on optimization solutions, providing more robust management strategies for decision makers. Including climate change scenarios constitutes another key avenue for future studies.

This study employs the Export Coefficient Model to estimate agricultural nitrogen loads. This method offers high computational efficiency for strategic, long-term analyses at the watershed scale, making it suitable for the rapid assessment of pollution risks under multiple scenarios. However, as a static, spatially lumped empirical model, it is inherently limited in its ability to accurately describe the transport and transformation processes of nitrogen within the soil-groundwater system and their spatial heterogeneity under different irrigation or fertilization regimes. Therefore, the model is better suited for comparative analysis and relative risk assessment among different management schemes, rather than for the precise, process-based quantification and prediction of pollution.

Although the GSFLOW model can simulate coupled surface water-groundwater processes, its representation of vegetation mainly relies on predefined crop coefficients and canopy parameters, which makes it difficult to fully capture the spatiotemporal dynamics of vegetation growth and their influence on the hydrological cycle. Recent studies have shown that assimilating satellite-derived vegetation indices (e.g., Leaf Area Index, LAI) into hydrological models can significantly improve simulations of evapotranspiration, soil moisture, and groundwater recharge. For example, studies by Paul et al. (2021), Bian et al. (2019), and Li et al. (2009) demonstrated that assimilating MODIS-LAI data can effectively improve watershed-scale runoff and evapotranspiration simulations. At the irrigation district scale, Zafarmomen et al. (2024) further showed that assimilating high-resolution Sentinel-2 LAI into hydrological models can more accurately represent the effects of vegetation dynamics on evapotranspiration, irrigation return flows, and groundwater recharge, thereby improving the simulation of surface water-groundwater interactions. Future research could build upon these approaches by assimilating multi-source remote sensing LAI data into the coupled modeling framework used in this study, thereby improving the representation of hydrological processes, particularly the effects of ecological water replenishment, and providing more reliable distributed hydrological state information for multi-objective optimization. The application of higher-resolution process-based models could further enhance the precision and reliability of system simulations.