Introduction

As the process of globalization accelerates, the rapid growth of the world’s population has led to a continuous rise in food demand, which has become one of the major challenges faced by the world1. Food production is highly water-intensive, accounting for over 70% of global freshwater withdrawals2. However, climate change and growing competition for water resources have exacerbated water scarcity, which has become a key constraint on food production3. Studies project that half of the world’s farmland will face water scarcity in the future4. As a global population and agricultural powerhouse, China nourishes nearly 20% of the world’s population with only 7% of the global cultivated land5,6. However, agricultural water consumption accounts for over 60% of the total water consumption in China, while the irrigation water use efficiency coefficient is only 0.576, significantly lower than the 0.7–0.8 level observed in developed countries7,8. Therefore, it is imperative to develop solutions that balance food production with optimal water use, ensuring food security while reducing water demand, thereby promoting sustainable agricultural practices.

Agriculture closely links water, food, energy, and carbon within an integrated water-food-energy-carbon nexus9. This nexus is closely related to the United Nations Sustainable Development Goals (SDGs), specifically SDG 2 (Zero Hunger), SDG 6 (Clean Water and Sanitation), SDG 7 (Affordable and Clean Energy), and SDG 13 (Climate Action), with water resources as the core element10. However, current research consensus indicates that China’s agricultural irrigation has not achieved ideal coordination among water, energy, and carbon emissions11. The waste of irrigation water resources and the irrational energy structure in agriculture, coupled with the low contribution of clean energy substitution, have led to persistently high agricultural energy consumption, further exacerbating carbon emissions and intensifying supply-demand imbalances among water, food, energy, and carbon emissions12. Under the commitment to carbon neutrality, the Chinese government places high importance on energy transition and carbon emission reduction, which brings new opportunities for agricultural sustainable development13,14. The water-food-energy-carbon nexus-based optimization of water resource allocation, coupled with enhanced irrigation techniques and energy structure transformation, effectively reduces agricultural water demand while lowering energy consumption and carbon emissions from irrigation. This integrated approach improves the water-food-energy-carbon nexus, facilitating the transition toward green, low-carbon, and sustainable agricultural development15,16.

However, the lack of long-term, reliable large-scale observational data and the spatiotemporal heterogeneity of farmland management pose major challenges to research on optimization of water resource allocation17. Agricultural systems exhibit notable spatial heterogeneity, with regional differences in climatic conditions, soil types, and water resource endowments leading to distinct variations in crop yields and planting structures18,19. For example, in northern China, where water resources are scarce, drought-resistant crops, such as wheat and corn, are predominant, whereas in southern China, where water resources are relatively abundant, the proportion of water-intensive crops like rice is higher20,21. This spatial heterogeneity not only creates marked differences in regional water demand but also complicates the matching between water supply and crop water requirement patterns22. Moreover, the continued climate change and increasing frequency of extreme weather events further amplify the uncertainty in crop yields and planting structures, posing even greater challenges for the optimal allocation of water resources23,24,25,26. To address these challenges, spatial-scale optimization models offer innovative solutions to reconcile the mismatch between heterogeneous water demand and spatiotemporally uneven water resource distribution27,28. Nevertheless, existing studies remain limited, as most focus on macro- or regional-scale water resource management with insufficient attention to fine-scale dynamics of meteorological conditions, soils, and cropping systems29. As a result, the complex linkages between crop distribution and irrigation water requirements are often insufficiently captured, and the changing dynamics of crop suitability under climate and management shifts are not fully considered. These limitations constrain the accuracy and adaptability of spatiotemporal irrigation optimization, highlighting the need for high-resolution approaches that integrate multi-source data and decision-analytic techniques. This tendency has hindered adequate representation of the interactions between high-resolution crop distribution and irrigation water requirements, and has also led to insufficient consideration of the dynamics of crop suitability under changing climate and management conditions. These oversights have severely constrained the progress in improving the accuracy and adaptability of spatiotemporal irrigation optimization. Therefore, integrating multi-source data and decision-analytic techniques at finer geographic scales to enable precise prediction and coordination of regional crop-water supply-demand relationships, and to facilitate spatiotemporally optimized water allocation, is critical for enhancing the overall efficiency of the water-food-energy-carbon nexus.

Here, we have developed a hybrid method that combines crop suitability assessment, mechanistic models of spatial planting structure simulation, machine learning, and multi-objective optimization algorithms to formulate optimal irrigation schemes at the kilometer scale for China’s staple crops and to evaluate their potential environmental impacts. Firstly, we quantified the crop water supply and demand at the kilometer scale. Based on the coupled model of the Maximum Entropy Model (Maxent) and Patch-generating Land Use Simulation model (PLUS), the impacts of spatial suitability for crop growth and changes in planting patterns on irrigation water requirements were analyzed to accurately predict water resource demands under different arable land patterns. Secondly, we developed a decision-making model for the water-food-energy-carbon nexus, optimizing irrigation water volumes at the monthly and kilometer scales under various scenarios, including future climate conditions, energy transitions, improvements in irrigation technologies, and changes in cropping patterns. Finally, based on the optimized water volume results, we explored the spatiotemporal dynamic impacts of agricultural energy consumption and carbon emissions under the optimal irrigation water usage conditions. This study contributes a framework that strengthens the link between high-resolution crop suitability, irrigation demand prediction, and resource–environmental sustainability. The framework and mechanisms of this study are detailed in Fig. 1.

Fig. 1: Mechanism diagram.
Fig. 1: Mechanism diagram.
Full size image

SDG symbols are credited to the United Nations.

Results and discussion

Historical evolution of irrigation water use and crop yield

The simulation of spatial planting structures serves as the foundation for grid-based water allocation optimization. We calibrated the Maxent-PLUS model using two phases of farmland dynamics (2000s and 2010s) and evaluated the model’s performance for future periods (2020 s and 2030 s). The regions where crop changes are expected were identified, based on the variation range of the spatial farmland distribution. Figure 2 presents the results of comparison by R2 and RMSE for the training and testing datasets across four machine learning models. Among the four models tested (XGBoost, Random Forest, SVR, and FNN), Random Forest (RF) consistently demonstrated the best balance between accuracy and robustness. For rice, maize, and wheat, RF achieved testing R2 values of 0.91, 0.98, and 0.95, respectively, with notably lower RMSE on the testing datasets (0.27, 0.23, and 0.29) compared with the other models, indicating stronger generalization performance. Although XGBoost and SVR also achieved high R2 values, they exhibited higher testing errors (e.g., RMSE of 0.61 for rice and 0.64 for maize in XGBoost), suggesting less stability. FNN, while competitive in training, suffered from larger testing errors (e.g., RMSE 0.76 for rice and 0.94 for wheat), indicating overfitting. Therefore, the RF model was ultimately selected for yield prediction due to its superior accuracy on unseen data and robustness across crop types, ensuring both reliability and practical applicability in subsequent optimization analyses.

Fig. 2: Results of XGBoost, RF. SVR, FNN.
Fig. 2: Results of XGBoost, RF. SVR, FNN.
Full size image

a, c, e represent the training and testing accuracy for rice, maize, and wheat, respectively; b, d, and f represent the feature vector and label for rice, maize, and wheat, respectively.

Figure 3 illustrates the simulation results for RCP245-2030s, with the spatial distribution patterns of the three crops shown in Fig. 3a. Based on the grid-scale crop distribution simulation results from the Maxent-PLUS model, we established a spatial dataset. By estimating crop water requirements and effective rainfall at the grid scale, we calculated the spatial irrigation water demand for each grid. Using the Random Forest method, we assessed the response effects of crop yields to spatial environmental variables. The different colors represent the gradient of crop yields and irrigation water demand. On a national scale, the average irrigation water demand per unit area for rice, maize, and wheat is 596 mm, 503 mm, and 461 mm, respectively, while their average yields are 7.31 t ha−1, 4.81 t ha−1, and 5.88 t ha−1, respectively.

Fig. 3: Simulation results.
Fig. 3: Simulation results.
Full size image

a Grid-based irrigation water use and crop yield under future scenarios. b Yield for different crops. c Crop planting area by province. d Irrigation demand of different crops by province. Trend analysis of e irrigation amount, f planting structure, g energy proportion, and h irrigation method proportion under different scenarios Note: Mainland China is primarily divided into six agricultural regions. Northern Region: includes Beijing, Tianjin, Hebei, Shanxi, Inner Mongolia, Shandong, and Henan. Northeast Region: includes Heilongjiang, Jilin, and Liaoning. Middle-Lower Yangtze Region: includes Shanghai, Jiangsu, Zhejiang, Jiangxi, Anhui, Hubei, and Hunan. Northwest Region: includes Qinghai, Shaanxi, Gansu, Xinjiang, and Ningxia. Southwest Region: includes Chongqing, Guizhou, Sichuan, Yunnan, and Xizang. South Region: includes Fujian, Guangdong, Guangxi, and Hainan.

Further analysis was conducted by aggregating the pixel-level spatial simulation results to the provincial and agricultural region scales. As shown in Fig. 3b, the Middle-Lower Yangtze region has the highest total rice yield, approximately 9.28 × 106 t. The Northeast region leads in total maize production, reaching about 8.40 × 106 t, while the North region has the highest total wheat yield, approximately 8.31 × 106 t. The planting area results for each region are shown in Fig. 3c. Hunan Province ranks first nationwide in rice planting area, with approximately 3.97 million hectares for early rice and 3.96 million hectares for double-cropped rice. Heilongjiang Province has the largest maize planting area, totaling 7.47 million hectares. The largest spring wheat planting area is in Xinjiang, covering approximately 620,000 hectares, while Henan Province holds the largest winter wheat planting area, reaching about 6.04 million hectares. As shown in Fig. 3d, Heilongjiang Province has the highest irrigation demand, reaching 61.28 × 109 m3, while Beijing has the lowest irrigation demand at 0.11 × 109 m3. Among the agricultural regions, the North region has the highest total irrigation demand at 175.79 × 109 m3, whereas the South region has the lowest total irrigation demand at 34.99 × 109 m3. Fig. 3e, f, respectively, show the planting area and irrigation demand every 10 years from the 2000s to the 2030 s. Over this 40-year period, the planting area of major staple crops (rice, maize, and wheat) in China steadily increased by 28.85%, expanding from approximately 7 × 107 ha in the 2000s to nearly 11 × 107 ha in the 2030 s. Growth slowed between the 2020 s and 2030 s, with similar trends observed across different SSP scenarios. Irrigation demand increased from approximately 15 × 1010 m3 in the 2000s to nearly 60 × 1010 m3 in the 2030 s.

The results of the planting area and irrigation demand indicated that future climate change and agricultural expansion would continue to drive the rise in irrigation demand. Figure 3g illustrates the evolving trend in energy structure. The proportion of coal usage has gradually declined, while the share of renewable energy sources, such as wind and solar, has increased markedly, indicating a transition toward a low-carbon energy system. Meanwhile, hydropower remains relatively stable, and the contributions of nuclear energy and diesel have slightly decreased. This adjustment in the energy structure is likely to have a profound impact on the sustainability of agricultural production. Figure 3h reveals the changing pattern of irrigation technologies. Although surface irrigation remains the dominant approach, its proportion is gradually decreasing. In contrast, the adoption rates of micro-irrigation and sprinkler irrigation are steadily increasing. This shift suggests that as agricultural water pressure intensifies, more efficient water-saving irrigation technologies are being promoted to optimize water resources utilization and meet the rising demand for agricultural water.

Optimization and water-saving efficiency analysis

Irrigation water productivity (IWP) is an important indicator for measuring crop irrigation water efficiency and is also one of the key parameters for evaluating the effectiveness of management measures. To quantitatively analyze the impact of the optimization model, we compared the changes in IWP before and after optimization. As shown in Fig. 4, the IWP markedly improved after optimization, indicating that the optimization model has a positive effect on enhancing water use efficiency. In particular, the improvement in IWP is most pronounced in the southern regions (Hainan, Guangxi) and some southwestern provinces (Yunnan, Guizhou). This may be due to the relatively abundant water resources in these areas, the large planting area of rice, and the high proportion of traditional irrigation technologies (such as surface irrigation) in some regions, which leads to lower water resources utilization efficiency. Over the decade (2000s–2010s), irrigation water productivity showed an improvement in several regions. For instance, in the middle and lower Yangtze River Basin, Zhejiang province experienced a 2.68% increase in IWP, Hubei province saw a 1.80% increase, and Hunan province improved by 1.30%. These improvements are likely related to the upgrading of irrigation infrastructure (Sprinkler irrigation in Zhejiang increased by 3.3%, and micro-irrigation increased by 3.1%) and the implementation of refined management measures in the region. In addition, the IWP in Northeast China (Heilongjiang, Jilin, and Liaoning) increased by 2.05%, 1.63%, and 1.20%, respectively. Heilongjiang and Jilin improved their irrigation systems, reducing energy consumption per unit of water by 30% and 16%, respectively. Meanwhile, in Liaoning, the promotion of water-saving irrigation technologies (with the proportion of micro-irrigation increasing by 22% points) led to a 20% increase in energy consumption for precision irrigation equipment, indicating that modern water-saving technologies may trade energy costs for improved water efficiency in the short term. Meanwhile, some regions experienced smaller growth, such as Gansu (0.40%) and Qinghai (0.30%) in the northwest, which may be due to water scarcity in these areas, limiting further optimization opportunities. Based on the previous analysis, we further optimized the model by implementing more precise irrigation measures and adjusting irrigation timing to improve water use efficiency. For example, in Chongqing, the IWP increased by 2.71% in the 2010s, highlighting the essential role of effective irrigation management strategies in enhancing water resources efficiency. This result clearly demonstrates the practical application potential of the optimization model, offering a basis for achieving sustainable agricultural water use.

Fig. 4: IWP before and after optimization in historical scenarios.
Fig. 4: IWP before and after optimization in historical scenarios.
Full size image

Note: This figure includes two sections (upper: 2000s; lower: 2010s). The maps display optimized irrigation water per unit area in historical periods; adjacent panels show optimized IWP values and their growth rates for each province.

This study developed a kilometer-scale grid-based irrigation water optimization model for China, demonstrating substantial advantages over traditional regional zoning approaches in both spatial resolution accuracy and technological applicability. By precisely characterizing the spatial heterogeneity of climatic conditions, cropping patterns, and soil properties at 1 × 1 km resolution, the grid-based model overcomes the limitations of homogenization assumptions inherent in county-level administrative units, thereby enabling precise quantification of crop water requirement differentials. Multi-climate scenario simulations supported by high-resolution datasets effectively decode spatiotemporal variation patterns in agricultural water demand, providing scientific foundations for dynamic optimization of irrigation scheduling. The established gridded decision-making framework incorporates a hierarchical irrigation water demand management system, which simultaneously identifies high water-consumption grids (>500 mm season−1) in southern paddy systems and low-efficiency water use units (IWP < 1.0 kg m−3) in northwestern rainfed agriculture areas. This spatially explicit approach facilitates the development of targeted “one-grid-one-policy” allocation strategies tailored to local hydro-agronomic constraints.

Future water consumption analysis on energy use and environmental impact

The proposed optimization method provides agricultural water-energy-carbon emission strategies under various future scenarios. Figure 5a shows the irrigation water consumption month by month and Fig. 5b illustrates the agricultural water flow patterns across various regions in China by month, along with the complex relationships between agricultural water use, energy consumption, and environmental impacts. From the overall structure of the Sankey diagram, the left side features six major regional nodes (e.g., North, Middle-Lower Yangtze, Northwest, Southwest, Northeast, and South), highlighting notable differences in agricultural water use across regions. According to the flow thickness analysis, the North region has the highest agricultural water consumption, reaching 238.99 × 109 m3, accounting for 31.26%. The Middle-Lower Yangtze region ranks second with 130.25 × 109 m3, making up 28%. This is closely linked to its intensive rice cultivation and extensive farmland (with the North region accounting for 35% of the country’s total). In contrast, agricultural water flow in the South and Northwest regions is relatively lower, at 52.9 × 109 m3and 64.36 × 109 m3, respectively. This may be attributed to smaller farmland areas in these regions and a predominance of dryland farming practices.

Fig. 5: Optimized agricultural water-energy-carbon strategies.
Fig. 5: Optimized agricultural water-energy-carbon strategies.
Full size image

a Monthly irrigation water consumption. b Water consumption impacts on energy and environment.

On the water supply side, surface water accounts for 82% of total irrigation water use, while groundwater contributes 17%. Additionally, the model’s monthly water consumption decision variables reveal the seasonal variations in agricultural water use. For instance, agricultural water flow peaks in summer (May to August), reaching 515.74 × 109 m3, which is closely linked to increased crop evapotranspiration demand under high temperatures. Notably, the peak growth period of winter wheat in North China usually occurs in spring, requiring a certain amount of spring irrigation to cope with drought. As a result, the total irrigation water use in northern regions in March was relatively high, reaching 4.877 × 109 m3, equivalent to 1.722 m3 ha−1 per hectare. In contrast, during winter (December to February), as crops enter dormancy, irrigation water use significantly decreases, with the total volume dropping to 7.693 × 109 m3, equivalent to 33 m3 ha−1 per hectare.

To further illustrate the optimization outcomes, our analysis emphasizes the spatial and temporal heterogeneity of irrigation demand. The North not only exhibits the largest total water use but also strong seasonal fluctuations, driven by winter wheat spring irrigation and summer peaks. In contrast, the South and Middle-Lower Yangtze show more concentrated seasonal patterns linked to rice cultivation, while overall demand in the South remains relatively stable due to higher rainfall contributions. These results highlight that the optimization needs and benefits vary considerably across regions and months, underscoring the necessity of differentiated water management strategies.

Greenhouse gas emissions related to irrigation activities include emissions from electricity consumption and degassing emissions from groundwater extraction, which are 5.66 × 106 t CO2eq and 48.61 × 106 t CO2eq, respectively. In particular, the provinces of Heilongjiang (Northeast region), Henan, Shandong, Inner Mongolia, and Hebei (Northern region) rely heavily on groundwater from March to May, leading to higher degassing emissions from groundwater extraction in these areas, with a total emission of 36.69 × 106 t CO2eq. Greenhouse gas emissions unrelated to irrigation activities account for a considerable portion, making up 81% of the total. Specifically, emissions from fertilizers (nitrogen, phosphorus, and potassium) amount to 174.35×106 t CO2eq, agricultural inputs contribute 45.54 × 106 t CO2eq, and land emissions (including CH4 from rice fields and N2O) total 14.99 × 106 t CO2eq.

The combination of energy types and irrigation technologies has a certain impact on improving resource efficiency and reducing carbon emissions. In the past, heavy reliance on coal and diesel to power surface irrigation resulted in both high energy consumption and elevated carbon emissions. For example, in the North region under the 245-2020s scenario, the energy consumption per unit area in crop cultivation was 304 kWh ha−1, with an IWP of 1.47. By the 245-2030s scenario, with the gradual integration of low-carbon energy and efficient irrigation technologies, energy consumption per unit area decreased to 190 kWh ha−1, IWP increased to 1.51, and carbon emission intensity per unit area reduced by 13%. This validates the feasibility of the “water-saving, energy-saving, and carbon-reduction” win-win pathway.

Sensitivity analysis of IWP, EQ, and GHG to irrigation water volume

This study examined the trade-offs between IWP, energy demand (EQ), and greenhouse gas emission (GHG) under different scenarios (RCP126, RCP245, and RCP585) across six regions during the 2020s and 2030s (Fig. 6a, b). Furthermore, a Sobol sensitivity analysis was performed to assess how different decision variables (including irrigation water sources, irrigation technologies, and seasonal factors) influence these objectives, with the aim of identifying region-specific dominant constraint patterns and establishing scientific foundations for developing spatially differentiated irrigation management policies (Fig. 6c).

Fig. 6: Objective function trade-off and sensitivity analysis.
Fig. 6: Objective function trade-off and sensitivity analysis.
Full size image

a 2020 s objective function trade-off, b 2030 s objective function trade-off, c sensitivity analysis.

From the results of the three-objective trade-off analysis, the overall performance during the 2020s and 2030s was relatively similar. Most scenarios showed IWP ranging between 50 and 80%, indicating that water resources use efficiency remained relatively stable without a notable improvement or decline in the future. The proportion of EQ mainly fell between 50% and 90%; some scenarios in the 2030s showed a slight decrease, suggesting potential reductions in irrigation energy consumption through energy-saving optimization measures. Similarly, GHG exhibited minimal variation, with both periods showing roughly equivalent proportions, indicating no marked increase or decrease in emissions. Overall, the solution sets for the 2020s and 2030s show similarities in the trade-offs between IWP, GHG, and EQ. The key difference lies in the reduction of EQ in the 2030s. The direct driving factor for this change is the decrease in irrigation energy consumption per hectare, from 124 kWh ha−1 in the 2020s to 120 kWh ha−1 in the 2030s (a 3.2% decrease), indicating that future irrigation system optimization will tend toward energy-saving technological advancements. Meanwhile, IWP and GHG remain relatively stable, reflecting the enhanced synergy between water conservation and emission reduction goals.

The Sobol sensitivity analysis results further revealed the influence of different decision variables on IWP, EQ, and GHG. The sensitivity of IWP to various combinations of water sources (surface water and groundwater) and irrigation technologies (surface irrigation, sprinkler irrigation, and drip irrigation) across different months of the year showed a moderate level with relatively even distribution, indicating that IWP responded without a clear bias towards any specific decision variable. In contrast, the sensitivity of EQ and GHG was concentrated around the groundwater combination, especially EQ, which was highly sensitive to the combined effect of groundwater and surface irrigation technologies (with some months having ST values exceeding 0.06). This highlighted the relationship between groundwater extraction and the energy consumption of surface irrigation systems. Additionally, for this combination, the ST/S1 ratio approached 1, indicating that the sensitivity of the objectives was mainly driven by a single variable, with minimal interaction between the variables.

Discussion

This study constructs a high-resolution (1 km × 1 km) water-food-energy-carbon emission coupling model to systematically reveal the complex feedback mechanisms between irrigation water use, energy consumption, and carbon emissions in China agricultural system. The study found that improving irrigation water productivity (by 2%) significantly reduced the energy consumption per unit of production (by 37%) and carbon emission intensity (by 13%), and reshaped the resource-environment benefit pattern of agricultural production at the regional scale. The irrigation scenarios in the 2020s maintained a relatively high-water productivity, while exhibiting higher energy consumption and lower greenhouse gas emissions. By the 2030s, energy consumption had decreased, but greenhouse gas emissions had increased, indicating the need for future irrigation systems to not only enhance water use efficiency but also further optimize energy structures to reduce carbon emissions. In optimizing irrigation strategies in the future, it is essential to consider the synergistic effects of multiple variables, rather than relying solely on improvements in a single water source or irrigation technologies. Specifically, these variables include but are not limited to: water source types (such as surface water, groundwater, rainwater collection, etc.), irrigation technologies (such as drip irrigation, sprinkler irrigation, micro-irrigation, etc.), and the utilization of renewable energy (such as solar, wind, hydro, etc.). By integrating these variables organically and constructing a diversified irrigation model, it is possible to ensure the stability of agricultural production but also to effectively reduce carbon emissions and achieve low-carbon development goals. Compared with Radmehr et al.30, who developed an optimization model for the water-energy-food nexus in arid regions and achieved 23–26% energy savings under drought scenarios, our study demonstrates similar energy-saving potential but extends the analysis to finer spatial and temporal scales while explicitly incorporating carbon emissions. In addition, Wu et al.31 emphasized that optimizing crop planting structures should address both resource efficiency and actual development demands, such as mitigating the environmental costs of rice expansion and reducing dependence on soybean imports. Our results are consistent with this perspective, as the irrigation optimization framework not only improves resource efficiency but also provides a basis for sustainable cropping adjustments under climate and resource constraints. Together, these comparisons indicate that while previous studies highlighted the benefits of macro-level optimization, our high-resolution approach captures spatiotemporal heterogeneity and carbon feedbacks, thereby offering more precise guidance for region-specific strategies. Especially in the context of the dual pressures of water scarcity and agricultural carbon reduction, optimizing irrigation strategies can serve as a key approach for balancing food security and environmental sustainability.

This study integrates machine learning algorithms with multi-source data fusion techniques to construct a high-precision agricultural system prediction framework. In terms of the technical approach, the Random Forest algorithm incorporates multiple variables, such as climate (temperature, precipitation, etc.) and soil factors (SOC, pH, etc.), which helps to improve the accuracy of crop yield predictions. At the same time, based on a quantitative analysis of the irrigation water supply-demand balance, the study achieved dynamic predictions of pixel-level irrigation water demand within grids. In terms of methodological innovation, this study combined the Maxent ecological niche model with the PLUS crop planting structure model to achieve spatial optimization and precise layout of the crop planting structure. Through these methodological and technological innovations, the study constructed a comprehensive prediction framework that covered planting suitability, water resource demand, and crop yield, considerably enhancing the agricultural system’s ability to adapt to climate change and improving resource management effectiveness.

This study utilized CMIP6 climate scenario simulations to systematically evaluate the effects of future climate change on agricultural water resource utilization, energy consumption, and carbon emissions, while optimizing irrigation strategies to regulate agricultural water resources under various climate conditions. However, under the backdrop of climate change, agricultural systems still face considerable uncertainty, especially with the increasing frequency of extreme weather events and concurrent adjustments in crop planting structures, which together are projected to drive a marked increase in irrigation demand during the 2020s–2030s, primarily through higher evapotranspiration under warming, a shift toward more water-intensive cropping patterns, and uneven impacts of technological adoption. While water-saving irrigation partly mitigates these pressures, it is insufficient to offset the combined effects of climate change and cropping adjustments, leading to a persistent imbalance between water supply and demand and affecting crop growth stability. Extending the projection horizon to the 2050s would likely amplify these increases under intensified climate change. The dynamic optimization model developed in this study is specifically designed to adaptively allocate water according to evolving climatic conditions, ensuring that the optimized irrigation amounts remain valid and applicable across different time horizons. Relying solely on irrigation optimization may not be sufficient to address these challenges. Therefore, future agricultural management needs not only to precisely control water resources utilization but also to integrate crop growth characteristics, enhancing the overall adaptability of agricultural systems. For instance, by incorporating crop growth models, such as AquaCrop or DSSAT, it is possible to more accurately simulate the growth performance of different crop varieties under varying climate conditions, thereby optimizing crop distribution and water allocation strategies, improving the stability and risk resilience of agricultural production.

This study constructed a deterministic irrigation water supply-demand balance analysis framework based on a grid-based approach, revealing the temporal evolution patterns of crop planting structure through scenario simulations, and achieving optimal water resources allocation through spatial analysis. This method has clear advantages in solving optimization problems by establishing a grid-based dynamic allocation framework for synergistic optimization of the water-food-energy-carbon nexus, enabling precise “one-grid-one-policy” decision support through kilometer-scale spatial resolution. However, while the framework enhances spatial precision, it continues to treat planting structure and irrigation decisions as independent variables, which limits the ability to capture their dynamic coupling. Moreover, achieving water savings may involve trade-offs with higher energy consumption and carbon intensity, complicating integrated policy design under the water-energy-carbon nexus. The implementation of high-resolution optimization also faces barriers such as data availability, computational costs, and governance capacity, which may hinder its real-world application. Although the framework shows potential applicability to other water-scarce regions, its effectiveness depends on local crop patterns, institutional settings, and socio-economic adoption of irrigation technologies. In addition, uncertainties in crop models and the feasibility of advanced irrigation practices remain critical challenges. Future research should further develop dynamic interaction models calibrated with machine learning methods, while policy-oriented extensions should focus on adaptive governance mechanisms that balance efficiency, equity, and sustainability. The dynamic optimization model for irrigation water use developed in this study is not limited to the Chinese context but also holds potential for broader application. Where data availability permits, the framework could be transferred to other water-scarce regions, such as the Mediterranean, India, and sub-Saharan Africa. These regions face similar challenges of intensified climate change, diverse cropping structures, and the need to balance the water–food–energy–carbon nexus. Therefore, the model’s high-resolution approach to dynamically optimizing crop distribution and irrigation demand could provide valuable decision support in these global contexts.

Materials and methods

This study developed a refined 1 km × 1 km pixel-level irrigation water optimization model for China, which integrated multi-source data, including meteorological, soil, and agricultural management conditions. The model’s objective function aimed to maximize irrigation water productivity while minimizing energy consumption and carbon emissions, addressing the inherent connections within the spatial water-food-energy-carbon nexus. In this study, a hybrid approach using the Maxent-PLUS model was employed to simulate changes in crop planting structures under the influence of climate change. Additionally, an effective rainfall estimation method and crop coefficient method were used to accurately quantify agricultural water demand at the grid scale in the future. Crop yield predictions were made using machine learning algorithms to adapt to future environmental changes.

To ensure the robustness of model simulations, all datasets were harmonized to a uniform spatial and temporal resolution before integration, with missing-value interpolation, cross-validation against ground observations, and consistency checks applied to minimize potential errors. These preprocessing and quality control steps ensured data comparability and reliability across sources, thereby supporting the accuracy of the optimization results. The detailed descriptions of data sources and coverage are provided in the corresponding sections of the manuscript.

Prediction of future staple crop yields

Machine Learning (ML) was used to predict the yield data for China’s 1 km × 1 km grid. The input variable dataset included climate parameters (temperature, precipitation, wind speed, radiation, and CO2 concentration) from Peng et al.32, soil property data (pH), Soil organic carbon, Total nitrogen, Total phosphorus, Total potassium, Cation exchange capacity, Coarse fragments, Bulk density, and soil texture from Liu et al.33 and FAO, and IIASA34, and fertilizer data (nitrogen, phosphorus, and potassium fertilizers) from Adalibieke et al.35 and the China Statistical Yearbook. It is important to note that the meteorological parameters within the crop growing season in different agricultural regions were processed to meet the data structure requirements for ML. Precipitation and radiation were summed, while temperature, wind speed, and carbon dioxide concentration were averaged. Four machine learning models with different complexities (XGB, RF, SVR, and FNN) were trained to fit the input meteorological, fertilizer, and soil datasets. The crop yield spatial data was obtained from Iizumi and Sakai36, which included yield data for major staple crops from 2000 to 2016. The multi-year cumulative data from this dataset were used as information for training in the frontend. The dataset was divided into 80% for training and 20% for testing, and chi-square validation was applied to remove outliers from the dataset.

Irrigation demand estimation

The estimation of crop water requirements is a crucial step in irrigation management and water resource optimization. Irrigation water demands were estimated based on the difference between effective precipitation and crop evapotranspiration (Eq. (1)):

$${{\mathrm{ID}}}=\left\{\begin{array}{lll} & \hskip -1.2pc {{\mathrm{ET}}}_{c}-{P}_{{\mathrm{eff}}}\, &{{\mathrm{ET}}}_{c}\ge {P}_{{\mathrm{eff}}} \\ 0&&{{\mathrm{ET}}}_{c} < {P}_{{\mathrm{eff}}}\hfill\end{array}\right.$$
(1)

Where \({\mathrm{ID}}\) is the crop irrigation requirement (mm), \({{\mathrm{ET}}}_{c}\) is the crop evapotranspiration (mm), and \({P}_{{\mathrm{eff}}}\) is the effective precipitation (mm).

Estimation of crop evapotranspiration

The crop coefficient method is a commonly used approach for calculating \({{\mathrm{ET}}}_{c}\)37. This study assumed that under adequate water supply conditions, crop water requirements were influenced solely by atmospheric and crop-related factors.

$${{ET}}_{c}={{ET}}_{0}\times {K}_{c}$$
(2)

Where \({{\mathrm{ET}}}_{0}\) is the potential evapotranspiration (mm), and \({K}_{c}\) is the crop coefficient.

The estimation of crop potential evapotranspiration in the model is based on the Penman-Monteith formula as follows:

$${{\mathrm{ET}}}_{0}=\frac{0.408\Delta \left({R}_{n}-G\right)+\gamma \left[900/\left(T+273\right)\right]{u}_{2}\left({e}_{s}-{e}_{a}\right)}{\Delta +\gamma \left(1+0.34{u}_{2}\right)}$$
(3)

where \(\Delta\) is the slope of the saturated vapor pressure curve (kPa °C−1); \({R}_{n}\) is the net radiation from the crop surface (MJ); \(G\) is the soil heat flux (MJ (m2 day) −1); \(\gamma\) is the humidity calculation constant (kPa °C-1); \(T\) is the temperature at heights of 1.5–2.5 m (°C); \({u}_{2}\) is the wind speed at 2 m (m s−1); \({e}_{s}\) is the saturated vapor pressure (kPa); and \({e}_{a}\) is the actual vapor pressure (kPa). Since this study focused on optimizing irrigation water resources allocation with a monthly time step, the potential evapotranspiration data used in this study were derived from a monthly potential evapotranspiration dataset32. In practice, the \({K}_{c}\) is influenced by soil conditions, climate, crop growth conditions, and management practices. If the field conditions differ from the standard conditions (management and environmental conditions), \({{\mathrm{ET}}}_{c}\) needs to be adjusted using \({K}_{c}\). For monthly irrigation planning studies, the FAO-recommended single crop coefficient correction method is applied to adjust \({K}_{c}\) (Eq.(4)). The detailed correction process can be found in Supplementary Text 1.

$${K}_{{cm}}={K}_{c}+0.04\left({U}_{2}-2\right)-0.004({{\mathrm{RH}}}_{\min }-45){\left(\frac{h}{3}\right)}^{0.3}$$
(4)

Where \({K}_{{cm}}\) represents the adjusted crop coefficient under non-standard climatic conditions, \({U}_{2}\) is the wind speed (m s−1), \({{\mathrm{RH}}}_{\min }\) is the average minimum daily relative humidity (%), and h is the average crop height (m).

Effective rainfall estimation

The equation for calculating the monthly effective rainfall was as follows38 (Muratoglu et al.):

$${P}_{{\mathrm{eff}}}=\left\{\begin{array}{ll}\frac{P\left(125-0.2\right)}{125}\hfill&P\le 250{{\mathrm{mm}}}\\ 125+0.1P\hfill&P > 250{{\mathrm{mm}}}\hfill\end{array}\right.$$
(5)

Where \({P}_{{\mathrm{eff}}}\) is the monthly effective rainfall (mm), and \(P\) is the monthly rainfall value (mm) from Peng et al.32.

Simulation of future crop planting structure

The Maxent-PLUS model, which combines the maximum entropy principle with cellular automata models, was used to model the spatial expansion of major food crops in the future. The Maxent model is a machine learning-based technique that simulates the potential distribution of species and their environmental niches39. It applies the maximum entropy principle, which selects the model with the highest entropy under the constraints provided by the given data, to predict the distribution probabilities of species under various environmental conditions. The output of the Maxent model includes maps of suitable areas for species distribution and analyses of the contribution of different environmental variables. The PLUS model is a land-use change model designed to simulate the changes in land use and land cover. Specifically, the two models serve complementary roles: Maxent focuses on identifying the potential spatial suitability of crops under different environmental drivers, while PLUS emphasizes the dynamic simulation of land-use transitions by considering spatial constraints and neighborhood effects. By integrating the probabilistic suitability maps from Maxent with the spatial allocation mechanisms of PLUS, the coupled Maxent-PLUS framework not only captures the environmental niches of crops but also represents realistic land-use change processes. This integration enhances the accuracy and adaptability of predicting future crop distribution, thereby improving the reliability of irrigation water demand estimation and the robustness of water-food-energy-carbon optimization strategies40. We focused on the main staple crops of China as the subject of the study, with the underlying dataset for crop simulation sourced from Luo et al.41. The crop combinations were classified into seven categories: rice, maize, wheat, rice-maize rotation, rice-wheat rotation, maize-wheat rotation, and rice-wheat-maize rotation. This classification was designed to ensure that each spatial grid contained only one type of cropping system. We simulated the spatial changes of staple crops at a 1-km resolution, which was commonly used in large-scale analyses and was matched with other spatial data. After simulation, the future crop planting scenarios were restored to spatial distributions of rice, maize, and wheat, with the results finally aggregated to provincial-level regions. The simulation process of the Maxent-PLUS model under different scenarios can be found in Supplementary Figs. S3, S4, and S5, with the environmental variables listed in Supplementary Table S2.

Model framework

The optimization model integrated four core objective functions: maximizing irrigation water productivity, minimizing energy consumption, minimizing carbon emissions, and ensuring fairness in resource allocation through the Theil fairness minimization objective (A quantitative indicator measuring the degree of irrigation water allocation inequality within grid cells, where lower values indicate fairer water resource distribution across different grid cells). Additionally, the model took into account the limited availability of regional water resources and the basic requirements for crop growth, setting two key constraint conditions at the grid scale: (1) the availability of water resources, and (2) the minimum irrigation requirements for crops, ensuring the model’s adaptability and rationality under different regional and environmental conditions.

(1) Irrigation water productivity (IWP)

IWP is an important indicator for measuring agricultural irrigation efficiency, reflecting the crop yield produced per unit of irrigation water, thereby improving the utilization efficiency of water resources. It was calculated using the following Eq. (6):

$${\mathrm{IWP}}=\frac{{\sum }_{c=1}^{C}\left({Y}_{c}\times {A}_{c}\right)}{{\sum }_{i=1}^{I}{\sum }_{c=1}^{C}{\sum }_{m=1}^{M}\left[\left(I{W}_{{\mathrm{icm}}}^{{\mathrm{sur}}}+I{W}_{{\mathrm{icm}}}^{{\mathrm{gro}}}\right)\times {A}_{c}\right]}$$
(6)

Where i represents different irrigation methods, including surface irrigation, sprinkler irrigation, and drip irrigation; c represents different crop types; m represents the months; \({Y}_{c}\) is the crop yield (kg ha−1); \({A}_{c}\) is the crop sowing area (ha); for the kilometer-scale grid optimization, 1 grid = 100 ha; \({{\mathrm{IW}}}_{{\mathrm{icm}}}^{{\mathrm{sur}}}\) is the surface water irrigation (m3 ha−1); \({{\mathrm{IW}}}_{{\mathrm{icm}}}^{{\mathrm{sur}}}\) is the groundwater irrigation (m3 ha−1).

(2) Energy consumption

The energy consumption of the irrigation process is influenced by the irrigation system, including surface irrigation, sprinkler irrigation, and drip irrigation, without considering water head losses in transport and field irrigation. The irrigation energy demand was calculated using the Eq. (7):

$${\mathrm{EQ}}={{\mathrm{EQ}}}^{{\mathrm{sur}}}+{{\mathrm{EQ}}}^{{\mathrm{gro}}}$$
(7)

Where \({\mathrm{EQ}}\) is the energy demand (kWh); \({{\mathrm{EQ}}}^{{\mathrm{sur}}}\) is the energy demand for surface water (kWh); \({{\mathrm{EQ}}}^{{\mathrm{sur}}}\) is the energy demand for groundwater (kWh). Specific explanations of the formulas are provided in Supplementary Text 2.

(3) GHG emissions

The greenhouse gas (GHG) emission target is another key objective function in this optimization model, aiming to minimize carbon emissions during agricultural production. This objective function covers multiple sources of greenhouse gas emissions, including electricity consumption (\({\mbox{GH}}{{\mbox{G}}}_{{\mbox{elec}}}\)), fertilizer use (\({\mbox{GH}}{{\mbox{G}}}_{{\mathrm{fer}}}\)), field emissions (\({\mbox{GH}}{{\mbox{G}}}_{{\mathrm{land}}}\)), agricultural input use (\({\mbox{GH}}{{\mbox{G}}}_{{\mathrm{input}}}\)), and the degassing process of groundwater (\({\mbox{GH}}{{\mbox{G}}}_{{\mbox{degassing}}}\))42. The calculation of this objective function was performed using the following Eq. (8), Specific explanations of the formulas are provided in Supplementary Text 2.

$${GHG}={{GHG}}_{{\mathrm{elec}}}+{{GHG}}_{{\mathrm{fer}}}+{{GHG}}_{{\mathrm{land}}}+{{GHG}}_{{\mathrm{input}}}+{{GHG}}_{{\mathrm{degassing}}}$$
(8)

(4) Water distribution fairness

The Theil index was used to measure the fairness of water distribution between different regions or water users:

$${{\mbox{Theil}}} = \mathop{\sum }\limits_{i=1}^{I}\mathop{\sum }\limits_{c=1}^{C}\mathop{\sum }\limits_{m=1}^{M}\left[\left(\frac{{{IW}}_{{icm}}^{{sur}}+{{\mathrm{IW}}}_{{icm}}^{{gro}}}{{\sum }_{i=1}^{I}{\sum }_{m=1}^{M}\left({{IW}}_{{icm}}^{{sur}}+{{IW}}_{{icm}}^{{gro}}\right)}\right) \right.\\ \left. \times {ln}\left(\frac{\frac{\left({{IW}}_{{icm}}^{{sur}}+{{IW}}_{{icm}}^{{gro}}\right)}{{\sum }_{i=1}^{I}{\sum }_{m=1}^{M}\left({{IW}}_{{icm}}^{{sur}}+{{IW}}_{{icm}}^{{gro}}\right)}}{\frac{{A}_{c}}{{\sum }_{c=1}^{C}\left({A}_{c}\right)}}\right)\right]$$
(9)

Constraints

(1) Water resource availability constraint

The model took into account the limited availability of regional water resources when allocating water resources.

$$\mathop{\sum }\limits_{i=1}^{I}\mathop{\sum }\limits_{c=1}^{C}\mathop{\sum }\limits_{m=1}^{M}\left({{IW}}_{{icm}}^{{sur}}\times {A}_{c}\right)\le {{TW}}^{{sur}}$$
(10)
$$\mathop{\sum }\limits_{i=1}^{I}\mathop{\sum }\limits_{c=1}^{C}\mathop{\sum }\limits_{m=1}^{M}\left({{IW}}_{{icm}}^{{gro}}\times {A}_{c}\right)\le {{TW}}^{{gro}}$$
(11)

Where \({{\mathrm{TW}}}^{{\mathrm{sur}}}\) is the available surface water (m3), and \({{\mathrm{TW}}}^{{\mathrm{gro}}}\) is the available groundwater (m3).

(2) The minimum irrigation requirement

For the irrigation water volume at the kilometer grid scale, it must be greater than the minimum irrigation water requirement:

$$\mathop{\sum }\limits_{i=1}^{I}\left({{IW}}_{{icm}}^{{sur}}+{{IW}}_{{icm}}^{{gro}}\right)\times {\alpha }_{i}\ge {{ID}}_{{cm}}$$
(12)

Where \({\alpha }_{i}\) is the water use efficiency of different irrigation methods, with data sourced from the China Water Resources Bulletin, and \({{\mathrm{ID}}}_{{\mathrm{cm}}}\) is the irrigation water demand (m3 ha−1).

Co-optimization of multiple management practices

We performed a multi-objective optimization algorithm (Nondominated Sorting Genetic Algorithm II, NSGA-II) to search for the global optimal management practices for the target. Unlike traditional NSGA models, our data structure design is more refined, integrating the foundational data of the optimization model into each spatial grid unit. This allowed each spatial grid to become an independent optimization unit, enhancing the model’s ability to capture spatial heterogeneity. Our algorithm features a dynamic adjustment mechanism for the fitness function, which more accurately analyzed the spatial heterogeneous contributions of different management practices. At the same time, it provided scientifically sound and rational water resources allocation plans for the target area.

NSGA-II divides the solution set into different Pareto levels through fast non-dominated sorting, calculated as follows:

$${rank}\left(p\right)=k,\quad{if\;p}\;{\in }\;{{thek}}^{{th}}\;{Paretofront}$$
(13)

Simultaneously, maintain solution diversity using crowding distance:

$${d}_{i}=\mathop{\sum }\limits_{m=1}^{M}\frac{{f}_{\!\!i+1}^{m}-{f}_{\!\!i-1}^{m}}{{f}_{\!\!\max }^{m}-{f}_{\!\!\min }^{m}}$$
(14)

Where \({f}^{m}\) is the value of the \(m\) th objective function. The algorithm progressively approximates the Pareto optimal frontier through selection, crossover, and mutation. In this study, NSGA-II is employed to identify a compromise solution among “maximizing irrigation water productivity, minimizing energy consumption, and minimizing carbon emissions”, thereby generating a set of optimal irrigation schemes.

Sensitivity analysis

This study used Sobol’s method based on variance decomposition to evaluate the contribution of model decision variables to the objective function of the model, quantifying the influence of each variable on the uncertainty of the model output43. Specifically, for the objective function \(Y=f\left({X}_{1},{X}_{2},....,{X}_{k}\right)\), its total variance \({\mathrm{Var}}\left(Y\right)\) was decomposed into the contributions of each decision variable \({X}_{i}\). The first-order sensitivity index \({S}_{1}\) was used to measure the independent effect of a single decision variable \({X}_{i}\) on the variance of the objective function. The calculation formula was:

$${S}_{1}=\frac{{{\mathrm{Var}}}_{{X}_{i}}\left({E}_{X{ \sim }_{i}}\left(Y\left|{X}_{i}\right.\right)\right)}{{\mathrm{Var}}\left(Y\right)}$$
(15)

where \({E}_{{X}_{i}}\left(Y\left|{X}_{i}\right.\right)\) represents the expected value of \(Y\) under fixed conditions of \({X}_{i}\). \({{\mathrm{Var}}}_{{X}_{i}}\) represents the variance calculation for \({X}_{i}\). The first-order index \({S}_{1}\) reflects the individual contribution of \({X}_{i}\) to the uncertainty of the objective function.

To further evaluate the interaction between decision variables, this study also calculated the global sensitivity index \({S}_{T}\), which was given by the formula:

$${S}_{T}=1-\frac{{{Var}}_{X{ \sim }_{i}}\left({E}_{{X}_{i}}\left(Y\left|{{X}_{ \sim }}_{i}\right.\right)\right)}{{Var}\left(Y\right)}$$
(16)

The global sensitivity index \({S}_{T}\) took into account the total contribution of \({X}_{i}\) and its interactions with other variables to the uncertainty in the objective function. By comparing the first-order and global sensitivity indices, it was possible to identify the decision variables and their interactions that had a notable impact on the objective function.

Scenario design

This study selected four time nodes at 10-year intervals from the 2000s to the 2030s, covering the period from the past to the future. The study period was selected primarily for two considerations: (1) Meteorological and land use data are obtainable and demonstrate good continuity within this timeframe; (2) Most climate scenario projections hold high credibility up to 2030s. This design helped us review the effectiveness of past policies while predicting and assessing the impacts of future climate change, energy policies, and land use changes. The aim was to analyze the effects of energy policy implementation and climate change on land use at different stages. By analyzing past trends (2000s and 2010s) and near-current years (2020s), a better understanding of the historical changes in climate policies and land use patterns was achieved. Additionally, using CMIP6 scenario models, future changes (2030s) were projected. This timeline arrangement provided a scientific basis for a comprehensive understanding of the dynamic interactions between the past, present, and future. First, CMIP6 provided multi-scenario simulation data, including SSP1−2.6 (low emission scenario), SSP2-4.5 (medium emission scenario), and SSP5-8.5 (high emission scenario), to obtain corresponding climate data for different time nodes, with data sourced from Yuan et al.44. Second, the spatial simulation method for future crop planting patterns was used to simulate the changes in staple crop patterns under different emission scenarios in the 2020s and 2030s. Third, data on the usage proportions of surface irrigation, drip irrigation, and micro-irrigation between 2009 and 2022 were obtained from the China Water Resources Statistical Yearbook, while the usage proportions of diesel, wind, hydroelectric, nuclear, solar, and coal were derived from the China Energy Statistical Yearbook. Linear regression prediction was used to explain the irrigation technologies proportions and energy proportions in the 2020s and 2030s. Lastly, eight optimization scenarios were established (as shown in Supplementary Table S4): (1) the status quo scenario, based on actual parameters, includes two scenarios: 2000s and 2010s; and (2) future scenarios, based on climate-crop structure-energy- irrigation technologies variations, include six scenarios: RCP126-2020s, RCP126-2030s; RCP245-2020s; RCP245-2030s; RCP8.5-2020s; RCP8.5-2030s.