Spatio-Temporal Optimization of Rice Irrigation at Raster Scale: Synergies Between Water Productivity and Methane Emission Reduction

Abstract

This study addresses the challenges of coordinating spatio-temporal water allocation to optimize water productivity and reduce carbon emissions in water resource management, particularly the lack of high-resolution, integrated optimization frameworks capable of simultaneously tackling water scarcity and greenhouse gas (GHG) emissions. We propose a modeling approach for large-scale regional rice irrigation that explicitly represents the physical-process-based relationships among irrigation water, yield, and methane (CH₄) emissions. Using GIS, a grid-based simulation domain was constructed at a 500 m × 500 m resolution, and the GIS-DSSAT and GIS-DNDC models were employed to simulate yield and CH₄ emissions under varying irrigation amounts. The Random Forest algorithm—selected for its ability to capture complex nonlinear interactions—was used to establish the response surfaces linking irrigation water, yield, and CH₄ emissions. A spatio-temporal irrigation optimization model was then developed to simultaneously reduce CH₄ emissions and enhance water productivity. This methodology was applied to the Sanjiang Plain in Heilongjiang Province, where the NSGA-II algorithm was used to derive optimal irrigation schemes for rice cultivation across 408,264 grid cells. The results revealed quadratic nonlinear relationships between irrigation water amount, yield, and CH₄ emissions. Compared to the conventional irrigation practice in the region, which typically involves 15–20 flood irrigation events per season, the optimized irrigation schedule comprised 7–14 events—with 12 events accounting for 42% of the cases—and an irrigation duration ranging from day 137 to 256. This led to a 10.3% reduction in total irrigation volume, a 9.6% decrease in CH₄ emissions per unit yield, and a 21.8% increase in water productivity. This study provides valuable decision support for optimizing regional water allocation and developing rice cultivation strategies that improve productivity while reducing emissions.

1. Introduction

Rice serves as a staple food for more than half of the global population. As the world’s largest rice producer and consumer, China accounts for 18.5% of the global rice harvested area [1,2]. However, ensuring food security in rice production heavily depends on substantial irrigation water resources. Rice irrigation consumes over 65% of China’s total agricultural water use, with water consumption 2–3 times higher than that of other major staple crops such as wheat and maize [3,4]. Consequently, optimizing water use is critical for achieving efficient rice production. Paddy rice fields contribute 10–13% of global anthropogenic methane (CH₄) emissions [5], and those in China account for 21.9% of the global total [6]. Moreover, CH₄ emissions constitute up to 94% of the total greenhouse gas emissions from rice paddies [7]. Given the urgency of mitigating global warming, optimizing irrigation regimes to enhance water productivity and reduce CH₄ emissions is essential for ensuring food security, combating climate change, and meeting the carbon neutrality target by 2060 [8]. However, operational frameworks capable of high-resolution, spatio-temporal optimization to address this water–carbon nexus at a regional scale are currently lacking.

Recent years have seen progress in optimizing rice irrigation. Studies have explored various techniques, from meta-analyses of water-saving irrigation [9] to optimization models combining stochastic methods and algorithms [10]. The integration of process-based models like DSSAT (Decision Support System for Agrotechnology Transfer) for water management [11] and DNDC (Denitrification—Decomposition) for emission estimation [12,13] marks a significant advancement. However, as summarized in

Table 1. Simulation model input parameters.

Datatypes	DSSAT Model	DNDC Model	Unit
Weather data	Site longitude (latitude), elevation, Day by day maximum (low) temperature, precipitation, solar radiation	Daily maximum (low) temperature, daily precipitation	°C, mm, MJ/m2
Soil data	Soil type, soil stratification, wilting point water content, field holding capacity, bulk weight, organic carbon content, soil pH, albedo, drainage rate, number of runoff curves, etc.	Soil texture, bulk weight, soil pH, organic carbon content, clay content	%, cm3/cm3, g/cm3
Crop data	Parameters of seedling growth characteristics, photoperiod sensitivity parameters, parameters of irrigating stage characteristics, maximum number of grains per plant, and parameters of potential irrigating rate	Crop type, maximum crop yield, optimum temperature for growth	kg/ha, °C
Field management data	Crop type, amount of straw returned to the field, date and method of sowing, sowing density, row spacing, date of irrigation, amount of irrigation, amount of fertiliser applied, application measures, depth of application, type of fertiliser	Cultivation method, date of planting and harvesting, date (amount) of fertiliser application (chemical/organic)	date, kg/ha, cm

, existing research often focuses on experimental plots or coarse regional scales, lacking the resolution to account for fine-scale spatial heterogeneity and optimal irrigation timing. This limits their direct applicability for precise, region-wide management.

Rice growth and CH₄ emissions are highly sensitive to spatially variable factors such as soil texture, organic matter, and microclimate [14]. This inherent spatial heterogeneity dictates that irrigation strategies must be site-specific to be effective. Consequently, accurately quantifying location-specific responses of yield and emissions to irrigation is fundamental. While distributed crop modeling on GIS (Geographic Information System) platforms provides a viable path forward [15,16], applications have been hampered by coarse resolutions (e.g., 10 km × 10 km) that cannot resolve critical sub-grid variability, a limitation detailed in Supplementary Material Table S1. This gap underscores the need for a high-resolution, physically based framework that can both simulate and optimize irrigation scheduling across vast areas while accounting for fine-scale spatial differences.

This study develops a physically based, high-resolution grid-driven irrigation optimization framework to improve water use efficiency and mitigate methane emissions at the regional scale. To capture the spatial heterogeneity of agricultural management and soil properties while ensuring computational tractability, a grid resolution of 500 m was selected. The specific objectives are to: (1) Overcome the coarse spatial resolution of traditional regional irrigation management by coupling the GIS-DSSAT and GIS-DNDC models on a 500 m × 500 m grid to simulate both rice growth and CH₄ emission processes; (2) Integrate crop models with machine learning techniques by employing the Random Forest algorithm to quantify nonlinear multi-factor interactions and establish the response relationships among irrigation water, yield, and CH₄ emissions; (3) Construct a spatio-temporally distributed multi-objective optimization model that simultaneously optimizes water productivity and CH₄ emission intensity using the NSGA-II algorithm, thereby achieving irrigation optimization for rice production across 408,264 response units at the regional scale. These contributions facilitate the optimization of spatio-temporal irrigation systems for regional rice cultivation within ecologically and economically sustainable agricultural practices.

The paper is structured as follows. Section 2 details the study area, data, and the coupled modeling-optimization methodology. Section 3 presents the optimization results, including changes in yield, CH₄ emissions, irrigation water use, and WP. Section 4 provides a combined discussion and conclusion, interpreting the findings, examining mechanisms and limitations, and outlining implications for sustainable rice production.

2. Materials and Methods

2.1. Overview of the Study Region

The Sanjiang Plain, located in the northeastern part of Heilongjiang Province, China (43°50′–48°24′ N, 129°11′–134°46′ E), is the province’s primary grain-producing region. It is a low-lying alluvial plain formed by the Heilong, Songhua, and Ussuri rivers, covering an area of approximately 108,900 km². The region experiences a temperate, humid, and semi-humid continental monsoon climate, with an average annual temperature of 1–4 °C. Annual sunshine duration ranges from 2400 to 2500 h, and the frost-free period lasts 120–140 days. The accumulated temperature above 10 °C is 2300–2500 °C. During the growing season (May–September), the daily average temperature ranges from 17 °C to 20 °C. The area is characterized by high latitude, rich soil organic matter, and an annual rainfall of approximately 384–886 mm, with about 400–600 mm occurring during the growing season. Elevation ranges from 34 to 1100 m above sea level. Precipitation is concentrated in the summer and autumn, and the region’s summer rain–heat synchronization, combined with its high soil fertility (soil organic matter content typically ranging from 3% to 5%), creates favorable conditions for rice cultivation. By 2022, the rice cultivation area in the Sanjiang Plain reached approximately 1.1 million hectares, accounting for 29% of the region’s arable land. The proposed model was applied to 21 counties (or cities) that constitute the core agricultural area of the Sanjiang Plain, including those within Jiamusi, Hegang, Shuangyashan, Qitaihe, and Jixi cities in Heilongjiang Province, as well as Yilan County in Harbin City. The study area is depicted in Figure 1.

2.2. Data and Methods

The data used in this study include model input parameters: meteorological, soil, crop, and field management data (Table 1). Meteorological data were obtained from the China Meteorological Data Service Center (http://data.cma.cn (accessed on 23 November 2024)) for the year 2022, covering 22 stations across the study area. The average annual temperature is 3.6 °C, and the average annual rainfall is approximately 580 mm. These data were used to characterize the regional agroclimatic conditions, including mean temperatures and precipitation amounts as summarized in Section 2.1. Station observations were interpolated using Thiessen polygons in GIS to create continuous spatial datasets. The processed meteorological data were then converted into formats compatible with the DSSAT and DNDC models, generating a gridded regional climate dataset for subsequent simulations. Soil data were extracted from the Harmonized Soil Database version 2.0 (HWSD v2.0), with single-factor extraction performed using GIS. Data screening was carried out in R (version 4.4.1), and the filtered data were rasterized in GIS to produce the multifactor raster data required for the model. The soil data include soil profile stratification, field water holding capacity, soil water content, bulk density, pH, and organic matter content. Crop and field management data were primarily sourced from relevant regional statistical yearbooks, national agricultural yield information, and related literature. The irrigation decisions in this study are based on a critical water consumption threshold parameter $\delta$.

2.3. Distributed Modelling of Crop Growth Processes

In this study, the GIS-DSSAT model was employed to simulate the spatial distribution of rice growth across the study region. The DSSAT (Decision Support System for Agrotechnology Transfer) model is a process-based crop growth simulation tool that quantitatively describes crop development, yield formation, and their interactions with climatic factors, soil properties, crop varieties, and cultivation practices. As one of the most widely used crop growth simulation systems, DSSAT provides quantitative tools for predicting crop growth and development, estimating yield, optimizing management practices, assessing environmental impacts, and analyzing climate change scenarios under diverse environmental conditions. The DSSAT system includes the CERES and CROPGRO model series and standardizes the input and output variables for these models [17,18,19,20].

The simulation employed a predominant japonica rice cultivar adapted to the cold region of the Sanjiang Plain. The genetic coefficients were derived from the standard “CERES-Rice” parameter library within the DSSAT model and were further calibrated and validated using local cultivar registration data and published regional calibration studies [17]. To establish robust irrigation–yield–emission response relationships, a series of 99 irrigation scenarios were simulated for each grid, with irrigation amounts ranging from 0 mm and increasing in 5-mm increments. This gradient was designed to comprehensively capture the full physiological response spectrum of rice, from conditions of water stress to waterlogging.

In this study, rice growth was simulated to assess yield under different irrigation water levels. The Cropping System Model (CSM), based on a modular simulation framework, was employed to simulate rice growth and development. The model integrates dynamic processes of soil moisture, nitrogen, and carbon cycling, with the CERES-RICE module specifically simulating rice growth. The simulation results were extrapolated to the regional scale using GIS (ArcGIS 10.8). A 500 m × 500 m grid resolution was applied across the Sanjiang Plain to compile spatial data—including meteorological and soil properties for each grid cell—thereby enabling simulation of environmental impacts on rice growth. Spatial data from meteorological stations were processed using R software to generate Rice. RIX files, which were subsequently integrated with the simulation models to simulate crop growth and water demand for each grid cell. This approach facilitated the estimation of rice yields under diverse meteorological conditions and irrigation scenarios, providing essential data for analyzing the yield response to irrigation water, as illustrated in Figure 2.

2.4. Distributed Modelling of Crop Methane Emissions

In this study, the GIS-DNDC model was employed to simulate the spatial distribution of CH₄ emissions during regional rice growth. The DNDC (Denitrification–Decomposition) model is a process-based soil environment simulation tool designed to assess the impacts of climate, soil, vegetation, and human activities on soil biogeochemistry. The model comprises six sub-modules: climate and soil, crop growth, soil organic matter decomposition, nitrification, denitrification, and fermentation. These sub-modules operate on daily time steps and exchange information to simulate the complex interactions among environmental conditions, plant growth, and soil chemical dynamics. Driven by ecological factors, the DNDC model dynamically simulates changes in soil conditions and evaluates gas emissions (e.g., CH₄) resulting from microbial activity within the plant–soil system. The model structure is divided into two main components. The first component simulates soil environmental conditions—including temperature, moisture, pH, redox potential, and the concentration gradients of relevant chemicals—which are influenced by ecological drivers such as climate and vegetation. The second component consists of three sub-models (nitrification, denitrification, and fermentation) that focus on how the soil environment affects microbial activity and quantify greenhouse gas emissions, including CH₄.

For the simulation, the key biogeochemical parameters governing methane production, oxidation, and transport in the DNDC model primarily adopted the model’s default settings recommended for paddy fields in East Asia [16]. Key parameters include the carbon allocation coefficients of rice plants, the decomposition rate constants of soil organic matter, and the optimal pH and redox potential (Eh) ranges for methanogenesis and methanotrophy (Supplementary Table S2).

In this study, we investigated CH₄ emissions under different irrigation conditions by simulating rice growth and its interaction with environmental factors. First, regional data were extrapolated to the Sanjiang Plain using GIS technology, which involved discretizing the region into a 500 m × 500 m grid to obtain spatial data—such as meteorological and soil properties—for each grid cell. These data served as ecological drivers in the model. The response of rice growth to soil chemical changes under varying environmental conditions was then simulated. Particular attention was paid to the processes of CH₄ production, oxidation, and transport in rice under different irrigation scenarios, which ultimately governed CH₄ emissions across meteorological stations and irrigation regimes. The resulting dataset provided key insights for establishing the relationship between CH₄ emissions and irrigation water, as illustrated in Figure 2.

2.5. Model Calibration and Validation

The model was calibrated and validated through two separate procedures: parameter adjustment and model evaluation. Station-specific data were first input into the model, and parameters were calibrated using a combination of the trial-and-error method and the GLUE (Generalized Likelihood Uncertainty Estimation) algorithm embedded within the model to achieve parameter localization. The calibrated genetic parameters were then applied to the validation datasets from each station to generate simulation results [16]. The validation results are illustrated in the Supplementary Materials (Figures S1 and S2). The model has been calibrated and validated using field data collected from over 30 stations distributed across the study region. Specifically, the DSSAT model was validated using actual values from the Heilongjiang Statistical Yearbook, while the DNDC model was validated using data from 37 irrigation districts. The validation results demonstrated good model performance. For yield simulation using DSSAT, the normalized root mean square error (nRMSE) was 5.08%, with a coefficient of determination (R²) of 0.87. For CH₄ emission simulation using DNDC, the nRMSE was 12.64%, with an R² of 0.83. These metrics indicate that both models satisfactorily captured the spatiotemporal dynamics of rice yield and methane emissions in the study region.

The integrated modeling framework (GIS-DSSAT, GIS-DNDC) was further validated at the regional scale by comparing aggregated outputs with independent estimates. The simulated regional average yield and total CH₄ emissions were within 10% of the values reported in the Heilongjiang Statistical Yearbook and published literature, indicating satisfactory performance of the coupled system.

This correlation is indicated by the Normalised Root Mean Square Error (nRMSE) and the Coefficient of Determination (R²). The equations are as follows:

$$RMSE=\sqrt{{\displaystyle \sum _{i=1}^n{\left(S_i-O_i\right)}^2}}$$

(1)

$$nRMSE={\displaystyle \frac{RMSE}{\overline{O}}}\times 100\%$$

(2)

$$R^2={\left\{{\displaystyle \frac{{\displaystyle \sum \left(O_i-\overline{O}\right)\left(S_i-\overline{S}\right)}}{\sqrt{{\displaystyle \sum {\left(O_i-\overline{O}\right)}^2{\displaystyle \sum {\left(S_i-\overline{S}\right)}^2}}}}}\right\}}^2$$

(3)

In the above equation, $O_i$ represents the i-th observation and $S_i$ represents the simulated value. $\overline{O}$ represents the mean value of the observation, while $\overline{S}$ represents the mean value of the simulated value. n represents the number of samples. When the normalized root mean square error (nRMSE) is less than 10%, it signifies that the simulation has a highly favorable effect. If the nRMSE falls between 10% and 20%, it indicates a decent simulation effect. However, if the nRMSE exceeds 30%, it suggests an inadequate simulation effect. The variable R² has a range of values from 0 to 1, with a higher value indicating a greater correlation between simulated values and measured values.

2.6. Irrigation Water Volume Fitted to Yield and Methane Emission Responses

To address the excessive computational cost arising from the direct coupling of the mechanistic model and optimization algorithms across more than 400,000 grid cells, this study employs a two-stage surrogate modeling framework. First, an irrigation scenario sample database is generated based on the DSSAT and DNDC models. Subsequently, a Random Forest (RF) algorithm is utilized to construct an efficient surrogate model that fits the complex response relationships between irrigation water quantity, yield, and methane emissions, thereby efficiently supporting subsequent large-scale optimization. RF is an ensemble learning method designed to enhance model accuracy and stability by constructing multiple decision trees and aggregating their predictions. During the construction of each decision tree, RF applies bootstrap sampling to input the simulation results of yield and CH₄ emissions under various irrigation conditions, as simulated by the DSSAT and DNDC models. It randomly selects samples from the training data, ensuring model diversity and mitigating the overfitting issues that may arise from a single decision tree. Rather than considering all features for each decision tree node, RF was selected to construct an efficient surrogate model. RF was preferred over SVM, ANN, and XGBoost because it requires no kernel/architecture tuning, provides intrinsic feature importance, and achieves competitive accuracy (R² > 0.75) with lower training cost. RF randomly selects a subset from the full set of features to split, thus enhancing the independence of each tree [21,22]. The final response relationship between irrigation water quantity, yield, and CH₄ emission is derived through majority voting (for classification problems) or averaging (for regression problems), as shown in Figure 2. These results provide crucial data to support the optimization of irrigation regimes.

The RF model was configured with 500 decision trees, and its hyperparameters, including tree depth and split criteria, were tuned through cross-validation. It was trained on a dataset comprising all 99 simulated irrigation scenarios to establish the nonlinear response surfaces. The model’s performance was rigorously evaluated, with key metrics such as R² reported in Figure 2. RF was selected over alternative methods like SVM and ANN due to its robustness against overfitting and its ability to quantify variable importance, which aids in interpreting the driving factors behind yield and emission responses.

2.7. Optimal Decision-Making for Spatio-Temporal Irrigation Water Quantity in Rice

This paper proposes a multi-objective regional spatio-temporal optimization model for irrigation water allocation. The model is based on spatially distributed simulations of rice yield and CH₄ emissions, generated using the GIS-DSSAT and GIS-DNDC models. It establishes the response relationships between irrigation water amount, yield, and CH₄ emissions, which are subsequently used to characterize water productivity and emission intensity. The model optimizes water resource allocation to maximize water productivity while promoting low-carbon rice cultivation. By analyzing the dynamic effects of irrigation water volume on yield and CH₄ emissions throughout the rice growth cycle, the model efficiently allocates limited water resources across grids. The objective function is formulated as follows:

2.7.1. Optimizing the Modelling Objective Function

(1) Water productivity:

Water productivity reflects the crop yield obtained per unit of water consumed. The objective function can be expressed as:

$$MaximizeWP={\displaystyle \sum _{i=1}^I\frac{Y_i}{W_i}}$$

(4)

$${\mathrm{Y}}_i=a_1{W_i}^2+b_1{W}_i+c_1$$

(5)

where WP represents water productivity [7,10], defined as the crop yield per unit of water consumed kg/m³). Y_i is the yield per unit area of grid i, the theoretical relationship between yield (Y_i) and irrigation water amount (W_i) is defined by a quadratic function (Equation (5)). The specific quadratic response relationship is generated through a dataset simulated by DSSAT, and the relationship is fitted using a Random Forest model. W_i is the amount of water applied per unit area of grid i (m³/ha), and a₁, b₁, and c₁ are the fitting coefficients.

(2) Methane emission:

The objective of methane emission is to minimize CH₄ emissions during rice production, thereby reducing environmental impacts and promoting sustainable agriculture. This objective function can be expressed as:

$$\min G={\displaystyle \sum _{i=1}^I{CH}_{4(i)}}$$

(6)

$$E_i=f_i\left(W_i\right)$$

(7)

CH_4(i) is obtained by fitting the relationship between CH₄ emissions and irrigation water (kg C/ha) using random forests based on DNDC simulations and E_i is the total CH₄ emission from grid i (kg C/ha). f_i denotes the non-linear response function of CH₄ emissions to irrigation water, established by the Random Forest model.

2.7.2. Constraints

(1) Irrigation water availability constraints

The total irrigation water allocated to each region shall not exceed the available water resources in the study area, and this constraint is expressed as follows:

$${\displaystyle \sum _i^{I}I{Q}_i\cdot A_i}\le Q_R$$

(8)

where IQ_i represents the irrigation quota (m³/ha) for rice in grid i, A_i denotes the area under rice in grid i (ha), and Q_i is the amount of water available in the area (m³).

(2) Water demand constraints

Irrigation water allocation must satisfy the water requirements of crops. This constraint can be expressed as: Safe food production is ensured by specifying the amount of water allocated to each unit to meet the crop’s water needs.

$${\displaystyle \frac{I{Q}_i-E{T}_i}{E{T}_i}}\le \delta$$

(9)

where IQ_i is the annual irrigation water volume (m³/ha) for rice in grid $i$ after optimization, ET_i is the water demand (m³/ha) for rice in grid $i$, the water demand decision parameter $\delta$ is set at 8% [23], so that irrigation allocation meets crop water demand within an acceptable deficit range.

The NSGA-II algorithm was configured with the following parameters: population size = 200; number of generations = 200; crossover probability = 0.9 (simulated binary crossover, SBX); mutation probability = 0.1 (polynomial mutation); distribution indices for crossover and mutation were set to 20 and 100, respectively. The Pareto front was selected based on the maximum hypervolume criterion. Convergence was assessed by monitoring the stability of the hypervolume metric over the last 100 generations. The final irrigation schedule for each grid was selected from the Pareto-optimal set as the solution with the minimum Euclidean distance to the ideal point (utopia point) in the normalized objective space, balancing water productivity maximization and methane emission minimization.

2.8. Decision-Making Process for Rice Irrigation Water Optimization

We coupled GIS with crop models to simulate rice growth and CH₄ emissions in the Sanjiang Plain under varying irrigation water amounts. The relationship between irrigation water amount and yield response was established using the Random Forest algorithm, while the corresponding relationship between irrigation water and CH₄ emissions provided the baseline data for the optimization model. Subsequently, a multi-objective optimization model was developed. The objective function included both water productivity and CH₄ emissions, with constraints that accounted for crop water demand and water availability. The decision variables were irrigation water amount and irrigation timing. The multi-objective model was solved using the NSGA-II algorithm, and irrigation schedules were determined based on the water demand of rice in each grid cell. Each grid cell served as the fundamental unit, enabling refined spatio-temporal optimization of irrigation decisions for rice cultivation across the Sanjiang Plain. The overall workflow of the multi-objective optimization for rice irrigation water allocation is illustrated in Figure 3.

3. Results

3.1. Comparative Analysis of Rice Yield Before and After Optimization

The analysis of rice yield distribution across cultivation grids in the Sanjiang Plain (Figure 4a,b) revealed that the optimized irrigation regime significantly increased rice yields, with yield improvements observed in most grids, underscoring its potential to enhance rice productivity. To assess the authenticity of simulated yields, we compared the spatial distribution patterns of optimised yields with regional averages and cross-referenced them against official rice production statistics published in the Heilongjiang Statistical Yearbook (2022). The comparison revealed that spatial trends were largely consistent. As shown in Figure 4c, 86.1% of the grids exhibited yield gains after optimization, while approximately 13.9% showed decreases—a pattern concentrated mainly in the Baoqing area.

This localized yield reduction is an expected outcome of the multi-objective optimization framework, which prioritizes regional-scale water productivity improvement and methane emission reduction over uniform yield gains across all grids. The Baoqing cluster exhibits a higher yield sensitivity to irrigation reduction; thus, when greater irrigation reductions are allocated to this area to maximize regional water savings and emission reductions, soil water deficits during critical growth stages become more pronounced, leading to quantifiable yield penalties. Under this trade-off, such grids are preferentially assigned to water-saving and emission reduction objectives despite moderate local yield loss. The impact of the irrigation optimization on yield varies across regions and soil conditions, as illustrated in Figure 4d. Yield increases are inconsistent among grids, with the largest gains (1.5–3.5 × 10⁴ kg) predominantly occurring in the northern part of the plain, which accounts for 31.1% of the total yield increase. This suggests that the irrigation optimization had a particularly significant effect in the northern region, especially in low-yielding areas such as Luobei, Suibin, and Huachuan, where the increase is more pronounced (2.0–3.5 × 10⁴ kg). Overall, the irrigation optimization led to an increase in total rice yield across the Sanjiang Plain, with an estimated yield gain of approximately 1.2 billion kg (8.4%), compared to pre-optimization levels. These findings highlight the potential of optimized irrigation technology to boost grain yield across diverse rice field scenarios, providing valuable insights for future spatio-temporal irrigation optimization and the formulation of policies for efficient regional agricultural water management.

3.2. Comparative Analysis of Methane Emissions Before and After Rice Optimization

The optimization of the irrigation system played a significant role in reducing rice CH₄ emissions (Figure 5a,b). After optimization, CH₄ emissions were effectively reduced in the majority of regions. However, in the northwestern areas—including Luobei, Huachuan, Hegang, Tangyuan, and Yilan—approximately 9.3% of the grids showed no observable reduction in CH₄ emissions (Figure 5c). This outcome represents an expected trade-off within the multi-objective optimization framework, which prioritizes regional-scale improvements over uniform gains across all grids. The absence of mitigation in these areas is primarily attributable to two factors: (1) heavier-textured soils and higher organic matter content reduce the sensitivity of methane production to irrigation reductions; and (2) in some grids, the algorithm constrained irrigation reduction to avoid yield penalties, thereby favoring global water productivity over local emission reduction. Such spatial heterogeneity underscores the value of high-resolution grid-based optimization in identifying the limitations of uniform strategies and supporting spatially differentiated management. In contrast, CH₄ emissions were reduced in 90.7% of the grids (Figure 5c). These results demonstrate that irrigation optimization measures effectively curbed CH₄ emissions, contributing to the rice GHG emission reduction target.

The network distribution within the optimized CH₄ reduction is shown in Figure 5d. Grids with significant reductions were primarily concentrated in the 150–200 kgC range, with 81.3% of the grids showing a reduction of over 100 kgC. This CH₄ mitigation is mechanistically attributed to the optimized irrigation regime’s regulation of soil redox potential. By reducing irrigation frequency and amount, the schedule prolongs aerobic phases, suppresses methanogenic activity, and enhances CH₄ oxidation. The spatial heterogeneity of this mitigation effect across the regional grid scale reveals substantial variation with local soil and management conditions. Overall, CH₄ emissions from paddy fields in the Sanjiang Plain were reduced by up to 1.5%. As a result of a significant increase in rice yield (by 8.4%, Section 3.1), the methane emission intensity per unit of production decreased even more notably, with CH₄ emissions per unit of production reduced by up to 9.6%. Therefore, with the widespread adoption of optimized irrigation systems, raster-scale spatio-temporal optimization of rice irrigation can play a key role in reducing greenhouse gas emissions in agricultural production, contributing to climate change mitigation efforts.

3.3. Comparative Analysis Before and After Optimization of Regional Irrigation Systems

The optimized irrigation system significantly reduced the irrigation water applied for rice cultivation (Figure 6a,b). Compared with pre-optimization levels, irrigation water decreased after optimization, with 90.9% of the grids achieving varying degrees of water savings. However, 9.1% of the grids—primarily in the Hulin area—did not exhibit the desired water-saving effect (Figure 6c). The distribution of water savings indicated that the majority of grids saved between 0.5 and 3 × 10⁴ m³ (Figure 6d), demonstrating the broad applicability of the optimized irrigation system for water conservation. The total regional irrigation volume under the baseline (pre-optimization) scenario was approximately 12.6 billion m³. After optimization, approximately 1.3 billion m³ of irrigation water was saved across the rice-growing region of the Sanjiang Plain, corresponding to a water-saving ratio of 10.3%. Although water savings were not achieved in all areas, the overall trend of water conservation was evident. Therefore, optimizing the irrigation system in the large rice-growing areas of the Sanjiang Plain resulted in substantial water conservation, helping to alleviate agricultural water stress and promoting the sustainable development of agricultural production.

The distribution of irrigation frequency under the optimized irrigation regime (Figure 7a,b) shows that the optimized irrigation frequency across grids ranged predominantly from 7 to 14 events, with 42% of grids adopting 12 irrigation cycles. This range aligns with—and in many cases rationalizes—common local farming practices, which often involve frequent flooding, by providing a scientifically based schedule that reduces total irrigation applications while targeting key growth stages. These results indicate that the optimized irrigation regime effectively maintained the necessary water conditions for rice growth while minimizing water waste. By reducing the number of irrigation events, substantial water savings were achieved without compromising rice productivity. Furthermore, the strategic adjustment of irrigation frequency helped reduce evaporative losses, thereby improving water use efficiency. Overall, the optimized spatio-temporal irrigation scheme not only reduced total water use but also fine-tuned irrigation timing and frequency to better match rice growth stages, achieving both water conservation and yield enhancement.

The distribution of irrigation timing (Figure 6 and Figure 7c) indicated that irrigation events were primarily concentrated between Julian days 137 and 256 (with 1 January as day 1), encompassing key growth stages from the greening stage to the milky-ripening stage. This scheduling was developed within the model framework, which considers crop water demand as the primary driver. For the regime with nine irrigation events, the irrigation period mainly spanned days 137–256, with variable timing for each event. For example, the first irrigation occurred between days 138 and 151, the second between days 151 and 161, and the third between days 160 and 179, and so on. When the number of irrigation events increased to 12, the schedule was further refined: the first irrigation occurred between days 137 and 149, the second between days 149 and 160, the third between days 158 and 172, the fourth between days 163 and 182, and subsequent events followed similar patterns. Analysis revealed that optimal irrigation timing exhibited clear spatial patterns correlated with local conditions. For instance, grids with sandy soils or those in warmer sub-regions tended to receive slightly earlier and more frequent irrigations during the early growth stages to compensate for higher percolation or evapotranspiration rates, whereas clay-rich soils in cooler areas allowed for longer intervals between events. Irrigation timing was carefully scheduled to match the water demands of rice at different growth stages. During the greening stage, irrigation promoted root system development. During the tillering stage, three to five irrigation events ensured optimal tiller growth. Irrigation during the panicle initiation, spike, and spike-flowering stages supported the physiological activities of rice. Finally, irrigation during the milky-ripening stage ensured proper grain filling and yield quality.

3.4. Comparative Analysis of Moisture Productivity Before and After Optimization

The optimized irrigation regime significantly improved water productivity in most of the grids (Figure 8). In this study, water productivity (WP), also referred to as water use efficiency (WUE), is defined as the rice yield (kg) per unit of irrigation water consumed (m³), expressed as kg/m³. Except for 11.6% of the grids, 88.4% showed an increase in water productivity after optimization, with 19.5% of the grids experiencing an improvement of 0.2–0.3. Overall, the average water productivity increased from 1.1 to 1.3, representing a 21.8% improvement. This enhancement aligns with or exceeds the range of WP improvements (typically 10–20%) reported in prior meta-analyses on water-saving irrigation in Chinese rice systems [7,24], confirming the effectiveness of the proposed spatio-temporal optimization approach.

The increase in WP was primarily driven by two complementary mechanisms: A more precise matching of irrigation timing and volume to rice physiological demands across growth stages, which reduced non-productive water losses (e.g., evaporation and deep percolation); The synergistic yield enhancement (8.4% overall) achieved through optimized water supply. Spatial analysis further revealed that WP gains were strongly correlated with soil texture (higher improvements in loamy soils), seasonal precipitation (greater improvements in drier sub-regions), and antecedent irrigation efficiency (larger gains where baseline WP was lower). These factors collectively explain the spatial heterogeneity in WP responses.

The improvement in water productivity showed marked spatial variability, driven mainly by differences in soil texture, precipitation patterns, and irrigation infrastructure. For example, grids with sandy soils or lower seasonal rainfall exhibited smaller gains in water productivity due to increased percolation losses and higher water deficit. In contrast, areas with clay-rich soils and advanced irrigation systems achieved above-average improvements. Overall, the optimized irrigation regime improved water-use efficiency by aligning water supply with crop demand throughout key growth stages, thereby minimizing non-productive water losses. The 21.8% increase in water productivity also yielded considerable economic benefits, such as reduced pumping costs, lower energy use for water delivery, and higher economic return per unit of water consumed. Under water-limited conditions, this strategy not only promotes increased rice yields but also supports the sustainable management of agricultural water resources.

4. Discussion

This study developed a high-resolution (500 m × 500 m) grid-based optimization framework that successfully reconciles water conservation, yield enhancement, and methane mitigation in rice production on the Sanjiang Plain. Compared to earlier coarse-scale applications (e.g., 10 km) of coupled GIS-DSSAT/DNDC models [15,16], our 500 m discretization substantially improves the representation of local biophysical conditions and enables more precise irrigation prescriptions. The 21.8% improvement in water productivity achieved in this study aligns with or exceeds the range of 10–20% reported in meta-analyses of water-saving irrigation in Chinese rice systems [7,9], confirming the effectiveness of the proposed spatio-temporal optimization approach.

4.1. Interpretation of Synergistic Benefits

The optimized irrigation regime, applied across more than 400,000 grid cells, yielded significant synergistic benefits. Relative to conventional practices, it reduced total regional irrigation water use by 10.3%, increased average rice yield by 8.4%, enhanced water productivity by 21.8%, and decreased methane emission intensity per unit yield by 9.6%. These outcomes demonstrate that precision water management can simultaneously address water scarcity, food security, and climate objectives in major rice-growing regions. The 10.3% reduction in irrigation water use is primarily attributable to optimized scheduling, which reduces both irrigation frequency (from 13–15 to 7–14 events) and total volume while maintaining soil moisture within optimal ranges for rice growth. The concurrent 8.4% increase in yield suggests that conventional excessive flooding not only wastes water but may also create suboptimal growing conditions due to prolonged anaerobic stress. Furthermore, the 9.6% reduction in methane emission intensity reflects the suppression of methanogenic activity during aerobic soil phases induced by reduced irrigation frequency.

4.2. Comparison with Alternative Irrigation Strategies

The optimized schedules exhibit conceptual similarities with alternate wetting and drying (AWD), a widely studied water-saving technique in rice systems [25]. Meta-analyses have reported that AWD can reduce CH₄ emissions by 30–50% compared to continuous flooding [24,26]. The comparatively lower reduction of 9.6% in emission intensity observed in this study reflects the multi-objective nature of our optimization framework, which balances water productivity, yield maintenance, and emission reductions. Unlike uniform AWD recommendations that typically prescribe fixed drying periods, our framework generates spatially differentiated schedules tailored to local soil and climate conditions. This approach avoids yield penalties in sensitive areas (e.g., sandy soils) while maximizing regional water savings. This distinction is critical: the 13.9% of grids that experienced yield reductions (Section 3.1) would likely have suffered greater losses under a uniform AWD protocol, underscoring the value of a spatially explicit optimization approach.

4.3. Spatial Heterogeneity and Its Implications

Spatial analysis revealed that optimal irrigation timing and frequency exhibit clear patterns correlated with local conditions. Grids characterized by sandy soils or located in warmer sub-regions tended to receive slightly earlier and more frequent irrigations during early growth stages to compensate for higher percolation or evapotranspiration rates. Conversely, clay-rich soils in cooler areas allowed for longer intervals between irrigation events. This spatial differentiation is essential for maximizing regional-scale benefits while minimizing localized penalties. The yield reductions observed in 13.9% of grids, primarily concentrated in Baoqing, were mechanistically linked to soil texture and climatic anomalies, confirming that the optimization framework successfully identifies its own boundary conditions.

4.4. Limitations and Uncertainty

The optimization results presented herein are derived from deterministic model simulations that do not formally propagate uncertainties. Although the coupled models were calibrated and validated against available data (see Section 2.5), uncertainties inherent in input data, model structure, and parameter equifinality are not explicitly quantified in the final optimized schedules. As a result, the reported numerical values should be interpreted as indicative of synergistic potential within the defined modeling framework rather than as precise predictions with quantified confidence intervals. In addition to this core limitation, several further constraints exist: reliance on single-year (2022) meteorological data, which may not adequately capture interannual climate variability; the absence of multi-year field validation across diverse sites; the lack of an economic analysis to assess farm-level profitability and adoption incentives; and the necessity for stakeholder engagement to address practical implementation constraints.

4.5. Future Research Directions

Based on the identified limitations, future research should prioritize several interconnected directions. Multi-year field experiments (3–5 years) at representative sites are necessary to validate the robustness of optimized irrigation schedules under varying hydroclimatic conditions and to refine model parameters. Additionally, formal uncertainty quantification—utilizing methods such as Sobol sensitivity analysis and Monte Carlo simulation—should be implemented to assess how parameter uncertainties influence the stability of Pareto-optimal solutions and to establish confidence intervals around predicted benefits. Integrated economic assessments are essential to evaluate farm-level profitability, incorporating investment costs for monitoring equipment, potential savings from reduced pumping costs, and yield risks associated with imperfect implementation. Furthermore, the optimization framework must be coupled with downscaled climate projections to design adaptive irrigation strategies for future climatic conditions and to evaluate the long-term sustainability of proposed schedules. Stakeholder engagement through participatory research with local farmers and extension agents is critical for identifying practical barriers to adoption and for developing user-friendly decision support tools—such as mobile applications or web-based platforms—that translate optimized schedules into actionable irrigation recommendations.

5. Conclusions

This study successfully developed and demonstrated a high-resolution (500 m × 500 m) grid-based optimization framework for rice irrigation, integrating process-based crop and methane models (GIS-DSSAT, GIS-DNDC), machine learning (Random Forest), and multi-objective evolutionary optimization (NSGA-II). Applied to the Sanjiang Plain—a major rice-producing region in Northeast China—the framework yielded the following key findings:

(1)

The optimized irrigation regime reduced total regional irrigation water use by 10.3%, increased average rice yield by 8.4%, enhanced water productivity by 21.8%, and decreased methane emission intensity per unit yield by 9.6%. These results demonstrate that precision water management can simultaneously address water scarcity, food security, and climate mitigation objectives.

(2)

The optimal irrigation strategy involved 7–14 irrigation events concentrated during the critical growth window from tillering to milky ripening (days 137–256), reflecting a strategic alignment of water application with crop physiological demand.

(3)

Spatial differentiation based on local soil and climate conditions is essential: grids characterized by sandy soils or experiencing above-average rainfall required schedule adjustments to avoid yield penalties, thereby confirming that uniform irrigation recommendations are suboptimal at regional scales.

The framework serves as a scalable decision-support tool. Extrapolating the 10.3% water-saving ratio to China’s total rice area (~30 million hectares) suggests an upper-bound theoretical potential of approximately 30 billion cubic meters of irrigation water saved annually. However, this estimate assumes similar conditions across all rice-growing regions and should not be interpreted as a precise prediction.

This study advances irrigation management from uniform practices to a spatially differentiated paradigm, demonstrating that synergistic outcomes across water, food, and climate dimensions are achievable through integrated modeling and optimization. Future research should prioritize multi-year validation, uncertainty quantification, economic analysis, and stakeholder engagement to enhance the robustness and practical applicability of the framework, thereby supporting the transition toward sustainable rice production at regional and national scales.