Optimized Water Management Promotes Greenhouse Gas Mitigation in Global Rice Cultivation Without Compromising Yield

Abstract

Rice is a vital staple food crop worldwide and also one of the major sources of greenhouse gas (GHG) emissions, generating substantial methane (CH₄) and nitrous oxide (N₂O). As one of the key management practices for rice production, the GHG mitigation potential of water management has attracted extensive attention, whereas its global scalability remains to be further investigated. Based on 15,458 global observations of field experimental data, we employed advanced machine learning methods to quantify the GHGs and soil carbon sequestration of global rice systems around 2020. Furthermore, we identified the optimal spatial distribution of GHG mitigation for five rice water management practices (continuous flooding (CF), flooding–midseason drainage–reflooding (FDF), alternate wetting and drying irrigation (AWD), flooding–midseason drainage–intermittent irrigation (FDI), and rainfed cultivation (RF)) through scenario simulation, under the premise of no yield reduction. The results of machine learning simulation showed that optimizing water management could reduce global rice greenhouse gas emissions by 39.17%, equivalent to 340.46 Mt CO₂ eq, while increasing rice yields by 3.55%. This study provides valuable insights for the optimization of agricultural infrastructure and the realization of agricultural sustainable development.

1. Introduction

Rice (Oryza sativa L.), a staple food crop for more than half of the world’s population, boasts a long cultivation history [1]. While serving as a critical source of dietary energy, rice is also recognized as one of the largest sources of greenhouse gas (GHG) emissions globally [2]. On the one hand, the unique flooding process in rice fields facilitates anaerobic microbial reactions in the rhizosphere, thereby elevating methane (CH₄) emissions [3]. On the other hand, the excessive application of nitrogen fertilizers stimulates substantial nitrous oxide (N₂O) emissions from the soil [4]. Since the signing of the Paris Agreement, numerous countries have set ambitious GHG reduction targets, and carbon mitigation in rice production has emerged as a pivotal component in efforts to achieve global GHG abatement goals [5,6].

Water management is a crucial component in rice production [7]. Conventional continuous flooding irrigation induces the formation of a strongly anaerobic environment in paddy soils, which in turn provides favorable conditions for the proliferation and metabolism of methanogenic archaea, ultimately leading to substantial methane emissions [8]. In contrast, optimized water management practices can effectively interrupt the CH₄ generation chain by breaking the continuity of the soil anaerobic environment [9]. For instance, mid-season drainage restores aerobic conditions in the soil through temporary drying, which significantly inhibits the activity of methanogenic archaea while promoting the growth of methanotrophs, thereby accelerating the oxidative degradation of residual methane in the soil [10]. The alternate wetting and drying mode, through the cyclic regulation of “wetting–drying” processes, can significantly reduce methane emissions on the premise of meeting the water demand for rice growth. However, the aforementioned optimized management practices pose the risk of increasing N₂O emissions while reducing CH₄ emissions [11,12]. Additionally, rice ecosystems exhibit a certain capacity for carbon sequestration, and water management practices are highly likely to modulate this carbon sequestration potential of rice systems [13]. Most importantly, water management is closely associated with rice yield, and thus the optimization of rice water management essentially entails a tripartite balance among GHG emissions, soil carbon sequestration rate (SOCSR), and rice yield [14].

Traditional field experiments of water management trials were conducted within a single field plot, resulting in research conclusions with certain geographical limitations [15]. Process-based models, such as DNDC, APSIM, and DSSAT, can predict GHG emissions, SOCSR, and crop yields by simulating biochemical processes; however, they often require intricate parameterization and calibration to achieve high simulation accuracy [16]. As a data-driven statistical tool, machine learning models can learn from existing field management experiences, enabling large-scale field simulations with high precision in a relatively short time [17]. Xiao et al. [18] optimized the irrigation amount and nitrogen application rate for wheat and maize in the North China Plain through a framework integrating machine learning and genetic algorithms, and developed grid-level agricultural management measures that achieve both emission reduction and yield increase. Yao et al. [19] optimized the water and nitrogen management of rice under different future climate scenarios in China using machine learning, which realized the triple benefits of nitrogen fertilizer reduction, GHG mitigation, and rice yield improvement.

To address the aforementioned research gaps, this study aims to achieve three core objectives: (1) develop high-accuracy machine learning models to predict key indicators (CH₄, N₂O, SOCSR, and rice yield) and interpret the impacts of water management practices on these target variables; (2) quantify the spatial patterns of these indicators for global rice production around 2020; (3) identify region-specific optimal water management regimes that minimize net GHG emissions without compromising yield. By fulfilling these objectives, this study is expected to provide a scientific basis for the optimization of agricultural water management infrastructure, offer actionable insights for sustainable rice production worldwide, and contribute to the achievement of global climate action and food security goals.

2. Materials and Methods

In this study, we compiled a global dataset of field trial observations and constructed four machine learning models to predict the indicators CH₄, N₂O, SOCSR, and rice grain yield. Subsequently, SHAP (SHapley Additive exPlanations) analysis was employed to evaluate the impacts of different water management regimes on the respective prediction targets. We then conducted simulations of net greenhouse gas (nGHG) emissions and rice grain yields at a spatial resolution of 0.05° across the globe around the year 2020. On this basis, we identified the water management regime with the minimum nGHG emissions for each rice-growing grid cell, subject to the constraint of maintaining current yield levels.

2.1. Data Collection

In this study, CH₄ and N₂O, SOCSR, and grain yield in rice cultivation were considered. The relevant data were obtained from peer-reviewed articles sourced from CNKI and Web of Science. The search language used was “(rice AND (CH₄ OR methane)) OR (rice AND (N₂O OR “nitrous oxide”)) OR (rice AND (“soil organic carbon sequestration” OR SOCSR OR SOC)) OR (rice AND yield)”. All the literature included in our dataset was required to meet the following criteria:

(1)

Field-based experiments were considered, while pot or greenhouse studies were excluded.

(2)

Information on experimental duration, latitude, and longitude was provided.

(3)

Detailed descriptions of agricultural management practices were given, covering water management, fertilization, straw management, and tillage.

(4)

At least one season of measured data on CH₄ emissions, N₂O emissions, SOCSR, or grain yield was reported, with model-simulated results excluded.

(5)

For CH₄ and N₂O emissions, we exclusively collected data measured via the static chamber method, given its status as the most widely adopted technique for field-based greenhouse gas flux measurements [20].

(6)

For the SOCSR, we retained data of the topsoil layer (0–20 cm).

(7)

For rice yield, we only included data that reported, or allowed for the calculation of, grain yield on a dry-weight basis.

(8)

All data collected are based on measurements over a single growing season. Studies that neither reported nor allowed for the indirect calculation of any of these four target variables for one growing season were excluded from the dataset.

(9)

For references without direct reporting of the SOCSR, we calculated the SOCSR during the rice growing season using the following formula:

$$SOCSR=({SOC}_t\times {BD}_t-{SOC}_i\times {BD}_i)\times H\times {10}^{-1}/t$$

(1)

where ${SOC}_t$ and ${SOC}_i$ denote the soil organic carbon content (g kg⁻¹) at the time $t$ and in the initial time $i$, respectively. ${BD}_t$ and ${BD}_i$ represent the soil bulk density (g cm⁻³) at the time $t$ and in the initial year $i$, respectively.$H$ is the soil depth (20 cm here). ${10}^{-1}$ is the unit conversion factor and denotes the experiment duration (season).

Based on these criteria, a total of 2341 CH₄ records, 1819 N₂O records, 800 SOCSR records, and 10,498 grain yield records were compiled, covering almost all major geographic and climatic zones for global rice cultivation (Figure 1, Table S1).

2.2. Machine Learning Models

For the four variables, including CH₄ emissions, N₂O emissions, SOCSR, and grain yield, we employed machine learning models (details presented later in the text) to predict their global distributions. The input features for the models comprised climate parameters, soil properties, and agricultural management practices. Climate parameters, including mean annual temperature and total annual precipitation, were extracted from the CRU TS 4.0 dataset [21] based on the geographic coordinates and experimental years of each site. Soil parameters consisted of soil texture (silt, sand, and clay content), pH, soil organic matter, total soil nitrogen, and bulk density. These were primarily obtained from the filtered literature; where such data were missing, they were supplemented using the Harmonized World Soil Database v2.0 [22] according to the site coordinates. Agricultural management practices covered water management (continuous flooding (CF) and flooding–midseason drainage–reflooding (FDF), alternate wetting and drying irrigation (AWD), flooding–midseason drainage–intermittent irrigation (FDI), and rainfed (RF) cultivation), fertilizer application (synthetic nitrogen, phosphorus, and potassium fertilization; organic nitrogen input), tillage (conventional tillage vs. no-till), and straw return rate (0–100%). The five water management options were encoded using one-hot encoding, with 1 indicating the adoption of a given practice and 0 otherwise, ensuring that only one water regime was assigned per observation. Tillage was treated as a binary variable, where 1 represented conventional tillage and 0 represented no-till. All remaining variables were continuous numerical inputs.

To identify the most accurate machine learning model, four classical machine learning algorithms, including Random Forest [23], Support Vector Regressor [24] (SVR), Gradient Boosting Decision Tree Regression [25] (GBDT), and eXtreme Gradient Boosting Regression [26] (XGBoost), were employed as candidate models for predicting the four target variables. First, all data collected in the dataset were divided into training and testing sets in an 8:2 ratio. Subsequently, the input features of the training dataset were standardized using Z-score (Equation (2)), where $Z_i$ is the standardized value of the i-th feature, $X_i$ is its original value, and ${\mu }_i$ and ${\sigma }_i$ represent the mean and standard deviation of the i-th feature, respectively. We performed hyperparameter tuning using five-fold cross-validation on the normalized training set, selecting the hyperparameter set that yielded the highest coefficient of determination (R²) as the optimal configuration for each model (see the ranges of hyperparameters in Table S2). Using these optimized hyperparameters, the models were trained on the full training set. Their performance was then evaluated on the test set using R², root mean square error (RMSE), and mean absolute error (MAE) as metrics for model selection (Equations (3)–(5))

$$Z_i={\displaystyle \frac{X_i-{\mu }_i}{{\sigma }_i}}$$

(2)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_{T_i}-y_{P_i})}^2}{{\sum }_{i=1}^n{(y_{T_i}-\overline{y_T})}^2}}$$

(3)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(y_{T_i}-y_{P_i})}^2}$$

(4)

$$MAE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_{T_i}-y_{P_i}\right|$$

(5)

where $y_{T_i}$ and $y_{P_i}$ denote the observed and predicted values of each target variable for the i-th instance, and n denotes the number of observations in the testing set. For each target variable, the model that performed optimally across the R², RMSE, and MAE metrics was selected as the final model. The final models utilized to simulate global rice GHG emissions, SOCSR, and grain yield were retrained on the entire dataset [27]. Considering that our machine learning model does not account for factors such as natural disasters and socio-economic conditions, our yield predictions may deviate from the actual situation. We therefore applied a correction factor to the predicted yield of each rice grid cell in each country [15], aligning the national total predicted yield with the 2020 statistical data (Equations (6) and (7)). In these equations, $Y_{ij}^{pred}$ represents the directly machine-learning-predicted yield for grid cell $i$ in country $j$, ${\alpha }_j$ denotes the correction factor of country $j$, and $Y_{ij}^{adj}$ is the adjusted yield. The term $Y_{total-j}^{pred}$ corresponds to the uncorrected sum of predicted yields across all grid cells in country $j$, while $Y_{total-j}^{FAO}$ is the total rice production for that country in 2020 recorded in FAOSTAT [28]. To further investigate the uncertainty of the machine learning model, we performed a bootstrap uncertainty analysis on the prediction results, as detailed in Text S1.

$$Y_{ij}^{adj}=Y_{ij}^{pred}·{\alpha }_j$$

(6)

$${\alpha }_j={\displaystyle \frac{Y_{total-j}^{pred}}{Y_{total-j}^{FAO}}}$$

(7)

To investigate the impact of different water management practices on the target variables, we employed the SHAP (SHapley Additive exPlanations) framework, an interpretable machine learning approach, to explain their impact in the final machine learning model [29]. Based on cooperative game theory, SHAP quantifies the contribution of each input feature to the model output using Shapley values. A positive Shapley value indicates that the feature has a positive effect on the prediction, while a negative value signifies a negative influence [30]. The sum of the Shapley values across all features equals the model’s predicted outcome. Using the entire dataset as input to the SHAP interpretability framework, we analyzed the effects of the five water management practices on variables CH₄, N₂O, SOCSR, and rice grain yield.

2.3. Global Mapping

Based on the finalized machine learning model, we simulated the baseline conditions of CH₄ emissions, N₂O emissions, SOCSR, and rice grain yield for global rice production. We utilized the global CROPGRIDS dataset [31] at a 0.05° resolution and, based on the latitude and longitude of each cultivation site, matched climate and soil parameters for each location from the CRU TS v4.0 [21] and HWSD v2.0 [22] datasets, using the average of years 2018–2022 as the reference. Subsequently, inorganic nitrogen, phosphorus, and potassium fertilizer application rates, as well as organic fertilizer inputs, were obtained from the NPKGRIDS dataset [32] and the study by Adalibieke et al. [33], according to the geographic location of each site. Data on straw management and tillage practices were sourced from global datasets by Smerald et al. [34] and Porwollik et al. [35], respectively. For water management, we used the global raster dataset by Grogan et al. [36] to extract the proportion of rainfed area for each location. Due to the absence of a more detailed dataset on current water management practices, we assumed that, apart from the rainfed areas specified in the Grogan et al. dataset, all other areas employed traditional continuous flooding irrigation.

2.4. Water Management Optimization

Based on the aforementioned climate and soil parameters as well as local agricultural management practices, we optimized water management for global rice cultivation grids at a 0.05° resolution. The specific approach involved, for each grid cell, keeping all other management practices unchanged while sequentially applying the best machine learning models to estimate GHG emissions and grain yields under five water management regimes: CF, FDF, AWD, FDI, and RF cultivation. nGHG emissions were calculated using Equation (8), which is derived from the IPCC AR6 [37]. The coefficients 27 and 273 represent the 100-year global warming potentials used to convert emissions of CH₄ and N₂O into CO₂-equivalent emissions. The factor 44/12 is the stoichiometric conversion coefficient used to convert changes in soil organic carbon (expressed in mass of C, as in SOCSR) into equivalent CO₂ emissions or removals. Maintaining yields without compromise to ensure food security constitutes a fundamental prerequisite for agricultural environmental mitigation research [38]. For each grid cell, we selected the water management practice that minimized nGHG emissions (${nGHG}_i$) without reducing yield under the baseline scenario as the optimal water management mode (Equation (9)). In this equation, ${n\mathrm{G}\mathrm{H}\mathrm{G}}_i\left(W_i\right)$ represents nGHG emissions under water management $W_i$, while $Y_{baseline-i}$ and $Y_{optimized-i}$ denote rice yields under the baseline and optimized scenarios, respectively. By multiplying the nGHG emissions (${nGHG}_{total-j}$) and yields ($y_{total-j}$) of each grid $i$ with its respective harvest area ($A_i$) and aggregating the results within country $j$, we obtained the total value of nGHG emissions and rice production for each country under both baseline and optimized scenarios (Equations (10) and (11)) [15,39].

$$nGHG=27\times {CH}_4+273\times N_{2}O-{\displaystyle \frac{44}{12}}\times SOCSR$$

(8)

$$\left\{\begin{array}{c}min:{nGHG}_i\left(W_i\right)\\ s.t. { Y}_{optimized-i}\left(W_i\right)\ge Y_{baseline-i}\end{array}\right.$$

(9)

$${nGHG}_{total-j}={\sum }_{i=1}^m{A}_i·{nGHG}_i$$

(10)

$$y_{total-j}={\sum }_{i=1}^m{A}_i·y_i$$

(11)

To further analyze the driving factors behind the optimized water management strategies, we computed the Pearson correlation coefficients between the five global optimized water management schemes and other climatic, soil, and agricultural management variables as follows:

$$r_{xy}={\displaystyle \frac{\sum _{i=1}^n(W_i-\overline{W})(y_i-\overline{y})}{\sqrt{\sum _{i=1}^n(W_i-\overline{W})(y_i-\overline{y})}}}$$

(12)

where $r_{xy}$ represents the correlation coefficient between variable $W_i$ (water management regime) and variable $y$ (climatic, soil, or agricultural management factor), $n$ denotes the number of observations, and $\overline{W}$ and $\overline{y}$ are the respective sample means.

3. Results

3.1. Performance of Machine Learning Models

The training results of the machine learning models are presented in

Table 1. Coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) of four machine learning models for CH₄, N₂O, soil organic carbon sequestration rate (SOCSR), and rice yield datasets. “Train” and “Test” denote the model’s performance on the training set and test set, respectively.

Target	Model	Dataset	R2	MAE	RMSE
CH4	Random Forest	Train	0.93	0.04 Mg ha−1	0.07 Mg ha−1
		Test	0.58	0.09 Mg ha−1	0.15 Mg ha−1
	SVR	Train	0.47	0.10 Mg ha−1	0.18 Mg ha−1
		Test	0.32	0.11 Mg ha−1	0.19 Mg ha−1
	XGBoost	Train	0.95	0.02 Mg ha−1	0.05 Mg ha−1
		Test	0.63	0.08 Mg ha−1	0.14 Mg ha−1
	GBDT	Train	0.93	0.04 Mg ha−1	0.06 Mg ha−1
		Test	0.60	0.08 Mg ha−1	0.14 Mg ha−1
N2O	Random Forest	Train	0.93	0.24 kg ha−1	0.44 kg ha−1
		Test	0.59	0.61 kg ha−1	1.15 kg ha−1
	SVR	Train	0.82	0.29 kg ha−1	0.73 kg ha−1
		Test	0.63	0.61 kg ha−1	1.09 kg ha−1
	XGBoost	Train	0.97	0.14 kg ha−1	0.27 kg ha−1
		Test	0.70	0.54 kg ha−1	0.99 kg ha−1
	GBDT	Train	0.95	0.21 kg ha−1	0.37 kg ha−1
		Test	0.67	0.56 kg ha−1	1.03 kg ha−1
SOCSR	Random Forest	Train	0.92	0.07 Mg ha−1	0.12 Mg ha−1
		Test	0.38	0.20 Mg ha−1	0.32 Mg ha−1
	SVR	Train	0.78	0.12 Mg ha−1	0.20 Mg ha−1
		Test	0.26	0.22 Mg ha−1	0.36 Mg ha−1
	XGBoost	Train	0.98	0.03 Mg ha−1	0.09 Mg ha−1
		Test	0.43	0.20 Mg ha−1	0.31 Mg ha−1
	GBDT	Train	0.96	0.05 Mg ha−1	0.09 Mg ha−1
		Test	0.43	0.19 Mg ha−1	0.31 Mg ha−1
Yield	Random Forest	Train	0.91	0.47 Mg ha−1	0.67 Mg ha−1
		Test	0.81	0.73 Mg ha−1	1.00 Mg ha−1
	SVR	Train	0.52	1.18 Mg ha−1	1.57 Mg ha−1
		Test	0.50	1.22 Mg ha−1	1.61 Mg ha−1
	XGBoost	Train	0.92	0.43 Mg ha−1	0.62 Mg ha−1
		Test	0.82	0.68 Mg ha−1	0.94 Mg ha−1
	GBDT	Train	0.91	0.48 Mg ha−1	0.67 Mg ha−1
		Test	0.83	0.69 Mg ha−1	0.93 Mg ha−1

. Based on a comprehensive evaluation of test-set performance (R², MAE, RMSE), XGBoost was selected as the final predictive model for CH₄ and N₂O emissions, while GBDT was chosen for SOCSR and grain yield. For CH₄ and N₂O, XGBoost achieved the highest test-set R² (0.63 and 0.70, respectively) alongside the lowest or competitive MAE and RMSE among all candidates. For SOCSR and grain yield, GBDT attained the highest test-set R² (0.43 and 0.83, respectively) and the lowest error metrics (MAE and RMSE). This selection is attributed to their consistent outperformance over Random Forest and SVR, which can be explained by several factors: (1) both XGBoost and GBDT employ gradient boosting frameworks that iteratively correct errors from previous models, thereby enhancing predictive precision and robustness [40]; (2) they inherently handle non-linear relationships and complex interactions more effectively through additive model construction and advanced regularization techniques [41]; (3) compared to Random Forest, they often achieve lower bias and better generalization by optimizing a differentiable loss function, while outperforming SVR in scalability and efficiency when dealing with large-scale, high-dimensional agricultural dataset [25,26]. The regression results of the four optimal models on both the training and test sets are shown in Figure 2. The prediction accuracy for grain yield was the highest on the test set, reaching 0.83. The prediction accuracy for the two GHGs, CH₄ and N₂O, ranked second and third, achieving values of 0.63 and 0.70 on the test set, respectively. Due to the pronounced volatility of SOCSR, the R² of its optimal machine learning model on the test set was relatively low at 0.43. The uncertainty of the machine learning model can be found in Figure S1.

3.2. Effect of Water Management on Rice Production

The SHAP-based interpretability analysis for water management in the machine learning model is illustrated in Figure 3. For the majority of the datasets, the SHAP values for CF and FDF were predominantly positive with respect to CH₄ emissions, indicating a consistent positive association between these water regimes and CH₄ release in the model. Notably, data points under FDF generally exhibited lower SHAP values than those under CF, a comparatively weaker positive association with CH₄ emissions. In contrast, most observations under AWD and RF cultivation yielded negative SHAP values, reflecting a negative association with methane emissions in our model predictions. This pattern is consistent with field experiments [42] and meta-analyses [43,44], which attribute the observed reduction to decreased anaerobic conditions and thus lower methanogenic activity. Although FDI also showed predominantly positive SHAP values for CH₄, their magnitudes were substantially smaller than those under CF and FDF, indicating a comparatively weaker promotion effect and a certain degree of mitigation relative to these two regimes in the model’s representation. These SHAP-derived patterns directly correspond to the biogeochemical cycles governing CH₄ dynamics in paddy soils. Prolonged anaerobic conditions under specific water regimes sustain methanogen activity and organic matter decomposition cycles, while periodic aerobic phases disrupt these cycles to suppress CH₄ generation [2]. For N₂O, in contrast to CH₄, water management practices that reduce CH₄ emissions, including AWD, FDI, and RF cultivation, mostly produced positive SHAP values, suggesting a positive association with N₂O emissions. Conversely, CF and FDF showed negative SHAP values, which is consistent with the inherent trade-off effect between anaerobic conditions favoring CH₄ and aerobic phases promoting N₂O production. Regarding SOCSR, except for a few outliers, none of the five water management practices exhibited a strong linear relationship with SOCSR, though FDF showed a slight negative influence. For grain yield, RF was consistently associated with reduced yield in the model’s outputs, while CF, FDF, and AWD displayed complex nonlinear relationships with yield. In contrast, FDI was predominantly associated with yield increase in most experimental observations.

3.3. Global GHG Emissions and Rice Yield

We mapped the global patterns of CH₄ emissions, N₂O emissions, SOCSR, and grain yield in rice cultivation (Figure 4a–d). High CH₄ emissions were primarily concentrated in irrigated rice-growing regions, such as China, Japan, and Western Europe. Elevated N₂O emissions were observed in parts of the Middle East, India, sub-Saharan Africa, and South America, which aligns with our earlier finding that RF systems tend to promote N₂O emissions, as these regions have substantial RF rice areas [45].

Strong soil organic carbon (SOC) sequestration was evident in regions such as China and West Asia. This positive trend may be attributed to several factors, including sufficient fertilizer input, widespread adoption of residue return practices, and increased application of organic amendments in these areas [46,47,48,49]. In contrast, many regions across Africa generally exhibited a net loss of SOC. This pattern is likely driven by factors such as lower organic matter inputs due to limited crop residue management, higher rates of soil erosion under conventional tillage systems, and potentially faster mineralization rates in warmer climates [50,51]. These observations align with findings from previous studies [16,52]. High-yielding areas were mainly distributed in East Asia, the United States, and South America, which typically benefit from higher fertilizer inputs and mechanized farming conditions [1]. In contrast, grain yields in sub-Saharan Africa were notably lower than in other regions. Regions with higher nGHG emissions were concentrated in East Asia, parts of sub-Saharan Africa, and east-central Brazil. The status of CH₄, N₂O, and SOCSR for major rice-producing countries is summarized in Table 1.

3.4. Optimization Results and Driving Factor Analysis

The spatial distribution of optimized water management practices is illustrated in Figure 5. We found that AWD is predominantly implemented in southern China, northeastern China, India, southern South America, and the southern United States. FDI is mainly concentrated in the central, western, and eastern regions of China. CF remains the primary management practice in the Middle East and Europe. RF cultivation is largely adopted in areas with abundant precipitation, such as the Amazon rainforest, equatorial Africa, and parts of Southeast Asia.

The corresponding impacts of these optimized practices on CH₄ emissions, N₂O emissions, SOCSR, and grain yield are shown in Figure 4f–i and summarized in

Table 2. Methane (CH₄), nitrous oxide (N₂O), and soil carbon sequestration rate (SOCSR) per hectare under baseline (BSL) and optimized (OPT) scenarios.

Region	BSL CH4 (kg ha−1)	OPT CH4 (kg ha−1)	BSL N2O (kg ha−1)	OPT N2O (kg ha−1)	BSL SOCSR (t C ha−1)	OPT SOCSR (t C ha−1)
Bangladesh	188.02	170.25	0.99	1.00	0.21	0.51
China	327.83	239.18	1.12	1.39	0.32	0.47
Indonesia	191.53	170.74	0.80	0.78	0.37	0.63
India	151.98	109.01	1.55	1.87	0.16	0.53
Japan	486.81	348.96	0.73	1.22	0.38	0.58
Thailand	221.49	196.88	1.36	1.38	0.29	0.49
United States	218.78	146.46	1.32	1.75	0.42	0.72
Vietnam	295.62	181.00	1.07	1.15	0.37	0.55
Africa	203.54	192.45	1.51	1.52	0.09	0.34
Asia	222.03	172.14	1.30	1.48	0.26	0.52
Europe	353.68	281.64	1.49	1.70	0.34	0.45
North America	204.46	147.24	1.22	1.54	0.34	0.66
Oceania	197.18	180.39	1.62	1.66	0.47	0.50
South America	213.33	185.52	1.55	1.56	0.24	0.47
Global	220.20	174.74	1.33	1.49	0.24	0.50

and

Table 3. Rice grain yield and greenhouse gas (GHG) emissions under baseline (BSL) and optimized (OPT) scenarios.

Region	BSL Yield (kg ha−1)	OPT Yield (kg ha−1)	Yield Increase Rate (%)	BSL nGHG (kg CO2 eq ha−1)	OPT nGHG (kg CO2 eq ha−1)	nGHG Mitigation rate (%)
Bangladesh	4596.85	4703.47	2.32	4577.85	3006.43	34.33
China	6969.48	7120.89	2.17	7988.42	5116.51	35.95
Indonesia	5125.94	5369.91	4.76	4024.93	2512.53	37.58
India	4090.98	4276.41	4.53	3945.34	1519.55	61.48
Japan	6831.51	6993.51	2.37	11,957.53	7641.40	36.10
Thailand	2931.67	3008.14	2.61	5296.03	3912.44	26.13
United States	8540.19	8714.23	2.04	4725.01	1789.02	62.14
Vietnam	6004.59	6276.92	4.54	6904.21	3195.65	53.71
Africa	2398.18	2549.41	6.31	5574.66	4368.12	21.64
Asia	4870.82	5037.20	3.42	5407.33	3161.73	41.53
Europe	6416.11	6468.71	0.82	8696.24	6414.34	26.24
North America	7209.92	7425.40	2.99	4610.41	1980.93	57.03
Oceania	7683.86	7795.69	1.46	4052.23	3497.35	13.69
South America	5993.10	6205.90	3.55	5299.44	3703.36	30.12
Global	4670.56	4836.56	3.55	5426.18	3300.81	39.17

. Following the implementation of optimized water management schemes, global rice greenhouse gas emissions are reduced by 39.17%, equivalent to 340.46 Mt CO₂ eq, while rice yields increase by 3.55%. We observed a systematic trade-off in this emission reduction process: the reduction in CH₄ emissions and the increase in SOCSR are achieved at the expense of elevated N₂O emissions (Table 2).

Results showed that the most significant GHG reduction occurred in regions including southern China, across India, southern Brazil, and North America, where emission reductions exceeded 50% (Figure 4k, Table 3). In contrast, the lowest reductions were achieved in the Middle East and Europe, which is consistent with the continued use of CF management in these regions (Figure 5). Yield increases were most pronounced in Africa, with an overall rise of 6.31% (Figure 4l), followed by certain parts of India and Southeast Asia. We noted that regions with more advanced agricultural technology, such as China, the United States, and Europe, exhibited relatively lower proportional yield increases (0.82–2.17%) under optimal water management, which may be attributed to their already high baseline yield levels.

The Pearson correlation coefficients (r) between optimized water management parameters and various input features are shown in Figure 6. We found that CF exhibited the strongest positive correlation with the baseline irrigation ratio (r = 0.2987), indicating that regions with higher baseline irrigation ratios showed a co-occurrence with the selection of CF in our optimization results. CF also showed a strong positive correlation with soil pH (r = 0.2603). In contrast, the adoption of CF was negatively correlated with mean annual temperature (r = −0.2117) and soil clay content (r = −0.1896). Clay-rich soils, characterized by poor aeration and prolonged waterlogging under flooding, can promote anaerobic conditions that enhance CH₄ emissions [53,54]. Within the framework of our emission-reduction optimization objective, this inverse statistical relationship is consistent with the model’s output, which less frequently selects CF for regions with high clay content.

FDF exhibited weak correlations (|r| < 0.2) with most other variables. It showed a weak positive correlation with sand content (r = 0.1624), as sandy soils, which may be related to the better aeration of sandy soils, facilitating soil structure recovery after mid-season drainage [55]. FDF was also weakly negatively correlated with silt content (r = −0.1149) and inorganic nitrogen input (r = −0.1053). Soils with higher silt content retain water more effectively, reducing the necessity for mid-season drainage, while fields with lower inorganic nitrogen inputs may rely more on drainage to regulate soil nitrogen availability. The generally low correlations between FDF and other climatic or management variables suggest that this regime appears less strongly constrained by climate and management practices, offering greater flexibility in application.

AWD showed a strong negative correlation with sand content (r = −0.4524). Sandy soils, due to their poor water retention, lose moisture rapidly during wet–dry cycles, which can pose challenges for maintaining the alternating “wet–dry” rhythm required by AWD and pose a risk to yield stability [56]. Correspondingly, in our model results, AWD shows a low co-occurrence with areas with high sand content. Conversely, AWD demonstrated a strong positive correlation with clay content (r = 0.4484). Clay soils, with their high water-holding capacity, can maintain soil moisture during dry phases and retain water during wet phases, properties that are consistent with the requirements of AWD. AWD also showed moderate positive correlations with silt content (r = 0.2731) and soil total nitrogen (r = 0.2158). Soils with moderate silt content balance water retention and aeration, and soils with higher total nitrogen can release nutrients under wet–dry alternation, meeting the nutrient demands of this regime. AWD was weakly negatively correlated with the baseline irrigation ratio (r = −0.1113), a pattern that is consistent with its lower dependence on irrigation and its conceptual suitability for areas with limited irrigation access. This aligns with the core water-saving advantage of alternate wetting and drying.

FDI showed moderate positive correlations with inorganic potassium (r = 0.2898), inorganic phosphorus (r = 0.2735), and inorganic nitrogen (r = 0.2222) fertilization, a pattern that aligns with the reported higher demand for inorganic NPK fertilizers under this regime. The sequential water management stages (flooding, drainage, intermittent irrigation) can accelerate soil nutrient loss, which may require greater fertilizer supplementation to maintain yields [57,58]. FDI also exhibited a weak positive correlation with the baseline irrigation ratio (r = 0.1670), a relationship that corresponds to the multi-stage water management requiring controlled irrigation infrastructure to support phased water regulation.

RF showed a moderate negative correlation with the baseline irrigation ratio (r = −0.3684), which is consistent with its conceptual definition as a rainfed system and its prevalent adoption in regions with low irrigation dependency. It also showed moderate negative correlations with silt content (r = −0.3558) and soil pH (r = −0.3008), indicating a statistical association between RF and areas characterized by lower silt content (i.e., higher proportions of sand or clay) and more acidic soils. RF was moderately positively correlated with sand content (r = 0.3136), a pattern that aligns with the characteristic of sandy soils to drain well, which may reduce the risk of waterlogging under RF conditions. RF showed weak negative correlations with the application rates of inorganic nitrogen, phosphorus, and potassium fertilizers (r ranging from −0.2831 to −0.1904), consistent with the typically low fertilizer inputs in regions where RF is widely practiced, such as sub-Saharan Africa and parts of Southeast Asia [59,60,61]. Finally, we observed that none of the water management regimes showed significant correlations with tillage practices.

4. Discussion

In this study, we employed a machine learning-based approach to predict global rice GHG, SOCSR, and grain yield. Traditional process-based models require the extensive and complex parameter tuning of soil, meteorological, and other input parameters to achieve high prediction accuracy [62,63]. This process involves substantial trial and error and demands advanced expertise in mechanistic modeling [64]. In contrast, while machine learning models also require parameter adjustment, the tuning focuses on model hyperparameters, which can be efficiently optimized through methods such as grid search within a short timeframe, making it accessible even to researchers without specialized process-modeling experience. In terms of simulation efficiency, process-based models typically incorporate numerous differential equations and empirical functions to simulate complex biochemical processes starting from crop growth, resulting in longer computational times [65]. Global-scale simulations with such models often entail substantial time and resource investments. Machine learning models, particularly tree-based algorithms such as Random Forest, that are now widely adopted in environmental sciences, rely on relatively simpler statistical simulation procedures. These models can predict global environmental indicators within minutes on a standard personal computer. However, compared to process-based models, machine learning models often function as “black boxes,” lacking sufficient interpretability [66,67]. This limitation has been addressed by interpretable machine learning frameworks such as SHAP and Partial Dependence Plots (PDP). In our research, SHAP effectively elucidated the influence of water management practices on the four target variables. Unlike meta-analysis, SHAP operates as a case-driven factor analysis tool capable of revealing both global and individual contributions of input features to the target variables. Nevertheless, compared to meta-analysis, SHAP relies on high-accuracy machine learning models trained on large datasets to yield reliable results, which may limit its applicability in research domains with limited data availability.

We observed significant spatial heterogeneity in both the baseline conditions and optimized outcomes of global rice GHG emissions, SOCSR, and yield. This pattern results from the combined effects of climate, soil properties, and current agricultural management practices. This highlights the necessity of moving away from a “one-size-fits-all” approach to global rice GHG mitigation. Instead, region-specific management strategies must be tailored to achieve the dual benefits of emissions reduction and yield enhancement. Furthermore, trade-offs represent another key challenge in GHG mitigation through water management optimization. As identified in our driver analysis, reducing CH₄ emissions may often lead to increased N₂O emissions, while simultaneously exerting non-linear influences on SOCSR and yield. Identifying such trade-offs requires high-accuracy simulation models to capture the complex interactions among these interconnected processes. Our study is closely aligned with the United Nations Sustainable Development Goals (SDGs), where reducing GHG emissions from rice cultivation corresponds to SDG 13 (Climate Action), while enhancing yield distribution contributes to SDG 2 (Zero Hunger). The trade-offs identified above further illustrate that advancing multiple SDGs often entails balancing conflicting objectives. Optimizing for synergistic benefits, therefore, represents a critical focus for future research.

While our study identifies the optimal water management practice for each rice cultivation grid cell, its practical implementation remains a considerable challenge. First, switching water management regimes often requires substantial investment in additional irrigation infrastructure. Especially in regions with a high concentration of low-income countries, such as sub-Saharan Africa, although the optimized scenarios can enhance yield, the cost of adopting entirely new irrigation systems may be prohibitive for local farmers, necessitating further support from governments and international organizations [68]. Second, due to data limitations, we only classified water management into five broad categories. In reality, rice water management typically requires more precise, real-time adjustments based on actual weather conditions, for example, increasing drainage during heavy rainfall to prevent waterlogging or supplementing irrigation during droughts to avoid crop water stress [69]. Finally, water management is only one component of agricultural management practices. To achieve greater emission reduction and yield improvement, coupling water management with other management strategies in an integrated optimization framework should be considered in future research. It is therefore crucial to distinguish between the biophysical mitigation potential mapped in this study and its practical adoptability. Our optimization provides a spatially explicit benchmark of what is theoretically achievable under idealized conditions. Translating this potential into on-ground reality, however, is contingent upon overcoming the socioeconomic constraints noted above (e.g., infrastructure, cost, labor, institutions). Future policy and research should bridge this gap by integrating biophysical optimization with socioeconomic feasibility analysis to design context-specific implementation pathways.

Our study has several limitations. First, while we endeavored to include comprehensive data on rice GHG emissions, SOCSR, and grain yield, the training dataset remains deficient in regions such as sub-Saharan Africa, South America, and Southeast Asia. This may lead to higher uncertainty in the simulation results for these areas. Second, given the scarcity of observational data, we were compelled to adopt certain assumptions and simplifications in our calculations, which had the potential introduce a degree of uncertainty. For instance, baseline water management practices for rice had to be assumed to estimate net greenhouse gas emissions and yields, which may lead to an overestimation of the optimization potential. As current spatially explicit cropland datasets do not distinguish between upland and paddy rice systems, our analysis does not differentiate between these cultivation types. Third, although machine learning models generally achieve high predictive accuracy, their application across diverse global climatic, soil, and socioeconomic conditions inevitably introduces a degree of error. Therefore, the effectiveness of the optimal water management modes identified requires further validation through field experiments. Fourth, our optimization framework did not account for the influence of socioeconomic factors such as infrastructure, labor, and water rights, which should be a key focus of future research. Furthermore, the SHAP and correlation analysis used to infer potential drivers of optimized water management choices is inherently observational. The identified associations, while informative for pattern recognition, do not establish causation. They reflect statistical co-occurrences within our global model and dataset, and their interpretation should be tempered by the complex, multi-factorial nature of real-world agricultural systems. Finally, as noted earlier, our optimization does not account for local socioeconomic conditions. Consequently, the identified optimal water management practices may not necessarily translate into the most favorable economic or environmental outcomes in practice.

5. Conclusions

In this study, we developed a global-scale assessment framework to quantify the potential of water management for reducing greenhouse gas (GHG) emissions while maintaining rice yield. Using 15,458 field observations, machine learning models were trained to predict key sustainability indicators. The XGBoost model achieved the highest predictive accuracy for CH₄ and N₂O, while the GBDT model performed best for SOCSR and rice grain yield. SHAP analysis revealed the distinct and interpretable roles of five water regimes on these variables. Applying this framework globally, we simulated that optimized water management could achieve a model-projected 39.17% reduction in GHG emissions alongside a 3.55% increase in yields, with strategies influenced differentially by local soil, climate, and management contexts. This analysis provides a data-driven, spatially explicit roadmap for targeting water management interventions in rice systems. However, the identified potential represents an idealized benchmark. Translating this potential into practice will require future efforts that integrate our findings with local socio-economic feasibility studies. This research provides valuable insights for advancing agricultural carbon emission mitigation and promoting sustainable agricultural development.