Producing more rice with fewer emissions: a global meta-analysis

Rice sustains more than half the global population and supports the livelihoods of 144 million smallholder farmers across 163 million hectares, producing 715 million tonnes annually^1,2. It is a vital economic lifeline, providing employment and sustaining livelihoods, particularly in Asia and Africa^3,4,5. However, while global rice demand is projected to double by 2050, production systems face increasing pressure from climate change impacts, such as saltwater intrusion and extreme weather^6,7,8.

In contrast, rice cultivation contributes significantly to climate change, accounting for over 50% of grain-related greenhouse gas (GHG) emissions and approximately 10% of global agricultural emissions^9,10. It represents 1.3–1.8% of total anthropogenic emissions⁹. In flooded fields, anaerobic conditions significantly increase methane (CH₄) emissions, which constitute 90% of the total field emissions, giving rice a Global Warming Potential (GWP) that is three to four times higher than wheat or maize^11,12. Consequently, rice production is a critical focus of climate change mitigation efforts in many nations¹³.

Balancing food rising demand with reduced GHG emissions requires optimising Grain Yield (GY) while lowering emissions in the rice field. An increasing number of studies have focused on identifying farming practices that lower GHG emissions in rice production without compromising yields^12,14. However, these studies frequently reveal trade-offs between maximising yield and minimising emissions, making it challenging to develop solutions applicable across diverse rice farming systems^15,16.

Greenhouse Gas Intensity (GHGI), defined as the ratio of GHG emissions to crop yield, offers a robust metric to evaluate both productivity and environmental sustainability in rice production systems. By measuring emissions per unit of yield, GHGI identifies agricultural practices that balance high yields with low environmental impact, addressing the dual challenge of food security and climate change mitigation. Yet, its use in meta-analyses remains limited despite its potential to inform sustainable agricultural practices^{14,17,18,19,20,21,22}, with most studies focusing on isolated practices, such as water management or fertiliser application, rather than the integrated nature of rice systems^20,23,24,25. These analyses often fail to account for interactions among farming practices, climatic conditions, and crop seasons. Moreover, their scope is typically confined to single-country studies or small datasets, restricting their ability to propose solutions adaptable across diverse global contexts^{14,17,18,19,20,21,22}.

To address these limitations, we conducted a global meta-analysis of 5322 field experiments from 504 studies (1991–2024), using Bayesian multivariate methods to simultaneously model yield and emissions (CH₄, N₂O, total GWP, and GHGI) while controlling for soil type, climate, season, and co-occurring practices. By synthesising data across diverse regions, practices, it identifies “Win-Win” practices for the first time that simultaneously enhance yields and reduce emissions in all stages of rice cultivation. Critically, we assess how practice effectiveness varies across different environmental contexts, specifically soil types, climate zones, and production seasons, revealing substantial baseline emission gradients and context-dependent intervention priorities. These findings provide evidence-based guidance for developing sustainable, low-emission rice systems tailored to regional conditions and farmer circumstances. This research serves as a valuable resource for policymakers, researchers, and farmers, providing guidance on strategic interventions to strengthen food security while advancing global climate objectives.

Greenhouse gas emissions and rice grain yield nexus

At a global scale, the baseline GHG emissions from rice cultivation estimated at the lowest GY (387 kg/ha) are 3519 kg CO₂eq/ha [CI: 1960, 5114] (Fig.1a). This corresponds to a GHGI of 0.96 kg CO₂eq/kg paddy [CI: 0.68, 1.26] (Fig.1b). CH₄ accounts for around 93.9% of total emissions, with N₂O contributing only 6.1%. Each additional kilogram per hectare of GY increases GWP by 0.11 kg CO₂eq, driven by rising CH₄ emissions (Fig.1a, c). The correlation coefficient between GY and GWP was relatively low (Fig.1a). GHGI decreased as GY increased (Fig.1b), with yield improvements outpacing emission increases across the dataset.

Farming practices, grain yield and greenhouse gas emissions

We evaluated farming practices’ effects on GY and GHG emissions, grouping them into four categories:

• Pre-planting – including farming systems, growing duration (varieties) and tillage methods.

• Planting and water management – covering planting techniques and water regimes.

• Nutrient management – encompassing nitrogen application, manure and compost use.

• Postharvest – including residue management.

A total of eight cultivation practices, divided into 37 categories (Supplementary Table 2), were analysed, with differences expressed as percentages for clarity, focusing on statistically significant results.

Six farming systems were identified and assessed in this review: Single-Rice, Double-Rice, Triple-Rice, Rice-Aquatic-Dual-Systems, Rice-Food-Crops, and Rice-Green-Manure. Rice-Aquatic-Dual-Systems exhibit the lowest GHGI (0.53 kg CO₂eq/kg paddy) (Fig. 2a), representing 17.78–38.02% emission reductions compared to other systems (Fig. 2c, Supplementary Fig. 1b). However, this system produced 9.75% and 11.03% less GY than Single-Rice and Rice-Green-Manure systems, respectively, limiting its suitability where productivity is prioritised.

Among higher-yielding systems, Single-Rice, Rice-Green-Manure, and Rice-Food-Crops demonstrate the greatest productivity, with yields ranging from 5.72 to 5.99 tons/ha. Among these, the Rice-Food-Crops system emits 21.7–22.43% less CH₄ than Single-Rice and Rice-Green-Manure per area unit (Supplementary Fig. 1b). In contrast, the Double-Rice system not only produces a lower yield, 16.86% and 12.01% lower than the Single-Rice and Rice-Food-Crops systems, respectively. It also exhibits a higher GWP than the Rice-Aquatic Dual systems (38.02%) and Rice-Food-Crops (17.18%), resulting in a higher GHGI.

Rice varieties were classified into five categories based on their growth cycle duration in the field: Ultra-Short-Duration (90 days or fewer), Short-Duration (91–110 days), Medium-Duration (111–130 days), Long-Duration (131–150 days), and Ultra-Long-Duration (151 days or more). Medium-duration varieties sustained GY without increasing GHG emissions. Regarding GHGI, our results indicate no significant differences among Medium, Ultra-Long, Short, and Ultra-Short varieties. However, Long-Duration varieties show significantly higher GHGI, 1.93% and 1.44% greater than Ultra-Short and Medium-Duration varieties, respectively (Fig. 2f). In terms of GWP, Ultra-Short and Medium-Duration varieties emitted 5.07–8.14% less per unit area than other varieties, with N₂O emissions increasing by up to 1.21% as growth duration lengthened (Supplementary Fig. 1f). Regarding GY, Medium and Long-Duration varieties increase GY by 2.99–5.16% compared to Short-Duration and Ultra-Short-Duration varieties.

Four tillage methods were identified and evaluated: No-Tillage, Conventional-Tillage, Reduced-Tillage, and Rotary-Tillage. No-Tillage reduced GWP by 16.62%, by 22.66–31.61% reduction in CH₄ emissions, compared to other tillage methods (Fig. 2k, Supplementary Fig. 1h), while GY declined by 6.26–8.23% (Fig. 2l). In contrast, Conventional, Reduced, and Rotary-Tillage show no statistically significant differences in terms of both GWP and GY. However, based on absolute GHGI, GWP, and GY values, Rotary-Tillage showed the lowest absolute GHGI (0.75 kg CO₂eq/kg paddy) among tillage methods with comparable GY to Conventional-Tillage (Fig. 2i).

Four planting methods were identified and evaluated: Drill-Seeding, Dry-Direct-Seeding (DDS), Transplanting, and Wet-Direct-Seeding (WDS). DDS exhibits the lowest GHGI among all planting methods, reducing emissions between 10.05 and 13.40% compared to other practices (Fig. 3b). While DDS slightly increases N₂O emissions (1.58–2.58%), it significantly reduces CH₄ emissions by 18.64% and 20.84%, leading to reduces GWP by 13.05%, 18.43% and 21.87% compared to Drill-Seeding, Transplanting and WDS, respectively (Supplementary Fig. 2b, c, Fig. 3c).

In terms of GY performance, DDS produces comparable yields to Transplanting while generating 6.1% higher yields than WDS and 18.76% more than Drill-Seeding. WDS, despite generating the highest GWP emissions, produces 10% lower GY than Transplanting. Drill-Seeding offers greater GWP reduction potential than Transplanting and WDS, but results in significantly lower levels of 11.93–24.03% compared to other methods (Fig. 3d).

This study evaluated six distinct water management regimes: Alternate-Wetting-Drying (AWD), Continuous-Flooding (CF), Controlled-Irrigation, Intermittent-Irrigation, Mid-Season-Drainage (MSD), and Rainfed. CF exhibits the highest GHGI (1.2 kg CO₂eq/kg paddy), 14.17% higher than Rainfed (Fig. 3f, Supplementary Fig. 2e, f). However, CF outperforms Rainfed in yield, producing 13.72% more rice despite 38.04% higher GWP. Rainfed showed the lowest GHGI due to reduced CH₄ (despite higher N₂O emissions), but produced 12.06–18.1% less than the other irrigation methods.

Compared to CF, advanced irrigation methods (AWD, Controlled-Irrigation, Intermittent, MSD) reduce GHGI. These methods significantly decrease GWP, particularly Controlled-Irrigation (20.33% reduction), Intermittent (15.04%), and AWD (11.52%), without compromising GY (Fig. 3g). AWD achieves the highest yields, outperforming CF and MSD by 3.85% and 5.39%, respectively, while Controlled-Irrigation yields 2.83% more than MSD (Fig. 3h). Although these advanced irrigation methods increase N₂O emissions due to reduced waterlogging, CH₄ reductions are significantly greater, resulting in a net decrease in GWP (Supplementary Fig. 2d–f).

Nitrogen application was classified and evaluated across four levels: no application (Zero-Nitrogen), low (1–100 kg/ha), moderate (101–200 kg/ha), and high (over 200 kg/ha). GHGI decreased up to 1.44% from Zero-Nitrogen (1.02 kg CO₂eq/kg paddy) to higher application rates (Fig. 4a, b). Nitrogen application increased total GHG emissions by approximately 6% (8% higher CH₄ and elevated N₂O) (Supplementary Fig. 3b, c), while yields increased by 32.03–39.11% compared to Zero-Nitrogen application (Fig. 4d). No significant differences in GHGI were observed among the three nitrogen application levels (low, moderate, high).

Compost and manure applications were classified into three categories: Compost-Application, Manure-Application, and Zero-Application. Both treatments increased GY by ~10% while increasing GHGI. Manure-Application shows higher GHGI (0.73 kg CO₂eq/kg paddy) than Compost-Application (0.68 kg CO₂eq/kg paddy). Both increase CH₄ emissions by 8.62–11.37% and N₂O emissions by 0.48–0.58% (Supplementary Fig. 3d–f, Fig. 4g). Consequently, compost increases GHGI by 1.6% while manure increases it by 2.61% compared to Zero-Application (Fig. 4f).

Five residue treatment methods were identified: Incorporated-Green-Manure, Incorporated-Straw (including stubble), Incorporated-Stubble-Only, On-field-Burning, and Removal. Straw-Removal and On-field-Burning achieved the lowest GHGI among residue management practices (Fig. 5b). The amount of rice straw and stubble incorporated into the soil increased both GY and GWP proportionally. Incorporating straw (straw and stubble) into the field doubles emissions compared to removing all residues. Simply removing straw (but incorporating stubble) reduces emissions by 25.06% per unit area compared to full incorporation (Fig. 5c). In terms of GY, retaining straw results in a 4.53% GY increase compared to removing (Fig. 5d). Consequently, straw removal achieves 8.27–22.69% lower GHGI than other incorporation methods (Fig. 5b).

On-field-Burning and Incorporated-Green-Manure before planting improve GY, with green Manure being particularly effective, increasing GY by 5.53–10.31% compared to other methods. Green manure leads to higher CH₄ and N₂O emissions compared to Removal and On-field-Burning (Supplementary Fig. 4a–c), though total emissions remain 31% lower than full straw incorporation (Fig. 5c). Meanwhile, On-field-Burning reduces CH₄ emissions by 62.92%, achieving 59.5% and 27.3% lower total emissions than Incorporated-Straw and Incorporated-Stubble-Only, respectively (Fig. 5c, Supplementary Fig. 4b).

Context-dependent of greenhouse gas emissions and grain yield

To assess context-dependency, we analysed GHGI variations across soil types, climate conditions, soil pH ranges, and production seasons (Fig. 6). Soil type showed the strongest effects: organic-rich soils (high organic matter) exhibited 42% higher GHGI (0.88 kg CO₂eq/kg paddy, Fig. 6a) than fine-textured soils (predominantly clay, 0.62 kg CO₂eq/kg). Continental climates exhibited the highest GHGI (1.00 kg CO₂eq/kg, Fig. 6b), which was 66% and 54% higher than those of arid and temperate climates, respectively (0.34 and 0.65 kg CO2eq/kg). Tropical climates showed intermediate GHGI (0.57 kg CO₂eq/kg). Production season affected GHGI (Fig. 6d), with late-season cultivation showing the highest GHGI (1.35 kg CO₂eq/kg), followed by early-season (1.33 kg CO₂eq/kg), wet-season (1.23 kg CO₂eq/kg), and dry-season (1.16 kg CO₂eq/kg). Soil pH showed limited variation in GHGI (0.61–0.80 kg CO₂eq/kg across pH 5.0–8.5, Fig. 6c), with overlapping confidence intervals indicating no significant differences among pH ranges.

High-GHGI contexts (organic-rich soils, continental climates, late-season) exhibited higher absolute emissions than low-GHGI contexts (fine-textured soils, temperate climates, dry-season), though productivity differences were minimal or negative (Fig. 6e–l). For example, organic-rich soils emitted 3053 kg CO₂eq/ha, 33% more than fine-textured soils (2287 kg CO₂eq/ha), while producing only 8% higher yields (6078 vs 5645 kg/ha). Late-season cultivation generated 6193 kg CO₂eq/ha (44% higher than the dry season’s 4300 kg CO₂eq/ha) yet yielded 5% less (5269 vs 5569 kg/ha). Mixed soils achieved the highest yields (6197 kg/ha, Fig. 6i) with intermediate GHGI (0.68 kg CO₂eq/kg, Fig. 6a).

CH₄ was the primary emission component, varying across contexts. High-GHGI contexts generated 53–57% more CH₄ than low-GHGI contexts: organic-rich soils emitted 167 kg CH₄/ha (53% higher than fine-textured soils’ 109 kg/ha), continental climates emitted 141 kg CH₄/ha (54% higher than temperate climates’ 91 kg/ha), and late-season cultivation produced 160 kg CH₄/ha (57% higher than dry-season’s 102 kg/ha) (Supplementary Fig. 5a–d). N₂O emissions showed limited variation (0.27–1.07 kg/ha across most contexts, Supplementary Fig. 5e–h), except mixed soils (1.39 kg/ha, Supplementary Fig. 5e).

Our meta-analysis of 5322 field experiments demonstrates that optimised rice farming practices can reduce GHG emissions while maintaining or enhancing grain yield, thus reconciling food security and climate mitigation goals. This synthesis advances beyond previous meta-analyses^{14,20,22,23,24,26,27,28,29,30,31,32,33,34,35} through four key ways: (1) unprecedented scale and temporal coverage (504 studies, 1991–2024), providing the most comprehensive evidence base available; (2) holistic assessment of eight major practices spanning the entire production cycle, from pre-planting through post-harvest, versus isolated interventions; (3) novel focus on per-unit GHG emissions, capturing the yield-emission trade-off central to sustainable intensification; and (4) Bayesian multivariate framework simultaneously modelling yield and emissions while controlling for climate, soil, season, and co-occurring practices, yielding unbiased effect estimates. These advances enable systematic identification of “Win-Win” practices that enhance productivity while reducing emission intensity²⁵.

Rice paddy GHG emissions arise from distinct biogeochemical pathways that respond differently to management interventions. CH₄ production requires three essential components: an organic matter substrate, anaerobic conditions, and time^{36,37,38,39,40}. Continuous flooding creates oxygen-depleted environments where methanogenic bacteria decompose organic carbon into CH₄ through methanogenesis^16,41. Over 90% of produced CH₄ bypasses soil oxidation and reaches the atmosphere through plant-mediated transport via rice aerenchyma tissues, with less than 10% released via soil diffusion or bubble formation^37,42,43. N₂O production follows fundamentally different controls, requiring a nitrogen substrate and oxygen availability^44,45,46. Applied nitrogen undergoes nitrification (NH₄⁺ → NO₃⁻) in aerobic soil layers and denitrification (NO₃⁻ → N₂O) when oxygen becomes limited⁴⁷. Unlike CH₄, which requires sustained anaerobic conditions, N₂O emissions peak during aerobic-anaerobic transitions, particularly field drainage or wetting-drying cycles, when alternating oxygen availability enables sequential nitrification-denitrification^44,45,48.

Rice yield improvements result from enhanced photosynthesis, efficient nutrient uptake^49,50, which collectively increase biomass production. Greater biomass intensifies substrate availability for methanogenic bacteria in root zones, thereby elevating CH₄ emissions⁵¹. Simultaneously, enhanced soil nutrient cycling facilitates denitrification, which increases N₂O emissions⁵¹. These substrate-emission linkages predict that higher-yielding varieties should produce proportionally higher emissions. However, our analysis reveals a critical asymmetry: the correlation between grain yield and GWP is surprisingly weak (r = 0.11, Fig. 1a). This decoupling occurs because yield gains substantially outpace increases in emissions. Consequently, GHGI decreases as grain yield increases (Fig. 1b), demonstrating that varietal improvements reduce emissions per kilogram of rice even while absolute emissions may rise modestly. Since over 90% of CH₄ is plant-mediated, future rice breeding should integrate traits restricting aerenchyma conductivity⁵² alongside continued yield improvements.

Beyond genetic improvements, management practices represent the primary determinant of field emissions, offering more immediate and scalable mitigation by directly manipulating substrate availability, redox conditions, and temporal dynamics. Planting methods and water regimes jointly determine field hydrology. DDS eliminates initial field flooding, reducing emissions without yield penalties compared to transplanting (Fig. 3b–d). In contrast, WDS, the predominant in Southeast Asia⁵³, shows unfavourable outcomes for both emissions and yields (Fig. 3b–d). Among in-season water regimes, intermittent drainage strategies (AWD, controlled irrigation, and intermittent irrigation) reduce GWP by up to 11% compared to continuous flooding (Fig. 3f), with AWD and controlled irrigation achieving the lowest emissions intensity (Fig. 3e). These favourable outcomes occur because strategic water management satisfies both emission mitigation and crop physiological requirements simultaneously. Periodic aeration increases N₂O moderately but reduces CH₄ substantially, by up to 30% (Supplementary Fig. 2b). Simultaneously, controlled drainage during non-critical phases enhances root health and nutrient uptake, while limiting water loss during tillering, flowering, and grain filling^54,55. Reducing cumulative anaerobic time thus lowers emissions while improving rather than compromising growing conditions.

Substrate management practices influence emissions by altering organic matter and nitrogen availability, the primary substrates for CH₄ and N₂O production. Multi-season systems demonstrate substantial substrate accumulation: double-rice and triple-rice systems double CH₄ emissions compared to single-rice due to accumulated residues, while producing lower per-crop yields due to soil degradation and pest pressure (Fig. 2a, d)⁵⁶. Rice-upland crop rotations reduce aggregate emissions while maintaining productivity by breaking pest cycles and introducing varied residue chemistries (Fig. 2b, c). Residue incorporation (e.g., straw and stubble) and the application of organic amendments (compost and manure) increase substrate availability, thereby elevating both CH₄ and N₂O emissions while improving soil fertility. The processing method affects outcomes: straw incorporation doubles emissions for modest yield gains (~5%), whereas compost shows more favourable trade-offs, with yields increasing (~10%) outpacing emission increases (~6%) (Figs. 4f, g and 5). Nitrogen application influences emissions through dual pathways: directly via N₂O production and indirectly via biomass-mediated CH₄ generation. Critically, the indirect pathway dominates; nitrogen addition increases CH₄ emissions by ~8% while N₂O increases only around 2.5%. Despite these emission increases, optimised nitrogen management achieves net GHGI reductions (Fig. 4) because yield gains substantially outpace emission increases.

Multi-pathway factors influence emissions through mechanisms that extend beyond individual pathways. Growth duration shows complex patterns: short-duration varieties produce higher daily CH₄ emissions than medium-duration types despite shorter exposure, likely reflecting climate confounding where short-duration Indica dominates warm regions with accelerated methanogenesis⁵⁷ (Fig. 2). Medium-duration varieties balance compressed emission windows with improved harvest index (Fig. 2). Rice-aquaculture systems suppress emissions through enhanced soil aeration and CH₄ oxidation by aquatic organisms⁵⁸. Tillage alters soil structure and redox dynamics: no-tillage reduces CH₄⁵⁹ but limits yields⁶⁰, whereas rotary tillage improves drainage, reducing CH₄ while enhancing productivity (Fig. 2, Supplementary Fig. 1).

Soil type, climate, and season create substantial baseline emission gradients with asymmetric productivity effects. Fine-textured soils achieve the lowest emission intensity through reduced substrate diffusion and enhanced CH₄ oxidation in clay matrices⁶¹, but moderate yields reflect constrained fertility. Mixed soils demonstrate that optimal productivity requires substrate for crops, not minimising methanogenic substrate (Fig. 6a, i). Temperate climates limit methanogenic activity (0.65 vs 1.00 kg CO₂eq/kg paddy for continental climates) without yield advantages⁶² (Fig. 6b, j). Dry-season cultivation achieves optimal emission-yield balance through controlled flooding, while late-season production combines elevated emissions (+44%) with reduced yields (-5%) due to accelerated methanogenesis and pest pressure^62,63 (Fig. 6d,l). Soil pH exhibits minimal effects due to its buffering capacity (Fig. 6c, k). Critically, emissions respond more strongly to context than yields: high-baseline contexts generate 33–44% higher emissions while yields change modestly (+8% to −7%)⁶². This asymmetry creates strategic opportunities: low-baseline contexts (fine-textured soils, temperate climates, dry-season cultivation) offer inherently efficient production with naturally low emission intensities requiring minimal intervention, while high-baseline contexts (organic-rich soils, warm climates, late or wet-season cultivation) present the largest absolute reduction potential for mitigation interventions targeting anaerobic conditions or substrate availability, despite offering no inherent productivity advantages.

These mechanistic insights enable categorisation of farming practices through pairwise comparisons against best-performing benchmarks (Fig. 7). “Win-Win” practices simultaneously sustain yields while reducing emission intensity, representing optimal pathways for sustainable rice production. These include water management (AWD: 12% emission reduction, +4% yield; DDS: 10–13% emission reduction, +18% yield); single rice systems; optimised nitrogen management (+32–39% yield, -4–5% emissions); medium-duration varieties; rice-upland rotations; rotary tillage; and strategic residue management. These practices achieve optimisation rather than intensification by aligning emission reduction with plant physiological requirements (strategic water timing, efficient nutrient uptake) or ecological processes (pest disruption, soil improvement) without yield-emission trade-offs inherent in substrate addition or extended anaerobic periods.

“Yield-priority” practices enhance productivity at increased emission costs, which is relevant when food security outweighs emission concerns. These include straw incorporation (+4.5% yield, +25% GWP), green manure (+8.5% yield with proportional emissions), organic amendments, transplanting, conventional tillage, and long-duration varieties with continuous flooding (Fig. 7). These operate through substrate management and extend anaerobic exposure; organic additions improve fertility but fuel CH₄/N₂O production under flooding conditions. Trade-off magnitude depends on context; substantial yield responses in low-fertility soils may justify emission increases when food production is paramount, whereas marginal gains in organic-rich soils create unfavourable trade-offs.

“Emission-priority” practices reduce GHG intensity at yield costs, relevant when carbon mitigation outweighs production maximisation. These include rainfed systems (lowest GHGI, -12–18% yield), rice-aquatic systems, ultra-short varieties, no-tillage, drill seeding (-12–24% yield), and zero-input systems (Fig. 7). These limit emissions by restricting substrate availability, reducing anaerobic exposure, or employing biological suppression, but constrain productivity through nutrient limitations, water stress, or poor establishment. Such strategies are suitable for high-baseline contexts (organic-rich soils, warm climates) or systems with carbon pricing that justifies yield trade-offs.

“Lose-Lose” practices compromise both productivity and emissions. These include intensive multi-season systems (double/triple-rice: doubled CH₄, reduced per-crop yields^64,65), mismatched variety durations, continuous flooding, poorly-timed drainage, and wet direct seeding (Fig. 7). On-field straw burning has unacceptable impacts, including the destruction of soil organisms, the release of carcinogens, and the loss of nutrients⁶⁶. These practices fail by compounding emission drivers without yield justification or imposing constraints that compromise productivity and environmental quality. Their prevalence reflects a historical tendency toward short-term thinking over sustainability. Transitioning away represents low-hanging fruit where “Win-Win” alternatives exist.

Practice combinations may amplify or constrain effectiveness, though our analysis controlled for co-occurring practices rather than explicitly modelling interactions. Mechanistic reasoning suggests compatible pathways may compound benefits (DDS + AWD extending aerobic periods), while conflicting pathways create antagonisms (AWD’s aerobic phases amplifying N₂O from straw decomposition). Particularly problematic combinations include straw incorporation and continuous flooding, which create massive CH₄ release as abundant substrate meets prolonged anaerobic conditions, or green manure and intensive multi-season systems, where organic amendments fuel cumulative emissions across multiple crops. Combining multiple substrate additions (straw, manure, and green manure) may saturate the methanogenic capacity, creating exponential rather than additive increases in emissions. Conversely, rice-upland rotations exemplify system-wide synergies that exceed the sum of individual practices, simultaneously disrupting pest cycles and preventing substrate accumulation. However, these hypotheses require validation through a factorial experiment.

These insights translate into region-specific priorities. South/Southeast Asian systems (Mekong, Ganges deltas) represent high-baseline contexts (167 kg CH₄/ha, 1.23–1.35 kg/kg GHGI), where emissions are elevated while yields remain comparable to those in low-baseline systems. Water management interventions target prolonged anaerobic conditions from monsoon waterlogging and abundant substrate. Northeast Asian systems exemplify efficient intensive production: temperate climates and fine-textured soils create low baseline emissions (0.65 kg/kg) while advanced management delivers high yields. Mediterranean/Central Asian systems achieve remarkable efficiency (0.34 kg/kg GHGI, one-third of continental systems) through natural water limitation, paired with drought-tolerant varieties. Intensive double- or triple-rice regions require system diversification through rotations, thereby breaking the cycle of substrate accumulation. Critically, identical practices yield different outcomes by region: water management delivers maximum benefits in warm, wet deltas; integrated practices achieve efficiency in temperate systems; diversification addresses intensification penalties in multi-season areas. Context characteristics, not geographic boundaries, should guide intervention selection (Supplementary Table 1).

Translating evidence into practice requires matching “Win-Win” strategies to farmer capacity and institutional context. Adoption feasibility depends on three interdependent factors: agronomic certainty (proven yield and emission benefits), economic viability (positive returns on investment), and institutional support (infrastructure, markets, and extension). We organise practices into three adoption pathways, reflecting these requirements, from immediate-access options to system-wide transformations that require coordinated intervention.

Pathway A (immediate adoption) offers practices that require no infrastructure and deliver immediate returns. Optimised nitrogen management requires no infrastructure and delivers immediate returns, yielding 32–39% gains in yield that outweigh modest increases in emissions, reducing GHGI emissions by 4–5% (Fig. 4). This approach is accessible across farm scales through existing extension networks, representing a low-hanging fruit opportunity. Medium-duration varieties similarly offer easy switching from long-duration types without system reorganisation, improving GHGI by 1.4–1.9% while maintaining yields (Fig. 2). Rice-upland crop rotations provide multiple benefits, disrupted pest cycles, reduced cumulative emissions, and maintained system productivity, where market prices for rotation crops (wheat, legumes) are competitive with rice, creating economic incentives without subsidies⁶⁷.

Pathway B (transitional adoption) encompasses practices that require infrastructure but offer strong agronomic benefits, justifying a staged implementation. Water management transitions can progress: continuous flooding → mid-season drainage → intermittent irrigation → full AWD. Each step demonstrates manageable changes, building farmer confidence and institutional capacity (equipment rental, extension expertise) before advancing⁶⁸. Similarly, residue management transitions, from full straw incorporation → partial removal (leaving stubble) → complete removal with cover crops or composting, balance emission reduction with soil organic matter concerns^69,70, avoiding abrupt system disruptions while moving toward optimal practices.

Pathway C (system transformation) addresses intensive multi-season systems (double/triple rice) that generate “Lose-Lose” outcomes but persist due to policy incentives that reward annual production volume rather than efficiency^71,72. Transformation requires coordinated changes, including policy reforms that realign subsidies toward per-area productivity, the development of markets and infrastructure for rotation crops, and farmer retraining for diversified systems. These represent decade-scale transitions needing governmental commitment. Similarly, shifting from wet direct seeding (dominant in Southeast Asia) to DDS requires not just farmer behaviour change but water infrastructure development, making it infeasible without public investment^73,74.

Each pathway requires matched institutional support. Pathway A practices require strengthened extension services that deliver agronomic knowledge through farmer field schools and demonstration networks⁸, ensuring evidence-based guidance reaches smallholders. Pathway B practices require infrastructure development (water control systems, mechanisation) and risk mitigation (subsidies for equipment, crop insurance during transition periods), with input subsidy reforms redirecting support from flood irrigation toward water-saving technologies⁷⁵. Pathway C transformations demand policy reform aligning incentives with sustainability goals, particularly carbon pricing mechanisms that monetise emission reductions⁷⁶, and regulatory approaches like residue burning bans coupled with viable alternatives (straw markets, composting facilities) and transition support^77,78. Coordinating stakeholder engagement, farmers, extension services, policymakers, and industry through market-driven approaches remains essential for sustainable transitions.

While these findings offer actionable insights, several key considerations warrant attention. Our analysis estimated additive effects of individual practices rather than multiplicative interactions. While this provides robust estimates for single interventions, it cannot reveal whether practice combinations perform synergistically or antagonistically, critical information as farmers adopt multiple strategies simultaneously. This limitation reflects the data availability: few studies employ factorial designs that enable interaction assessment. Additionally, conventional/farmer practice nitrogen application rates were rarely reported in source studies, limiting our ability to benchmark optimised management against practical baselines. Geographic representation is severely uneven, with Asian rice systems dominating the literature while Africa and Oceania remain underrepresented despite substantially different production contexts, limiting the generalisability of global estimates to these regions (Supplementary Text 1). Technical confounding affects some comparisons: tillage effects may conflate with water management when land preparation involves flooding; rice-aquaculture systems vary widely in animal types and intensities beyond current data resolution; and data quality concerns persist for triple-rice systems^79,80. Emerging practices such as drone seeding lack sufficient peer-reviewed evidence for inclusion. These constraints limit subgroup analyses and interaction assessment but do not affect primary findings for major practices where evidence is robust across contexts.

We focused on yield-emission trade-offs as core metrics for sustainable intensification, appropriately reflecting the most widely reported outcomes in source literature. However, this scope excludes critical sustainability dimensions. Long-term trajectories, soil organic carbon dynamics, cumulative soil health impacts, and sustained productivity over 5–10+ years remain largely uncharacterised. Broader agroecosystem functions, including soil biodiversity, water quality, and climate resilience, were not systematically assessed due to limited reporting. Moreover, practice-specific environmental risks extend beyond emissions: high-N applications cause soil acidification and heavy metal accumulation⁸¹; intensive systems reduce biodiversity^82,83; AWD alters microbial communities with uncertain nutrient cycling consequences⁸⁴; and straw removal depletes soil organic carbon, though composting offers alternatives for maintaining organic inputs^64,85. These multifunctional considerations demonstrate that “Win-Win” classifications based solely on yield-emission metrics represent necessary but insufficient sustainability criteria. Future research must develop standardised protocols integrating soil health, water quality, and climate resilience alongside productivity and emissions to capture full agroecosystem performance.

Five interconnected needs emerge for future studies: (1) future meta-analyses should assess synergistic or antagonistic effects when implementing multiple practices simultaneously; priority combinations include DDS × AWD (compounding aerobic periods) and AWD × straw incorporation (where drainage timing modulates decomposition); (2) Long-term monitoring networks spanning 5–10+ years across representative sites are essential to track soil organic carbon trajectories, cumulative soil health impacts, and sustained productivity beyond single-season measurements; (3) Multifunctional assessment protocols integrating soil health (biodiversity, organic matter), water quality (nutrient leaching), and climate resilience (yield stability) alongside productivity and emissions would enable comprehensive sustainability evaluation across contexts; (4) Geographic expansion through region-specific field studies in Africa, Oceania, Latin America, and other underrepresented areas is critical to ensure recommendations reflect diverse production contexts and avoid overgeneralisation from Asian systems that dominate current literature (Supplementary Text 1); (5) Nitrogen management interactions (application methods, fertiliser types, co-applied nutrients: phosphorus, potassium, zinc) should be assessed through factorial meta-analytic models against regional conventional rates to provide practice-relevant recommendations for integrated nutrient management; (6) As factorial field experiments accumulate, future meta-analyses should assess emerging technologies, including drone seeding, as sufficient field data become available.

Implementing research findings in real-world settings demands tackling economic and social obstacles that extend beyond agronomic factors. Priority areas include carbon credit mechanisms accessible to smallholders; input subsidy reforms redirecting support toward water-saving technologies; strengthened extension services through farmer field schools; and regulatory approaches (residue burning bans) coupled with viable alternatives⁸⁶. Given that rice-upland rotations emerge as “Win-Win” strategies, future research should extend to rotation system performance, assessing rotation crop outcomes (wheat, maize, legumes) and cascading effects across agronomic, economic, and social dimensions. Context-specific solutions integrating scientific evidence with local knowledge, supported by enabling policies and robust institutions, remain essential for scaling sustainable intensification⁸⁶.

Farming practice selection and framework

We applied the SRP guidelines version 2.1 to select farming practices based on emissions performance indicators, providing a structured and comprehensive approach to evaluating practices across all stages of rice production. The SRP is benchmarked using 12 performance indicators and 41 requirements derived from three sustainability dimensions – social, environmental, and economic⁸⁷. These indicators allow us to assess sustainability improvements from improved farming practices and identify potential areas for further enhancement⁸⁷. Building on this, we developed a comprehensive framework that includes eight selected farming practices, including Farming Systems (FaS), Growing Duration (GrD), Tillage Method (TiM), Planting Method (PlM), Water Regime (WaR), Nitrogen Application Rate (NiA), Compost and Manure Application (CoMa), and Residue Management (ReM) (Supplementary Fig. 6). These practices correspond to seven key requirements, encompassing all stages of rice production, such as land conversion and biodiversity, water management, organic fertiliser use, inorganic fertiliser use, rice stubble management, and rice straw management⁸⁷.

Data establishment

The study database was established using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses method (PRISMA)⁸⁸. To enhance the efficiency and accuracy of this process, we employed the Covidence application (app.covidence.org), which streamlined the screening, extraction, and management of data, while maintaining full compliance with PRISMA’s selection criteria.

Five databases, including GeoRef (www.engineeringvillage.com), Web of Science (www.webofknowledge.com), Scopus (www.scopus.com), Science Direct (https://www.sciencedirect.com), EBSCOhost (www.ebscohost.com) were used to search for relevant peer-reviewed papers. The last search was done on July 14, 2024, with the general equation: “(rice OR padd*) AND (rotation* OR intercropping OR “farming system” OR “water management” OR irrigat* OR drainage OR fertili* OR manure OR stubble OR straw) AND (“environment* impact*“ OR “carbon footprint” OR emission* OR “global warming” OR ghg)”. A total of 15,991 potentially relevant literature was retrieved (Supplementary Fig. 7).

Starting with 15,991 relevant literature retrieved, 504 papers were extracted based on the inclusion outlined criteria: (1) studies must be peer-reviewed and published in English; (2) at least one farming practice related to our framework must be included in the experimental design; (3) cumulative Greenhouse Gas (GHG) emissions must be measured over the entire rice season using validated tools and methods; (4) both Grain Yield (GY) and GHG emissions (seasonal cumulative) must be determined simultaneously, or calculable based on author-provided formulas; (5) data units must be consistent and convertible to a per-hectare basis; (6) the location and experimental conditions must be clearly stated in the study.

A total of 5322 experimental results were extracted from selected papers. For each study, we extracted three response variables: methane (CH₄), nitrous oxide (N₂O) and GY, replications, and precision indicators of the effect sizes (standard deviation (SD) or standard error (SE)). Since the carbon dioxide (CO₂) absorbed by rice photosynthesis is higher than that produced by respiration, direct CO₂ emissions from rice fields are excluded from GHG emissions⁸⁹. In cases where data were only presented in figures, values were extracted using the Get Data Graph Digitizer (http://getdatagraphdigitizer.com/). Unidentified Error bars were assumed to represent SE, given that the error bar typically appeared in SE format^79,90. In cases where the same data were used in multiple publications, we recorded it only once. Each experiment was recorded independently if a publication reported seasonal, year, or repetition differences. We defaulted to using three replicates for studies that did not report the number of experiment replicates. Emails were sent to the corresponding author to collect the original data that could not extracted by the above method. The no-response papers were removed from the database.

In addition, for each independent experiment, the cultivation practices were extracted, including farming systems, planting methods, cultivars (growing duration), water regimes, tillage methods, nutrient management (nitrogen, phosphorus, potassium, compost, manure), and residue management. Our database also incorporated site characteristics, such as experiment location, climate, soil type, pH, season, and GHG emission measurement methods. The description of each cultivation practice is provided in Supplementary Table 2.

All the effect sizes were converted to standard units of kg/ha, limited to the scope of a single experimental season. GY was recorded at 14% moisture; the studies did not report a grain moisture content determined to be 14%. The formula will convert studies using another moisture rate:

Where: GY: Grain yield at Ma moisture rate

GY_b: Grain yield at Mb moisture rate

M_a: Standard moisture rate (14%)

M_b: Reported moisture rate in the study

If standard deviation (SD) was reported, the following equation was used to calculate the standard error (SE):

Where: “n” is the number of replicates.

Studies that did not report a statistical variance were also included by using an arbitrary SD value based on a coefficient of variation (CV) by formula³¹:

Global Warming Potentials (GWP) in the 100-year period – the standard mid-term timeframe used to assess the relative impact of different GHG on global warming – was calculated based on the IPCC⁹¹ formula:

SE of GWP calculated based on the SE of CH₄ and SE of N₂O by formula:

Greenhouse Gas Emission Intensity (GHGI) or Yield-Scale Greenhouse Gas Emission was calculated by formula:

SE of GHGI calculated based on the SE of GWP and SE of GY by formula:

The key independent variables of each farming practice were selected based on their prevalence in research datasets and their relevance and frequency in real-life practices (Supplementary Table 2). We also categorised variables related to the experimental conditions, including climate (ClC), soil pH (PhT), soil types (SoC), and season (Sea). The Köppen climate classification using the Köppen-Geiger climate classification layer integrated with Google Earth’s World Map was used to classify ClC based on the longitude and latitude or location of the study⁹². PhT was classified into levels ranging from Very Acidic (<4.5), Acidic (4.5–5.5), Slightly Acidic (5.6–6.5), Neutral (6.6–7.3), Slightly Alkaline (7.4–8.0), to Alkaline (>8.0) based on soil pH reporting on the study⁹³. Based on the characteristics and soil types reported for each experiment, we classified them into Fine-textured soils, mixed soils, and organic-rich soils to reflect the soil’s drainage capacity and organic content, significantly impacting GHG emissions⁹⁴. Although the reported seasons have different names across various countries, the most common factor used to classify them is rainfall, which includes the ‘wet season’ and ‘dry season’. However, ‘early season’ and ‘late season’ were also added to our classification list, as they frequently appear in the database, even though there is no clear difference in rainfall between these two seasons and the wet season in the main growing regions, primarily in China.

Bayesian hierarchical meta-analysis

Due to the high heterogeneity in the global dataset, a Bayesian Hierarchical Meta-Analysis was employed to account for the substantial differences between studies. This method provides a flexible framework to model study-level variation, helping to mitigate the influence of outlier studies and improve the model’s predictive capacity⁹⁵. The model uses Markov chain Monte Carlo (MCMC) methods to sample from the posterior distribution. MCMC allows the model to approximate the posterior distribution when explicit calculation is not feasible while also managing the complexity of the data through iterative sampling, ensuring accurate parameter estimation⁹⁶. This approach incorporates prior information to efficiently model heterogeneity, handling outliers by adjusting the model parameters to account for peripheral variables⁹⁷. Since the outcome measures across experiments were consistent throughout the dataset, each experimental result was treated as an independent effect size in our meta-analysis⁹⁸. All data analysis was conducted in RStudio (version 4.4.0), using the “brms” and “rstan” packages, with ggplot2 for visualisation^{99,100,101,102}.

The form of the models was

Estimate| se_Estimate ~ Predictor_Variable + (Other_farming_practice_variables) + (Cultivation_condition_variables) + (1 | Random Variables)

In which:

Estimate| SE_Estimate: refers to the estimate of the dependent variable (CH₄, N₂O, GY, GWP or GHGI) and the corresponding SE for each experiment. The standard error plays a role in adjusting the weight of each experiment based on the reliability of the results, thereby improving the model’s accuracy in estimation^103,104.

Predictor_variable: is the main predictor variable influencing the dependent variable.

Farming_practice_variables and Cultivation_condition_variables:

are auxiliary variables, including other farming practices and cultivation condition variables that act as controls for external factors that could affect the outcome. These help to reduce unwanted variation and provide more accurate estimates of the main predictor variable’s effects¹⁰⁵.

1 | Random Variable: helps the model control for variation within and between studies (StudyID) and clusters of studies by country (Site), ensuring that results are not biased by unobserved differences between experiments^104,106.

Based on the overall model, 40 different models were independently developed (see Bayesian Model Optimisation and Sensitivity Analysis) to assess the impact of farming practices on five key dependent variables: CH₄, N₂O, global warming potential (GWP), GY and greenhouse gas intensity (GHGI). We applied a logarithmic transformation to the effect sizes to stabilise the distribution of the dependent variables and minimise errors from outliers¹⁰⁷. We did not include an intercept in the model to avoid the influence of any fixed average values, providing an unbiased and comprehensive view of each category¹⁰⁸.

Subsequently, four models were created to evaluate the correlation between GY and GWP, GHGI, CH₄ and N₂O (Supplementary Table 3). In these four models, we did not apply a logarithmic transformation to the independent variables, as the model estimated the correlation between two continuous variables, and the large sample size ensured that the estimates remained undistorted and stable during the analysis¹⁰⁹. At the same time, we included an intercept to define baseline parameters in the model, which provided a standard reference point for evaluating changes in GHG emissions globally. This ensured that the predictions were not affected by random factors from individual studies^109,110.

We developed a protocol for integrating model optimisation and the sensitivity assessment of variables in the model. First, we conducted exploratory data analysis by performing descriptive statistical checks and examining the distribution of each dependent variable to identify potential priors and family types for inclusion in the sensitivity assessment and model optimisation process (Supplementary Fig. 9). Simultaneously, we carried out a Variable Contribution Analysis to determine the potential impact of various farming practices on the dependent variables (Supplementary Fig. 10). Next, seven key steps in the sensitivity analysis and model-building process were executed, with 15 independent models tested. Leave-one-out (LOO) cross-validation was used to iteratively compare intermediary models, focusing on the predictive performance of each model, with evaluations based on the elpd_loo (Expected Log Predictive Density), p_loo (Effective Number of Parameters), and looic (LOO Information Criterion) metrics¹¹¹.

Step 1: Simple model setup: A simple model was established with only the dependent and predictor variables (Estimate| se_Estimate ~ Predictor_variable) as a foundation for sensitivity assessment and model optimisation.

Step 2: Prior Identification: To determine the appropriate priors for the model parameters, three priors, including non-informative priors, Student-t with three degrees of freedom and a wide scale, and narrow priors were evaluated (Supplementary Table 3). Wide priors were used to avoid overly strict initial assumptions, thereby reducing the risk of the model being influenced by overly narrow priors⁹⁷. However, narrow priors were more suitable when reliable observational data were available, allowing for tighter model adjustment, minimising overfitting, and reducing unwanted variation¹¹².

Step 3: Family Identification: Using the identified priors in step 1, we tested the sensitivity and predictive capability of the model with two families, Gaussian and Student. The Gaussian family was used for normally distributed data with fewer outliers, while Student’s t-distribution was chosen for data with more outliers or non-normal distribution, providing better control over the influence of extreme values¹¹³.

After identifying the appropriate priors and family, we tested the effects of farming conditions, farming practices, and random variables.

Step 4: Identification of Farming Condition (auxiliary) Variables: We added environmental variables such as climate type (ClC), soil types (PhT), season (Sea), and soil conditions (SoC) to identify the most important farming condition variables for inclusion in the overall model.

Step 5: Identification of Farming Practices (auxiliary) Variables: Similar to step 4, we conducted a sensitivity analysis with other farming practice variables (in addition to the main predictor variable) identified through the variable contribution analysis (Supplementary Fig. 9).

To avoid overfitting and reduce complexity, we limited the final model to include a maximum of two farming condition variables and two additional farming practice variables.

Step 6: Identification of Random Effect: We tested different random effect structures, including StudyID, Site/StudyID, and Site + StudyID, to evaluate their suitability for improving the predictive performance of the models.

Step 7: Final Model Establishment: We combined the results from the six previous steps to build the overall model. During model optimisation, we gradually increased the number of iterations, warmup, chains, adapt_delta, and max_treedepth to ensure that the model’s MCMC (Markov Chain Monte Carlo) chains had no divergent transitions and achieved acceptable levels for Rhat and Effective Sample Size (Supplementary Table 4).

All models were evaluated by plotting the posterior samples to confirm chain convergence and check for divergent transitions (Supplementary Fig. 8). We examined the R-hat values, ensuring they were below 1.1, and checked the Effective Sample Size (ESS), with a threshold of >400 considered acceptable and >1000 recommended for robustness^99,114,115. Pareto k diagnostic values were estimated using the “loo” package in R¹¹⁶. To manage observations with high Pareto k values, we employed the loo function with the moment_match = TRUE option, which automatically applies moment-matching adjustments. This approach helps maintain acceptable levels of Pareto k (0.7–1: less than 20% of observations, and >1: less than 5% of observations)¹¹¹. In our results, the highest Rhat, lowest Bulk_ESS, lowest Tail_ESS, and the proportion of observations with unacceptable Pareto k values were 1.01, 961, 1111, and 0.60, respectively, across all models (Supplementary Table 4). Most selected models (84%) of models had R² values greater than 0.5 (ranging from 0.51 to 0.78), 14% of models had R² values between 0.31 and 0.42, and only one model had an R² below 0.2. In summary, the results indicate that the models are of high quality regarding predictive capacity, reliability, and the ability to explain data variability, while effectively managing heterogeneity and differences between studies.