Eighteen rice-producing countries were selected for our analysis (Supplementary Table 1). Those countries account for 88 and 86% of global rice production and harvested rice area2, respectively (FAOSTAT, 2015–2017). We followed two steps to select the dominant cropping systems in each country. Within each country, our study focused on the main rice-producing area(s) (Supplementary Tables 2 and 3). For example, in the case of Brazil, we selected the southern and northern regions, which together account for nearly all rice production in this country. In the case of Vietnam, we selected the Mekong Delta region, which accounts for nearly 60% of national rice production57. While we tried to cover all major rice cropping systems in each country, this was not possible in the case of rainfed lowland rice cropping systems in northeastern Thailand and eastern India because of lack of reliable estimates of yield potential and access to farmer yield and management data. Once the main rice-producing region(s) in each country was (were) identified, we then determined the dominant rice cropping system(s) for each of them (Supplementary Table 3). We note that “cropping system” refers to a unique combination of a number of rice crops planted on the same piece of land within a 12-month period (and their temporal arrangement), water regime (rainfed or irrigated), and ecosystem (upland or lowland) (Supplementary Fig. 1 and Supplementary Table 2). In our study, rice cropping systems are single-, double-, or triple-season rice; none of the cropping systems are ratoon rice. Following the previous examples, two cropping systems were selected for Brazil (rainfed upland single rice and lowland irrigated single rice in the northern and southern regions, respectively) and two systems (double and triple) were selected for the Mekong Delta region in Vietnam. These systems account for nearly all rice harvested areas in these regions. We distinguished between rice-based cropping systems sowing hybrid versus inbred cultivars in the southern USA. Across the 18 countries, this study included a total of 32 rice cropping systems, which, in turn, covered 51% of the global rice harvested area (Supplementary Tables 1 and 3). Note that the area coverage reported here corresponds to that accounted by 32 cropping systems (and not by the countries where the cropping systems were located). These systems portrayed a wide range of biophysical and socio-economic backgrounds (Supplementary Figs. 1 and 2 and Supplementary Tables 1 and 2), leading to average rice yields ranging from 2–10.4 Mg ha−1 (Supplementary Fig. 3). For data analysis purposes, rice cropping systems were classified into tropical and non-tropical9,58,59 and also based upon water regime and crop season.
Agronomic information was collected via structure questionnaires completed by agricultural specialists in each country or region (Supplementary Table 6). The collected data included field size, tillage method, crop establishment method, degree of mechanization for each field operation, seeding rate, crop establishment, and harvest dates, nutrient fertilizer rates, manure type, and rate, pesticides (number of applications, products, and rates), irrigation amount (in irrigated systems), energy source for irrigation pumping, labor input, and straw management (Supplementary Tables 4 and 5). Average values for each cropping system reported by country experts were retrieved from survey data available from previous projects (Supplementary Table 7). Rice grain yield was reported at a standard moisture content of 140 g H2O kg−1 grain, separately for each crop cycle, using data from, at least, three recent cropping seasons in each cropping system. In the case of irrigated rice cropping system in Nigeria and Mali, data were only available for one crop cycle in double-season rice. In this case, we assumed management and actual yield to be identical in the two crop cycles.
In all cases, and wherever possible, data were cross-validated with other independent datasets (e.g., FAOSTAT, World Bank, IFA, and published journal papers), which gives confidence about the representativeness and accuracy of the survey data. For example, we estimated area-weighted national yield according to actual yield provided for each cropping system and annual rice harvested area in each system for each of the 18 countries. Comparison of these yields against those reported by FAOSTAT2 showed a strong association and agreement between data sources (Supplementary Fig. 10). We also cross-validated actual yield, N fertilizer, labor, and irrigation from our database with those reported by previous studies (published after the year 2000) based on on-farm data collected in ten selected countries. Due to the lack of on-farm data on irrigation, we used published data collected from experiments that follow typical farmer irrigation practices. In the case of irrigation, our cross-validation differentiated between crop seasons (wet versus dry) in the case of irrigated double-season rice cropping systems. In all cases, average yield, N fertilizer, labor, and irrigation from our database fell within (or very close) the range of values reported in previously published studies for those same cropping systems (Supplementary Table 8). Measured daily weather data, including daily solar radiation, minimum and maximum temperatures, and precipitation, were derived from representative weather stations in each region (Supplementary Fig. 2 and Supplementary Table 9). Data on per-capita gross domestic product (GDP) during 2015–2017 were retrieved for each country to explore relationships between yield gap and economic development60 (Supplementary Fig. 9 and Supplementary Table 1).
Estimation of yield gaps
The yield gap is defined as the difference between yield potential and average farmer yield. Estimates of yield potential for irrigated rice or water-limited yield potential for rainfed rice were adopted from Global Yield Gap Atlas (GYGA)61 (Supplementary Table 7). Yield potential simulation in GYGA was performed using crop growth and development model ORYZA2000 or ORYZA (v3) (except for APSIM in the case of India) and based on actual data on crop management, soil data, measured daily weather data, and representative rice varieties planted in each region (see details for yield potential simulation in Supplementary Information Text Section 1). Data on yield potential were not available for Australia (AUIS) in GYGA; hence, we used estimates of yield potential from Lacy et al.62. Yield potential (or water-limited yield potential for rainfed rice) and average yields were computed separately for each rice crop in each rice cropping system (Supplementary Fig. 3). The coefficient of variation (CV) of yield potential (or water-limited yield potential) was estimated for each cropping system (Supplementary Fig. 4). In this study, average rice yield was expressed as percentage of the yield potential (or water-limited yield potential for rainfed rice) for each cropping system (Fig. 1 and Supplementary Fig. 5). In those cropping systems where more than one rice crop is grown within a 12-month period, we estimated yield potential and average yield on both per-crop and annual basis by averaging and summing up the estimates for each rice crop, respectively. In the case of per-crop averages, for those cropping systems in which the harvested rice area changed between crop cycles, we weighted the values for each cycles based on the associated harvested rice area. However, for simplicity, the main text reports only the values on a per-crop basis; annual estimates are provided in the Supplementary Information. Normalizing average yield by the yield potential at each site provides a direct comparison of yield gap closure across systems with diverse biophysical backgrounds (e.g., variation in solar radiation, temperature, and water supply). Without this normalization, one might make biased assessment in relation to the available room for improving yield. For example, an actual yield of 8 Mg ha−1 is equivalent to 80% of yield potential in the cropping system of central China, whereas a yield of 8 Mg ha−1 achieved by irrigated rice farmers in Brazil only represents 55% of yield potential (Supplementary Fig. 3).
Quantifying resource-use efficiency
We assessed the performance of rice production by calculating the following metrics: global warming potential (GWP), fossil-fuel energy inputs, water supply (irrigation plus in-season precipitation), number of pesticide applications, nitrogen (N) balance, and labor input, each expressed on an area and yield-scaled basis (Figs. 2, 3 and 4 and Supplementary Figs. 6, 7 and 11). We estimated metrics on both per-crop and annual basis and report the values on a per-crop basis in the main text while the annual estimates are provided in the Supplementary Information. In the case of GWP, it includes CO2, CH4, and N2O emissions (expressed as CO2-eq) from (i) production, packaging, and transportation of agricultural inputs (seed, fertilizer, pesticides, machinery, etc.), (ii) fossil-fuel energy directly used for farm operations (including irrigation pumping), and (iii) CH4 and N2O emission during rice cultivation63. Emissions from agricultural inputs were calculated on application rates and associated GHG emissions factors (see details in Supplementary Information Text Section 2, Supplementary Table 10). In the case of fossil fuel used for field operations, it was calculated based on the number and type of farm operations and associated fuel requirements (Supplementary Table 11). Total N2O emissions were calculated as the sum of direct and indirect N2O emissions. A previous meta-analysis including rice showed that direct soil N2O emissions can be estimated from the magnitude of N-surplus, which was calculated as applied N inputs minus accumulated N in aboveground biomass at physiological maturity21. Therefore, direct soil N2O emissions for a given rice crop cycle were estimated following van Groenigen et al. N-balance approach21. Indirect N2O emissions were estimated based on the Intergovernmental Panel on Climate Change (IPCC) methodology64, assuming indirect N2O emissions represent 20% of direct N2O emissions. The CH4 emissions from rice paddy field were calculated following IPCC65. Following this approach, CH4 emissions are estimated considering the duration of the rice cultivation period, water regime during the cultivation period and during the pre-season before the cultivation period, and type and amount of organic amendment applied (e.g., straw, manure, compost) based on a baseline emission factor. We assumed no net change in soil carbon stocks as soil organic matter is typically at steady state in lowland rice66. We did not attempt to estimate changes on soil C in the upland rice system in Brazil. All emissions were converted to CO2-eq, with GWP for CH4 set at 25 relatives to CO2 and 298 for N2O on a per mass basis over a 100-year time horizon67. For each rice crop cycle in each of the 32 rice systems, GWP was calculated as the sum of CO2, CH4, and N2O emissions expressed as CO2-eq. (Details on N2O and CH4 emissions estimates and GWP calculations are provided in Supplementary Information Text Section 2).
Calculation of energy inputs was similar to that of GWP and was based on the reported rates of agricultural inputs and field operations and associated embodied energy (see details for energy input estimates in Supplementary Information Text Section 2, Supplementary Table 12). Human labor was also included in the calculation of energy inputs. There was a strong positive relationship between energy input and GWP on both per-crop (r = 0.81; p < 0.01) and annual basis (r = 0.92; p < 0.01), so we only showed results on GWP in the main text. Results on energy input and net energy yield (the difference between energy output and input) on a per-crop or annual basis can be found in Supplementary Fig. 11.
The N balance was calculated as the difference between N input from synthetic N fertilizer, manure, and biological N fixation minus N removal with the harvested grain (and straw if it was burned or removed from the field) following Dobermann and Witt68 (see details for N balance estimates in Supplementary Information Text Section 3). The N input and N removal were estimated for each rice crop cycle. The N input via manure was calculated based on the amount and source of manure and N concentration. An average input of N from biological N fixation of 30 kg N ha−1 crop−1 was assumed for lowland rice systems69, while biological N fixation in upland rice was assumed to represent 10% of that in lowland rice70. Grain N removal was calculated based on average grain yield and rice grain N concentration. The N removal with straw was estimated assuming a typical percentage of straw remaining in the field and percentage of N lost from the crop residue in different straw managements (Supplementary Table 13). In our N balance calculation, we assumed N losses via lixiviation and denitrification to be similar to the amount of N inputs via irrigation water and atmospheric deposition52. A threshold of N balance of 75 kg N ha−1 was used in this study to estimate N excess (and potentially large reactive N losses), as potential N losses increase substantially when N balance exceeds 75 kg N ha−1 as reported by previous studies32,33,34.
We estimated the amount of active ingredient and environmental impact quotient (EIQ) of pesticides including insecticide, herbicide, and fungicide applied per hectare per crop following Kovach et al. environmental risk assessment methodology71 (see details for toxicity estimation in Supplementary Information Text Section 4). The two metrics showed a significant and positive relationship on a per-crop basis (Supplementary Information Text Section 4, r = 0.96; p < 0.01), and EIQ was also significantly and positively correlated with the number of pesticide applications (r = 0.87; p < 0.01). Given the uncertainty in EIQ estimates associated with sketchy reporting of products and application rate of pesticides, and considerable variation in the reliability of such data among countries or regions, the number of pesticide applications is used to evaluate environmental risk among cropping systems.
Labor requirement is a key driver explaining changes in rice area, systems, and profit63,72. Our labor data included labor involved in land preparation, seed preparation, crop establishment, water irrigation, fertilization, pesticide application, weeding, harvesting, threshing, and drying (Supplementary Table 4). Given the intrinsic relationships among labor input, mechanization level, establishment method (direct seeding versus transplanting), and field size72, we characterized each rice cropping system in terms of these parameters and expressed labor input on both area and yield-scaled basis (see details for labor input and degree of mechanization in Supplementary Information Text Section 5, Fig. 4 and Supplementary Fig. 6; Supplementary Table 4).
Estimation of overall performance index
We computed a semi-quantitative index to quantify the performance of each cropping system in relation to six metrics, including the yield gap (as percentage of yield potential) and yield-scaled metrics including GWP, water supply, number of pesticide applications, N balance, and labor input (Fig. 5 and Supplementary Fig. 8). For each metric, the score was calculated by normalizing the data relative to the maximum value among all 32 cropping systems. An exception was the yield-scaled N balance, which was expressed as an absolute deviation from 8 kg N Mg−1 grain. This value corresponds to the average yield-scaled N balance estimated for Australia and California, which we assumed here to be a reasonable target to achieve the dual goal of minimizing the N balance and closing the yield gap, while avoid soil N mining (Fig. 3). Finally, we estimated an overall performance index for each rice cropping system by averaging the individual scores associated with the six metrics. Four out of the six metrics are related with yield-scaled metrics (GWP, water supply, number of pesticide applications, and N balance), one with yield gaps, and another one with labor. To avoid biases, we weighted each individual score so that yield gap, resource-use efficiency, and labor will have a similar impact on the computation of the overall performance index. Lower (higher) overall index indicates better (worse) overall performance (Fig. 5). Following previous assessments of sustainability in cropping systems25,55,73,74,75, radar charts were used to show the performance of each cropping system in this study in relation to yield-scaled metrics, yield gaps, and labor. Separate analyses were performed based upon climate background (non-tropical and tropical) and also by crop season (wet and dry) in the case of tropical environments (Fig. 5 and Supplementary Fig. 8). Finally, Pearson’s correlation coefficients were calculated to investigate associations between resource-use efficiency and yield gaps (Supplementary Table 14). Statistix 8 and SigmaPlot 12.5 were used for statistical analysis.
To illustrate the potential of our assessment to serve as basis to prioritize agricultural R&D, we explored a scenario in which there is an explicit effort to (i) increase average yield from current level to 75% of yield potential in cropping systems with relatively large yield gaps (defined here as average yield < 60% of yield potential), and (ii) reduce current N balance to 75 kg N ha−1 crop−1 in cropping systems that currently exhibit a large N balance (defined here as N balance > 100 kg N ha−1 crop−1). Selection of this N balance target (i.e., 75 kg N ha−1 crop−1) was based on data from the literature showing that large N losses occur when N balance exceeds that value32,33,34. Following these criteria, we selected a total of 19 cropping systems with large gaps (mostly in Sub-Saharan Africa and Asia) and eight cropping systems with large N balance (mostly in China and South Asia). Following previous studies, the excess N was calculated as the amount of N balance exceeding 75 kg N per ha76. We also explored a second scenario in which, for the same set of 19 systems, average yield increases from current level to 75% of yield potential without changes in current yield-scaled N balance (Table 1). Because the goal of these two scenarios was to understand how to produce more while reducing the environmental impact on existing global rice area, we calculated the potential extra rice production and changes in excess N across the 32 cropping systems considering current rice harvested area, cropping intensity, and proportion of irrigated area (Table 1).
Uncertainty and limitations
We acknowledge the uncertainty related with estimation of yield potential and collection of actual yield and management data. In all cases, we used estimates of yield potential derived using well-calibrated models and best available sources of weather, soil, and management data. To the extent that it was possible, we cross-validated estimates of yield potential with measured yield data collected from well-managed crops that grew without nutrient limitation and without yield reductions due to biotic stresses. In the case of survey data, there is always uncertainty in relation to the representativeness of the regions and years included for the analysis. The analyses presented herein focused on the most intensively cropped area of each cropping system using data from at least three cropping seasons for each area. We note, however, that we could not include drought-prone rainfed lowland rice cropping systems in northeastern Thailand and eastern India in our analysis due to lack of robust estimates of yield potential and data on inputs and management. Hence, our findings for rainfed lowland rice apply to those systems in Sub-Saharan Africa and Indonesia. Future work should include these cropping systems. While variation still exists within each of cropping system in terms of crop sequence configuration, management, and inputs, we also note that some level of spatial aggregation was needed for the purpose of the cross-system comparison presented in this paper and also to be effective at orienting agricultural R&D at the national and regional levels. Similarly, to make a fair comparison of the cropping systems and for interpretation of the results, we have to categorize cropping systems according to water regime (irrigated versus rainfed), climate background (tropical and non-tropical), cropping intensity (single, double, and triple), crop season (dry and wet), and, in the case of labor, establishment method (transplanting versus direct seeding). Our categorization can be further improved, for example, by adding other biophysical and socio-economic factors that can help interpret the results and design interventions. Likewise, estimation of GWP required a number of assumptions in relation to GHG emissions; in those cases, we relied on the most recent literature to derive appropriate emission factors. Overall, our assessment should be considered as an initial step, which could be further refined as more spatially granular data on yield, inputs, and agronomic management, and methods to estimate GHG emissions, become available in the future. However, we do not expect these sources of uncertainty and limitations to affect the conclusions from this study. Detailed description of data sources and estimations of yield gaps, resource-use efficiency, and labor inputs can be found in Supplementary Information.
Further information on research design is available in the Nature Research Reporting Summary linked to this article.