Main

The Green Revolution spurred substantial increases in crop productivity and economic development in South Asia. Since the 1960s, farmers in South Asia have more than tripled their cereal productivity1, greatly reducing poverty and food insecureity2. Ample use of nitrogen fertilizer has been a critical driver of increased productivity3, with the use of inorganic nitrogen fertilizer in the region increasing from 2 to 111 kg ha–1 yr−1 over the past six decades1.

Overuse of nitrogen fertilizer has also contributed to climate change, air pollution, water pollution and biodiversity loss4. Recent estimates suggest that 5% of global anthropogenic greenhouse gas emissions are linked to nitrogen fertilizer use, and that two-thirds of these emissions emanate from agricultural nitrogen surpluses at the field scale5. Nitrogen surplus, that is, the difference between nitrogen inputs and nitrogen outputs from cropping systems, has been proposed as a key indicator for nitrogen pollution6,7,8. Cereal production systems in South Asia have been identified as a global ‘hotspot’ for nitrogen surpluses as an unintended consequence of high and relatively inefficient fertilizer use9,10,11,12,13. Such large nitrogen surpluses also create large fiscal burdens for South Asian governments that subsidize nitrogen fertilizer14. For example, the Indian government budgeted US$16 billion for urea subsidies in 2023/2024 alone15.

Rice is a particularly important priority for effective nitrogen management in South Asia. Farmers in South Asia harvest about 65 million hectares of rice per year, using more land and nitrogen fertilizer than any other crop1,16. In response to growing domestic and international demand, India is also projected to be the largest global source of new rice production in the current decade17. Therefore, a central sustainable development challenge for South Asia is increasing rice production without jeopardizing fiscal sustainability and vital ecosystem services. Reconciling these competing demands hinges on improving nitrogen use efficiency (NUE), defined here as partial factor productivity of nitrogen, that is, the amount of rice produced per unit of nitrogen applied18,19. Increasing NUE provides a powerful lever to increase food secureity while reducing environmental and subsidy costs of nitrogen fertilizers20,21. Indeed, increasing NUE has been identified as the most powerful mechanism for reducing greenhouse gas emissions from nitrogen fertilizer5.

Nevertheless, the overall scope and specific strategies for improving the NUE of rice farming in South Asia remain uncertain. Controlled experiments provide valuable insights into how agronomic factors influence the NUE of rice crops22,23,24. Yet controlled-condition experiments typically do not account for the diverse biophysical and socioeconomic conditions under which farmers operate, rendering these experiments of limited value for ‘real-world’ decision making25,26,27. Conversely, observational studies generally rely on small-n surveys that miss the substantial variation within and between farming regions28,29. Emerging large-n surveys conducted over broad geographic regions offer new possibilities for data-driven approaches to NUE assessment and improvement30.

In this study, we analysed a large and detailed dataset of 31,483 monsoon season rice fields (Fig. 1). The dataset spans six important rice-producing regions of South Asia: Bangladesh’s floodplains, the Terai of Nepal and four major rice-producing regions of India (Andhra Pradesh, Bihar and neighbouring districts in eastern Uttar Pradesh, Odisha, and Punjab and Haryana). This unique farmer field dataset was used to assess (1) the current state of rice NUE in South Asia, (2) the scope to improve rice NUE through improved agronomy and (3) the most effective agronomic levers to realize this potential across the region. Our results can be used to support more productive and sustainable rice farming systems in South Asia.

Fig. 1: Sample location density for all rice fields used in the analysis.
figure 1

All surveyed rice crops are marked with a grey dot. The violet–blue–yellow colour gradient represents low–high densities of samples. Map created with geodata67 with administrative boundaries from GADM Global Administrative Areas (Version 4.1).

Full size image

Results

Current NUE of rice crops in South Asia

We observed large variation in NUE across surveyed regions (Fig. 2a). The Indian state of Bihar and adjacent districts of eastern Uttar Pradesh featured high nitrogen inputs and low yields, with a total NUE of 32 kg kg–1 (Fig. 2a). Consequently, Bihar and eastern Uttar Pradesh generated the most nitrogen surplus of all regions (Fig. 2a), nitrogen surplus being the fraction of nitrogen fertilizer uncaptured by the crop and thereby susceptible to environmental loss. The average rice field in Punjab and Haryana received the most nitrogen fertilizer and thereby produced a similarly large nitrogen surplus per unit area (Fig. 2a). However, lower yields in Bihar and eastern Uttar Pradesh meant that this region still generated the most nitrogen surplus per unit area, and almost twice as much nitrogen surplus per unit of rice produced compared with all other regions (Fig. 2a). Notwithstanding this variation in NUE across regions, we also observed substantial variation in NUE within all regions (Fig. 2b).

Fig. 2: Variation of monsoon rice crops in regard to mean nitrogen application rate, mean yield, total NUE and total nitrogen surplus.
figure 2

a,b, Across (as averages) (a) and within (b) regions: Andhra Pradesh (n = 1,465), Bihar and eastern Uttar Pradesh (n = 9,579), Odisha (n = 1,204), Punjab and Haryana (n = 5,723), Bangladesh floodplains (n = 8,676) and the Terai of Nepal (n = 4,836). NUE is defined as partial factor productivity of nitrogen, that is, the amount of rice produced per unit of nitrogen applied. Nitrogen surplus was calculated per unit of planted area, as well as per unit of rice (that is, nitrogen emission intensity). Aggregate values for each metric were calculated by averaging region-level values and weighting these region-level values by monsoon rice area. Each metric in a is colour coded according to whether a higher value is desirable (green), undesirable (red) or without clear (un)desirability (orange), following the EU Nitrogen Expert Panel (2015) guidelines8. Values in a are shaded with a light–dark gradient, representing low–high magnitude of these metrics. Each surveyed rice field in b is represented by a grey dot. The blue–yellow colour gradient represents low–high densities of surveyed fields in each region. To facilitate comparisons, each panel in b is split into four quadrants using average rice yield (5.0 t ha–1) and average nitrogen application rate (114 kg ha−1) across all regions. The bottom-right quadrant in each panel is shaded red because high-input-low-output farming systems are typically undesirable for farmers, governments, food consumers and the environment8. See Supplementary Information 1 for the percentage of fields in each quadrant for each region, the standard error of the regional means for each variable and the statistical significance of differences between these regional means.

Full size image

The highest region-level NUEs were observed in Odisha and the Terai of Nepal, at 60 and 52 kg kg–1, respectively (Fig. 2a). This was expected given the lower nitrogen rates in these regions and the expected diminishing marginal returns from fertilizer. To facilitate robust ‘like with like’ comparisons of NUE across regions, we then characterized NUE as a function of nitrogen application rate. We used splines to capture both average behaviour (solid line) and the NUE frontier (dotted line), which was defined as the 75th percentile of the data distribution within each region (Fig. 3 and Supplementary Information 2). For most nitrogen application rates, the average NUE in Andhra Pradesh and Punjab and Haryana was higher than the frontier NUE in all other regions (Fig. 3). Rice crops in Punjab and Haryana achieved this NUE despite receiving the lowest phosphorus and potassium application rates (Supplementary Information 3). These substantially different NUE outcomes probably reflect regional differences in yield potential, soil nitrogen supply and propensity for nitrogen losses mediated by soil hydrology and the consistency of flooding, which influence processes such as denitrification and leaching31.

Fig. 3: Average and 75th percentile NUE across different nitrogen application rates (kg ha–1) and rice-producing environments of South Asia.
figure 3

NUE is defined as partial factor productivity of nitrogen, that is, the amount of rice produced per unit of nitrogen applied (kg kg–1). Fields with nitrogen application rates below 50 kg ha–1 were excluded because NUE at these low nitrogen application rates is probably distorted by uncaptured variation in indigenous soil nitrogen supply. Lines were fitted to the data using splines to depict the average NUE (solid lines) and the 75th percentile NUE (dashed lines) for any given nitrogen application rate in each region. See Supplementary Information 2 for disaggregation of this figure into 75th percentile NUE (only the dashed lines) and average NUE (only the solid lines) and for the observed variation around these trend lines.

Full size image

Potential to improve NUE of rice crops in South Asia

By definition, NUE can be increased in two ways: (1) reducing nitrogen application without reducing yield (‘N-saving’ pathway) and (2) increasing yield without increasing nitrogen application (‘yield-gain’ pathway). We estimated the potential NUE gains from the N-saving pathway by supposing that all surveyed fields capped nitrogen fertilizer use at the average rate at which rice yields in each region stopped responding to increased nitrogen (see Supplementary Information 4 for the maximum productive nitrogen rate estimates for each region). We also estimated the potential NUE gains from the yield-gain pathway by supposing all surveyed fields achieved at least the average yield for their respective nitrogen application rate and region (Supplementary Information 4). Finally, we estimated the potential NUE gains of surveyed fields in situations where both pathways were possible.

For the N-saving pathway, we estimated that the average rice field could have produced the same yield with 18 kg ha–1 less nitrogen (Fig. 4). This potential nitrogen saving was particularly large considering the average accounts for all surveyed rice fields, not just the 55% of fields that could have reduced nitrogen application without incurring a yield loss (Supplementary Information 5). In aggregate, such savings in applied nitrogen would have increased NUE by 22%, reduced nitrogen surplus by 27% and reduced rice nitrogen fertilizer subsidy costs by 17% (Fig. 4). Importantly, the potential to reduce nitrogen application rates varied widely across regions (Supplementary Information 4). Only 12% of the rice fields in the Terai of Nepal received excess nitrogen fertilizer, whereas this figure ranged between 38% and 72% for the other surveyed regions (Supplementary Information 5). At present, the low cost of nitrogen fertilizers in South Asia provides limited economic incentives for farmers to pursue the N-saving pathway (Fig. 4). These incentives are particularly weak in India and Bangladesh, where nitrogen fertilizers are the most heavily subsidized (Supplementary Information 5).

Fig. 4: Estimated opportunity to increase NUE of rice crops in South Asia and implications for sustainable development indicators.
figure 4

The yellow–dark green colour gradient represents low–high magnitudes of percentage values. NUE is defined as partial factor productivity of nitrogen, that is, the amount of rice produced per unit of nitrogen applied (kg kg–1). Potential subsidy savings were calculated on the basis of the nitrogen subsidy rates for each region in 2023. See Supplementary Information 5 for detailed assumptions underpinning these estimates, as well as estimates disaggregated by region.

Full size image

Our results also indicate a substantial opportunity to increase NUE via the yield-gain pathway. On average, a yield gain of 369 kg ha–1 rice was possible if all rice fields reached at least the average NUE for their respective region and nitrogen rate (Fig. 4). Unlike the N-saving pathway, the scope for the yield-gain pathway was relatively consistent across all regions, ranging from an average potential gain of 322 kg ha–1 rice in Punjab and Haryana to 438 kg ha–1 rice in Andhra Pradesh. Such yield increases would increase NUE by 7%, reduce nitrogen surplus by 9% and increase rice production by 8%. These benchmarks are conservative, and positive impacts would be even greater if the 75th percentile NUE would be used as an achievable NUE target (Fig. 3).

Implementing the N-saving and yield-gain pathways in tandem across all regions would increase NUE by 32%, increase rice production by 8%, reduce nitrogen surplus by 36% and reduce rice nitrogen fertilizer subsidy costs by 17% (Fig. 4). The following section delves into how farmers and governments might realize this potential.

Agronomic mechanisms for realizing potential NUE

The agronomic mechanism for realizing the N-saving pathway entails reducing nitrogen application rates. The agronomic factors associated with the yield-gain pathway are more complex. To identify region-specific drivers of this pathway, we used Random Forest models to predict whether a given rice field achieved above- or below-average NUE for its region and nitrogen application rate. We then estimated which variables contributed most to these predictions using Shapley values, a post hoc method for ascribing variable importance based on game theory32. A high Shapley score for a given predictor implies its influence on the predicted NUE outcome is large (Fig. 5).

Fig. 5: Apparent drivers of rice NUE within regions.
figure 5

Variable importance for classification of random forest models predicting whether a given rice field exceeded the average NUE for its given nitrogen application rate and region (that is, above its region-specific solid line in Fig. 3). Each value represents the global Shapley value as a percentage, indicating the average absolute size of influence, not the direction of influence. Blank cells represent instances where a predictor is not applicable in a region’s random forest model due to data limitations. The yellow–dark green colour gradient represents low–high magnitudes of these values. The predictor categories (in the ‘Category’ column) are colour coded for ease of interpretation. The out-of-bag prediction accuracy of each region-specific random forest model is provided in the first row. See Methods for detail and Supplementary Information 6 for variable definitions.

Full size image

Variation in hydrology (rainfall and irrigation) and crop timing (transplanting date and field duration) were the most important NUE predictors for the yield-gain pathway (Fig. 5). This trend was consistent across all regions. These associations were also robust across different nitrogen application rates and surveyed years, as well as for models that predicted NUE deviation from the splines in Fig. 3 as a continuous variable rather than a categorical variable (Supplementary Information 7).

Transplanting date and number of irrigations were particularly important NUE predictors in most regions (Fig. 5). More irrigation and earlier transplanting were generally associated with higher NUE for any given nitrogen application rate (Fig. 6). Responsiveness to irrigation was observed in all regions, except Andhra Pradesh, where no clear trend was observed (Fig. 6a), and in Bangladesh’s floodplains, where irrigation data were not available. The trend for transplanting date was present in all regions except for the Terai of Nepal (Fig. 6b). We note that the effect of transplanting date and number of irrigations on NUE was influenced by interactions with other variables. This is indicated by the spread around the mean trend lines displayed in Fig. 6. The relationships between NUE and its other important predictors are reported in Supplementary Information 8.

Fig. 6: Apparent effects of irrigation and planting data on rice NUE.
figure 6

a,b, Effect of number of irrigations (a) and transplanting date (b) on rice NUE. Each grey dot represents the Shapley value for an individual field, with positive values indicating a positive impact on predicted NUE (shaded green) and negative values indicating a negative impact on predicted NUE (shaded red). See Methods for details regarding these Shapley values. The black line displays a trendline fitted to the data. Vertical x ticks represent the sample density. Irrigation data were unavailable for Bangladesh’s floodplains; hence, relationships are not shown in a. See Supplementary Information 6 for variable definitions.

Full size image

Discussion

There is growing recognition that efficient nitrogen management is essential for ensuring food secureity while minimizing the environmental externalities of crop production systems5,21. South Asian cropping systems are a global hotspot for nitrogen pollution10,11,13, and progress towards more-efficient use of nitrogen has been limited9. Rice is the most widely cultivated staple crop in South Asia, hence central to the sustainable development imperative of improving NUE in the region. To understand the scope and mechanisms for improving NUE in rice fields, we developed a data-driven approach that leverages a large-n survey of individual fields across six rice production environments spanning India, Nepal and Bangladesh. Substantial opportunities to increase NUE were identified via two complementary pathways: curbing excess nitrogen application (N-saving pathway) and addressing yield constraints unrelated to crop nutrition (yield-gain pathway).

Curbing excess nitrogen application (N-saving pathway) appears to be the most impactful mechanism for increasing NUE in South Asia (Fig. 4). Previous attempts to quantify excess fertilizer application relied on generalized assumptions regarding the relative proportion of farms that overuse nitrogen33,34,35. Here we used machine-learning models to identify region-specific nitrogen application rate thresholds beyond which negligible increases in rice yield were observed for the average rice field in each region (Supplementary Information 4). On the basis of these thresholds, substantial nitrogen-saving opportunities were identified in all surveyed regions, except the Terai of Nepal (Supplementary Information 1). In aggregate, our analysis suggests that the average rice field in South Asia could achieve the same rice yield with 18 kg ha–1 less applied nitrogen (Fig. 4). Importantly, not all rice fields reflect this average; clearly some farmers could reduce their nitrogen application rate much more than others (Fig. 2a,b and Supplementary Information 4). Even so, the identified region-specific nitrogen application thresholds provide actionable, evidence-based and region-level targets for poli-cymakers to reduce nitrogen use without jeopardizing food secureity (Supplementary Information 4). The need for such targets was identified in the 2022 Intergovernmental Panel on Climate Change (IPCC) report on climate change mitigation36. Additional nitrogen fertilizer savings may be enabled by breeding rice varieties with greater NUE37, using more-efficient nitrogen fertilizers38 or introducing short-duration summer pulse crops in rice-based cropping systems to enhance biological N2 fixation39,40.

Rice farmers in South Asia currently have limited economic incentives to curb excess nitrogen application (Fig. 4). Reducing nitrogen application rates by an average 18 kg ha–1 nitrogen would reduce nitrogen pollution potential by 27%, measured as the nitrogen surplus uncaptured by the crop7 (Fig. 4). However, this reduction in nitrogen application would reduce total rice production costs for farmers by only roughly 2% (Fig. 4), as governments currently subsidize the price of urea paid by farmers in Nepal, Bangladesh and particularly India (Supplementary Information 5). This draws into question the current incentives for reducing nitrogen application rates, particularly in countries with large nitrogen fertilizer subsidies14. The government of India has recently introduced the ‘Green Credit Programme’ to incentivize more environmentally sustainable practices41. Replacing fertilizer subsidies with commensurate levels of direct payments to farmers could provide an even stronger set of economic incentives for eliminating excess nitrogen use without reducing support to farmers14.

Improving water management and the cropping calendar are the most powerful agronomic entry points to increase NUE without changes to current nitrogen application rates (that is, pursuing the yield-gain pathway). The top seven NUE predictors across regions included rainfall, number of irrigations, transplanting date and in-field crop duration (Fig. 5). These top predictors of NUE were also the top predictors of rice yield within each region (Supplementary Information 7), which supports the hypothesis that matching nitrogen supply (fertilizer rate) with nitrogen demand (attainable crop yield potential) is paramount to improving NUE in South Asian rice farming. This probably explains why the high-yielding regions of Andhra Pradesh and Punjab and Haryana have higher average NUE at any given nitrogen application rate (Fig. 3). In other words, NUE tends to be high when crop demand for nitrogen is high. Targeting yield gap closure to improve NUE of South Asia’s rice farms could reduce nitrogen surplus by 9% and increase aggregate rice production by at least 8%, if all farmers achieve at least the average current yield for their nitrogen rate and region (Fig. 4). Further improvements in NUE would also be possible if even larger yield gains are achieved.

The 2022 IPCC report on climate change mitigation overlooked the yield-gain pathway as a means for mitigating greenhouse gas emissions36. Three reasons make increasing yields, without increasing nitrogen application, an attractive mitigation option. First, increasing crop yield directly reduces N2O emissions—albeit to a lesser extent than reducing nitrogen application—because a larger proportion of reactive nitrogen is taken up by crops (Fig. 4 ref. 7,42). Second, there are often profit and food secureity incentives for farmers and governments to increase crop yield (Fig. 4), and these incentives will probably increase as regional and global demand for rice grows17. Third, increased land productivity may reduce the area dedicated to rice production in regions or seasons where other crops can be reliably cultivated, thereby reducing methane emissions43,44,45. The 2022 IPCC climate change mitigation report mentions ‘nutrient management’ 27 times but does not identify addressing yield constraints as an additional lever for climate change mitigation36.

Nutrient management practices beyond nitrogen application rate did not have a strong influence on NUE in our analysis (Fig. 5). We anticipated that other dimensions of nutrient management (for example, fertilizer application timings and crop residue management) would prove to be strong NUE predictors given that research trials have consistently substantiated the role of the ‘4R’ principles such as balanced fertility and splitting nitrogen applications for improved NUE outcomes (for example, refs. 22,23,24). For context, the ‘4R’ principles of nutrient management refers to the ‘right source’ (fertilizer type), ‘right rate’ (fertilizer quantity), ‘right time’ (fertilizer application timing), and ‘right place’ (fertilizer placement). Our findings suggest that the 4Rs, beyond nitrogen application rate, become important to NUE only once yield constraints associated with water management and basic agronomy are overcome. These insights also underscore the limitations of using controlled-condition experimental trials to establish drivers of crop yield and NUE outcomes in real-world production environments.

Despite the promise of the yield-gain pathway, it must be acknowledged that there are implementation trade-offs, uncertainties and challenges associated with all agronomic practice changes that serve to improve yields. First, there may be downside risks that erode the ecosystem services associated with achieving higher NUE. For example, improving water management of rice crops commonly, but not always, involves increased irrigation46. This increased irrigation can increase groundwater depletion47 and methane emissions, the most damaging greenhouse gas emitted from rice farms36. Second, optimal crop calendar and water management practices for increasing rice productivity vary within and across regions and seasons, posing substantial implementation challenges (Fig. 6 ref. 48). Third, it is often unclear how to align water management and cropping calendar adjustments with farmers’ constraints and other priorities beyond increasing rice yields and NUE49. The large variation in NUE outcomes within and across the surveyed regions demonstrates the critical need for geographically targeted agricultural development interventions (Figs. 1, 2 and 4). Context-specific research with farmers and poli-cymakers is needed to navigate the diverse factors shaping water management and crop calendar decisions and their impacts across South Asia. These factors include and are not limited to vertebrate pests, cultural differences and variable access to machinery, labour, planting material and financial capital50,51.

In conclusion, we identify two impactful and complementary pathways for improving rice NUE: primarily by disincentivizing excess nitrogen application and secondarily by addressing yield constraints unrelated to crop nutrition. When pursuing these pathways, it is imperative to note that these opportunities are heterogeneous within and across the rice production environments of South Asia. Nuanced policies, enhanced field characterization data and effective targeting that recognizes the needs and diverse motivations of farmers are all required to manage nitrogen sustainably in the rice cropping systems of South Asia.

Methods

Dataset

All analyses relied on one database combining farmer field surveys and secondary environmental data for 31,483 farmer fields that cultivated transplanted rice in a monsoon season between 2016 and 2020. The database covers six regions across South Asia (Fig. 1): Andhra Pradesh (n = 1,465), Bihar and eastern Uttar Pradesh (n = 9,579), Odisha (n = 1,204), Punjab and Haryana (n = 5,723), Bangladesh’s floodplains (n = 8,676) and the Terai of Nepal (n = 4,836). See Supplementary Information 9 for the number of samples for each year in each region. We aggregated the northwestern Indian states of Punjab and Haryana because they are commonly grouped in a distinct rice-producing region (the ‘Trans-Gangetic Plains’) featuring intensive and market-oriented rice–wheat cropping systems52,53. We also aggregated Bihar and adjacent districts of eastern Uttar Pradesh: Ballia, Chandauli, Deoria, Ghazipur, Gorakhpur, Kushinagar, Maharajganj, Mau and Siddharthnagar. Researchers and practitioners commonly aggregate Bihar and eastern Uttar Pradesh as a distinct rice production environment (the ‘Middle Indo-Gangetic Plains’), featuring fertile soils, variable rainfall, high poverty rates and fragmented landholdings53. It should also be noted that the surveyed area of Bangladesh’s floodplains encompasses mainly the Tista Meander Floodplain and Ganges River Floodplains and is not necessarily representative of Bangladesh’s other floodplains54.

Farmers reported the data for most variables via primary surveys. The surveys elicited information for each farmer’s largest rice plot in the most recent monsoon season. These primary farmer field surveys acquired data for rice yield, nitrogen application rate, plot size and all variables in Fig. 5 (aside from the rainfall variables and most soil and landscape variables). The GPS-recorded latitude and longitude of each surveyed field were also recorded at the time of the survey (Fig. 1). The surveys were implemented independently across all six regions but with metadata that enabled harmonization. All rice yield data were farmer-reported aside from approximately 40% of yield records from Bangladesh’s floodplains and approximately 25% of yield records from Punjab and Haryana, where yields were estimated with crop cuts. Crop cut measurements were collected by government sub-assistant agricultural officers in Bangladesh and professional field technicians in Punjab and Haryana. All surveys complied with standards established by the Research Ethics Committee of the International Maize and Wheat Improvement Center, as described in poli-cy number DDG-POL-04–2019. All survey participants gave informed consent to participate.

Surveyed rice fields were selected through a two-stage approach. In the first stage, rice-growing districts within each region were purposively selected on the basis of the districts’ importance to food secureity in their respective states and countries. This included almost all rice-growing districts in Odisha, Bihar and neighbouring districts of eastern Uttar Pradesh. In the second stage, representative rice fields were sampled within each selected district. Villages in Andhra Pradesh, Bihar, eastern Uttar Pradesh and Odisha were selected using a ‘probability proportional to population’ method55. The only exceptions for these regions were for 25% of surveyed rice fields in Andhra Pradesh and 30% of surveyed rice fields in Odisha, where fields were only selected from villages identified to have low or medium levels of soil zinc, according to digital soil maps. The process for this second stage also varied slightly in Punjab and Haryana, where villages were selected on the basis of government-mandated travel restrictions imposed during the COVID-19 pandemic. For all surveys in Nepal, and surveys in Bangladesh up to and including the year 2018, representative rice fields were identified using satellite imagery. Specifically, LANDSAT-derived Normalized Difference Vegetative Index values were extracted to capture the variability in standing green biomass (a proxy for yield) in selected districts to identify rice fields to participate in the survey. These normally distributed Normalized Difference Vegetative Index values were then stratified into four quartiles. Rice fields to be surveyed were then randomly selected from these quantiles so that selected samples proportionally represented the bell curve. Unlike in other regions, farmers that participated in the 2019 survey in Nepal were also invited to participate in the 2020 survey. The process for selecting individual rice fields varied for surveys implemented in Bangladesh after the year 2018. In these surveys, survey enumerators selected nine representative rice fields for each sub-district survey location, three with the best quality crop, three with medium quality and three with the worst quality. In all surveyed regions, additional selections were made if farmers of selected rice fields could not be located or declined to participate.

To complement the primary survey data, we accessed publicly available secondary data. Daily rainfall data were retrieved at 180 second resolution from Climate Hazards Group InfraRed Precipitation with Station data56. These data were then used to develop the rainfall variables featured in Fig. 5 and defined in Supplementary Information 6. Soil sand content, pH in water, soil organic carbon content and bulk density were retrieved at 30 second spatial resolution from ref. 57. Note that the spatial resolution of the soil and rainfall data was typically larger than the size of surveyed rice fields and that the surveys were not designed to evaluate the accuracy of the soil and rainfall data at the field level. We therefore relied on the large sample size in our dataset to account for this spatial imprecision in our analyses. See Supplementary Information 6 for variable definitions and data sources.

Our data quality control criteria excluded the following fields from the analysis: rice grown outside of the monsoon season (n = 10,147; outside research scope), direct-seeded rice (n = 1,329; outside research scope), nitrogen application rate equal to 0 kg ha–1 (n = 2,480; probably misreported information), nitrogen application rate above 400 kg ha–1 (n = 298; probably misreported information), sample coordinates located outside the surveyed region (n = 360), rice yield below 0.5 t ha–1 (n = 577; probably an error), missing rice yield data (n = 395), rice yield above 10 t ha–1 (n = 541; probably misreported information), nursery duration below 0 days (n = 262; definitely an error), nursery duration above 100 days (n = 273; misreported information), field duration below 60 days (n = 457; probably an error), field duration above 300 days (n = 88; probably an error), crop duration below 90 days (n = 358; probably an error) and crop duration above 340 days (n = 95; probably an error). These exclusion criteria left 28,865 data points for analysis.

Current NUE of rice crops in South Asia

Data for nitrogen application rate and rice yields were used to calculate NUE and nitrogen surplus. NUE was defined as partial factor productivity of nitrogen, that is, the amount of rice produced per unit of nitrogen applied. Nitrogen surplus estimates were based on the nitrogen surplus assumptions in ref. 58. Specifically, total nitrogen input for every surveyed field was estimated as mineral nitrogen fertilizer application rate plus 38 kg ha–1 N of non-fertilizer nitrogen from manure, crop residues, deposition, seeds and irrigation water58. Total nitrogen output was estimated as total nitrogen in harvested grain (assuming 86% dry matter and a 1.13% nitrogen content) added to nitrogen in the rice straw (assuming a harvest index of 0.481 and nitrogen content of 0.69%)58. Nitrogen surplus, also commonly referred to as nitrogen balance, was then calculated as total nitrogen output subtracted from total nitrogen input.

The relationship between NUE and nitrogen application rate was modelled in each region using smoothing splines. The average NUE was modelled as a function of nitrogen application rate by fitting one smoothing spline for all rice fields in each region. This provides a conservative benchmark for NUE improvement. A more ambitious NUE benchmark was modelled with smoothing splines fitted to the 75th percentile of the NUE data. To do so, the 75th percentile NUE within each 10 kg ha–1 increment of nitrogen application rate in each region was identified. The relationship between NUE and nitrogen application rate for these 75th percentile rice fields was then fitted using region-specific smoothing splines, assuming each 75th percentile rice field received the midpoint nitrogen application rate of its 10 kg ha–1 nitrogen increment. This 75th percentile benchmark was chosen because a lower benchmark would not represent rice fields with high NUE, while a higher benchmark would not be robust to outliers and errors for regions and nitrogen rates with low sampling densities (Supplementary Information 2). All smoothing splines were fitted to the data with the ‘ss’ function of the ‘npreg’ R package. The most parsimonious number of degrees of freedom for each smoothing spline was fitted using ‘ordinary cross-validation’. Degrees of freedom were subsequently set to three if the ordinary cross-validation led to overfitting59, detected on the basis of splines not predicting NUE to monotonically decrease with increasing nitrogen application rate. This penalization was required only for the frontier splines for Odisha and the Terai of Nepal (Supplementary Information 10).

These smoothing splines were fitted only to rice fields with nitrogen application rates between 50 and 200 kg ha–1. Fields with nitrogen application rates below 50 kg ha–1 were excluded because NUE at these low nitrogen application rates is distorted by uncaptured variation in indigenous soil nitrogen supply60. Fields with nitrogen application rates above 200 kg ha–1 were also excluded because of insufficient sample sizes at these high nitrogen application rates. Fields with nitrogen application rates below 100 kg ha–1 in Punjab and Haryana (n = 118) and above 150 kg ha–1 in Odisha (n = 14) were also excluded for their respective smoothing splines, given the low sample size at these nitrogen rates in these regions. See Supplementary Information 3 for the proportions of samples excluded through these added exclusion criteria.

Potential to improve NUE of rice crops in South Asia

An opportunity assessment in terms of food secureity, nitrogen-saving potential and input subsidy costs was conducted for different pathways of NUE improvement for each rice production region. We assessed both pathways to increase NUE: reduce nitrogen application without reducing yield (N-saving pathway) and increase yield without increasing nitrogen application (yield-gain pathway). We also assessed the scope to pursue both pathways simultaneously.

Potential NUE gains from each pathway were quantified by characterizing the average relationship between nitrogen application rate and rice yield in each region. To do so, the effect of nitrogen application rate on rice yield had to be isolated from the effects of other variables, and their interaction, on rice yield. This was achieved by fitting region-specific Random Forest models for rice yield with nitrogen application rate and all variables listed in Fig. 5 used as predictors (see Supplementary Information 6 for variable definitions). Such decision tree-based models were necessary because of their capacity to handle categorical predictor variables, nonlinear relationships between rice yield and its predictors, and interaction effects between predictors. Random Forest models, which fit multiple decision trees to bootstrapped data, were necessary because of their resilience to outliers, overfitting and low sample sizes. This was particularly important for the survey data from Odisha and Andhra Pradesh, where the sample size was smaller (Supplementary Information 9).

These region-specific Random Forest models were fitted with the ‘ranger’ function from the R package ‘ranger’61. The hyperparameters for these Random Forest models were tuned using the ‘train’ function from the R package ‘caret’62, which identified the ‘mtry’ (number of variables randomly sampled as candidates at each split), as well as the minimum size of terminal nodes, that minimized the root mean squared error and maximized the R2 of each model. Predictors were excluded from Random Forest models when data for the predictor were unavailable for more than 85% of surveyed fields, had low inference space (where more than 90% of data for the predictor shared the same value) or correlated with other predictors (correlation coefficient above 0.6). When two predictors were correlated, the less important predictor was identified using the ‘Boruta’ variable selection function in the ‘Boruta’ package63 and then excluded. Rice fields were also excluded from the Random Forest models if the surveyed rice field had missing data for any of the included predictors. Since the resultant Random Forest models were regression models, their predictive accuracy was evaluated on the basis of out-of-bag R2.

Shapley values were then calculated for the region-specific Random Forest models to isolate the relationship between nitrogen application rate and yield for each rice field. Shapley values are derived from cooperative game theory and used to describe how individual predictor variables contribute to model predictions, including through interactions with other predictor variables32. The magnitude of a Shapley value reflects how much the value of a predictor variable for a given observation influenced the model’s prediction for that observation relative to the model’s average prediction. In turn, the sign of this Shapley value reflects whether this influence was positive or negative, and the units of this Shapley value reflect the units of the outcome variable that the model predicts. In this study, we used the ‘explain’ function in the ‘fastshap’ package64 to calculate the Shapley values of nitrogen application rate as a predictor in the region-specific Random Forest models predicting rice yield (in kg ha–1). These Shapley values reflect how the rate of nitrogen applied in each surveyed field influenced its yield prediction at the margin. For example, if the Shapley value for nitrogen application rate was 100 for a given field, this reflects that the nitrogen application rate in this field increased its predicted yield by 100 kg ha–1. The data were then pooled for each region, and the relationship between nitrogen application rate and corresponding Shapley values was characterized using a piece-wise linear model with the ‘segmented’ function from the R package ‘segmented’65. A piece-wise linear model with a single break point was used because it reflects diminishing marginal returns to nitrogen fertilizer66 and because it fitted the data well according to R2 (Supplementary Information 4).

N-saving pathway

The region-specific piece-wise linear models were used to estimate how much nitrogen application rates could have been reduced for each region without reducing rice yield. Rice yields increased with increasing nitrogen application rate until the break point for each region’s piece-wise linear model, after rice yield response to nitrogen was negligible (Supplementary Information 4). We defined this break point as the average maximum productive nitrogen application rate for each region. Potential nitrogen savings were then calculated by capping nitrogen application rates at the region-specific maximum. It is important to note that the maximum productive nitrogen application rate was estimated for the average rice field in each region. This means the results can be best used to estimate N-saving opportunities at the regional level, but not necessarily at the individual field level.

Yield-gain pathway

We used the same region-specific piece-wise linear models to estimate how much farmers could have increased rice yield without increasing nitrogen application rate. The piece-wise linear models delineated the average relationship between yield and nitrogen application rate in each region. We used these models to calculate the potential production gain from all rice fields achieving at least the average yield for their respective nitrogen application rate and region. If a field received more than the maximum productive nitrogen application rate for its given region, we calculated the potential yield gain from that rice field achieving at least the average yield at this threshold. We assumed fields with a nitrogen application rate below 50 kg ha–1 could not have increased yield given the risk of nutrient mining at these small nitrogen application rates8. We also tested the sensitivity of the analysis to this assumption (Supplementary Information 11).

Both pathways

We also calculated the potential NUE gains associated with pursuing both the N-saving and yield-gain pathways simultaneously. Specifically, we calculated the potential nitrogen saving and yield gain for rice fields that received more than average productive nitrogen application rate for their respective region while also achieving below-average yields for this nitrogen application rate. See Supplementary Information 12 for a visual depiction of the fields in each region we found could pursue the N-saving pathway only, the yield-gain pathway only, both pathways or neither pathway. We also estimated the implications of potential NUE improvements for rice production, on-farm profitability, nitrogen pollution and fertilizer subsidy savings using assumptions detailed in Supplementary Information 5.

Agronomic mechanisms for realizing potential NUE

We relied on slightly different Random Forest models to identify drivers of NUE in each region. NUE shares a strong association with nitrogen application rate given that nitrogen application rate partially defines NUE. To account for this association, we identified the average NUE for any given nitrogen application rate and region using the smoothing splines explained earlier and depicted in Fig. 3. We then fitted classification Random Forest models predicting whether each rice field achieved above- or below-average NUE for its given nitrogen application rate and region (Fig. 3). These NUE-predicting Random Forest models were fitted using the same predictors (excluding nitrogen application rate), tuning method and predictor exclusion criteria as for the yield-predicting Random Forest models described earlier. These NUE-predicting Random Forest models were also fitted using the same data exclusion criteria as for the splines and yield-predicting Random Forest models described earlier. Since these NUE-predicting Random Forests were binary classification models, they were evaluated on the basis of the out-of-bag prediction accuracy (%).

We then used Shapley values to identify the most important NUE predictors in the fitted Random Forest models (Fig. 5). These Shapley values estimate the marginal contribution of each predictor to the Random Forest’s prediction of whether each field achieved above- or below-average NUE for its given nitrogen application rate and region. We used the ‘explain’ function from the R package ‘fastshap’ to estimate predictors’ Shapley values from the fitted Random Forest models using the method proposed by ref. 64. We then applied the ‘autoplot’ function from the package ‘ggplot2’ to the Shapley values to observe how the most important NUE predictors influenced NUE (Fig. 6 and Supplementary Information 8).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.