nature food

Article https://doi.org/10.1038/s43016-023-00913-8

Global spatially explicit yield gap time

trends reveal regions at risk of future
crop yield stagnation

Received: 10 October 2022 James S. Gerber 1,2 , Deepak K. Ray 1, David Makowski 3,
Ethan E. Butler 4, Nathaniel D. Mueller 5, Paul C. West 2,6,
Accepted: 6 December 2023
Justin A. Johnson 6, Stephen Polasky 6, Leah H. Samberg7, Stefan Siebert 8

Published online: 26 January 2024 & Lindsey Sloat9

Check for updates

Yield gaps, here defined as the difference between actual and attainable
yields, provide a framework for assessing opportunities to increase
agricultural productivity. Previous global assessments, centred on a
single year, were unable to identify temporal variation. Here we provide
a spatially and temporally comprehensive analysis of yield gaps for ten
major crops from 1975 to 2010. Yield gaps have widened steadily over most
areas for the eight annual crops and remained static for sugar cane and oil
palm. We developed a three-category typology to differentiate regions of
‘steady growth’ in actual and attainable yields, ‘stalled floor’ where yield is
stagnated and ‘ceiling pressure’ where yield gaps are closing. Over 60% of
maize area is experiencing ‘steady growth’, in contrast to ∼12% for rice. Rice
and wheat have 84% and 56% of area, respectively, experiencing ‘ceiling
pressure’. We show that ‘ceiling pressure’ correlates with subsequent yield
stagnation, signalling risks for multiple countries currently realizing gains
from yield growth.

The green revolution coincided with a doubling of global crop produc- all of which compare current yields (measured or modelled) to a yield
tion from the late 1960s to 2000 (ref. 1), alleviating fears of a Malthusian ceiling11. Conceptualizations of yield ceiling range from agronomic
crisis. However, these production increases have come at a substantial potential12,13 to best-in-class regional yield6,14–16. We use as yield ceiling
environmental cost, and steadily increasing food demand is placing an ‘attainable yield’ defined as the 95th percentile observed regional
additional pressure on natural resources2,3. Closing yield gaps is widely yield, intending to estimate the highest yield attained somewhere
cited as a pathway for increasing production while minimizing envi- in the world in each set of biophysical conditions. This definition of
ronmental impacts1,2,4–9 and is directly related to several of the United attainable yield follows that of Evans and Fischer17 and is effectively the
Nations Sustainable Development Goals10, including no poverty, zero same as the ‘feasible yield’ defined by van Dijk et al.18 or the ‘plateau’ in
hunger, decent work and economic growth, climate action and life exploitable yield as articulated by van Ittersum et al.8, which remains
on land. 15–25% below17 agronomic potential8,18–21.
Considerable efforts have been devoted to calculating yield gaps Most yield gap studies to date have been snapshots in time and
at local to global scales, using various complementary methodologies, lead to limited policy recommendations. Van Oort et al.22 classify yield

1
Institute on the Environment, University of Minnesota, St Paul, MN, USA. 2Project Drawdown, . 3Université Paris-Saclay, INRAE, AgroParisTech, Palaiseau,
France. 4Department of Forest Resources, University of Minnesota, St Paul, MN, USA. 5Department of Ecosystem Science and Sustainability, Department
of Soil and Crop Sciences, Colorado State University, Fort Collins, CO, USA. 6Department of Applied Economics, University of Minnesota, St Paul, MN,
USA. 7Rainforest Alliance, https://www.rainforest-alliance.org. 8Department of Crop Sciences, University of Göttingen, Göttingen, Germany.
9
World Resources Institute, Washington DC, USA. e-mail: [email protected]

Nature Food | Volume 5 | February 2024 | 125–135 125

Article https://doi.org/10.1038/s43016-023-00913-8

gap studies into ‘narrow scope high detail’ relevant to analysing spe- Table 1 | Percentage of harvested area with growth in
cific interventions but with limited range of applicability, and ‘broad attainable yield and average annual percentage growth in
scope low detail’ that can address questions about the envelope of attainable yields
sustainable production possibilities or indicate locations to target
Percentage of harvested Average annual percentage
agronomic intervention22. Many authors have evaluated the envelope
area with significant growth growth in attainable yield
of sustainable production based on static yield gaps5,6,22,23, but there in attainable yield
have been limited attempts to derive policy recommendations from
Entire First and last Entire First and last
these studies22. Van Oort et al.22 identify five categories based on a period decades period decades
combination of economic, climatic and humanitarian considerations,
1975– 1975– 2000– 1975– 1975– 1975–
and recommend specific policy interventions ranging from increasing 2010 1985 2010 2010 1985 2010
resource use efficiency to agricultural research and development22.
Barley 69% 43% 1% 0.8% 1.3% 0.7%a
Fischer and Connor assess yield gaps and divide the world into two
typologies: low input–large yield gap, indicating a need for improving Cassava 85% 0% 75% 0.8% 0.1%a 3.4%
farmers’ access to tools for management, and intensified small yield Maize 100% 88% 97% 1.4% 1.4% 1.5%
gap, indicating a need for increases in potential yields24.
Oil palm 7% 0% 1% 1.3% 1.6%a 0.3%a
A few previous studies have considered yield gap trends through
time. A monograph by Fischer et al. surveys trends in yields and yield Rapeseed 99% 86% 1% 1.6% 2.2% 2.1%
ceilings based on case studies covering multiple crops and regions23. Rice 100% 94% 65% 1.2% 1.8% 0.7%
However, this analysis does not have consistent global coverage. Fis- Sorghum 94% 53% 0% 0.1% a
0.5% a
0.3%
cher et al.23 reach policy conclusions consistent with those of Fischer
Soybean 100% 89% 86% 1.1% 1.1% 1.0%
and Connor24. Hatfield and Beres derive yield gaps for wheat for ten
countries based on a quantile regression analysis of national time series Sugar cane 56% 6% 0% 0.4% 0.6%a −0.2%a
of yields supplemented by state- and county-level data16. Such studies Wheat 98% 78% 4% 1.3% 1.8% 0.7%a
are valuable contributions and can provide a greater understanding Areas for which attainable yields increase over the indicated period with 95% confidence
of both production possibilities and specific policy prescription than intervals (first three numerical columns) and average annual per-year linear growth in globally
snapshot-in-time studies. We aim here to provide a study that is more averaged attainable yield (last three numerical columns). Statistical analysis was carried out
independently for the full interval as well as the first and last decades. Calculations of area
comprehensive in number of crops, spatial resolution of data and
and attainable yield growth are relative to 2000 for the 1975–2010 and 2000–2010 intervals,
global coverage and draw policy-relevant conclusions on trends in and relative to 1975 for the 1975–1985 interval. Confidence intervals and regional results
global production potential and indications of desirable interventions are shown in Supplementary Table 11. Results with fixed-area counterfactuals are shown in
for assuring food security. Supplementary Table 3. aNot significant (95% confidence intervals encompass zero).

We calculate spatially explicit global time trends in attainable yields

and yield gaps from 1975 to 2010 for ten crops comprising 83% of global the first and last decades in the study. Barley, rapeseed, sorghum and
calories (maize, wheat, rice, oil palm, soybean, barley, sugar cane, sor- wheat have seen an order of magnitude drop in area of attainable yield
ghum, rapeseed and cassava). We use a quantile regression model with growth, while cassava went from 0 to 75%. Maize and soybean show an
year-specific coefficients to calculate the area-weighted 95th percentile increase in attainable yield over 85% of area in all decades (Table 1 and
yields for each year across the world given local climate, soil characteris- Supplementary Table 1).
tics and irrigation management. These 95th percentile yields are designed The rate of attainable yield growth shows variation across crops,
to quantify the best yields in each set of biophysical conditions, denoted regions and time periods with attainable yield growth at the global
the ‘attainable yield’. The present method extends climate analogue scale over 1975–2010 for all crops except sorghum (Table 1 and Sup-
approaches6,14 with inclusion of a broader set of biophysical variables plementary Tables 3, 10, 11 and 15–24). Maize, rapeseed and soybean
and methods that result in continuous yield surfaces, few parameters show >1% growth in attainable yield over the 1975–2010 period and in
relative to process-based models and calculation of confidence inter- the latest decade. While wheat and rice also experienced >1% growth
vals. Our analyses are based on a high-resolution historical crop dataset over the full interval, growth rates have fallen in the most recent decade.
derived from census and survey information across ∼20,000 political Cassava stands out for a large increase in attainable yield growth in the
units. We use a static climatology that leads to more accurate models of most recent decade (Table 1 and Supplementary Table 2).
yield gap trends than those based on yearly data. To facilitate comparison Yield gaps have increased over areas ranging from 10% (oil palm)
between crops and time periods, we quantify growth as percentage of to 71% (maize) of the 2000 harvested area (Table 2 and Supplementary
linear change relative to 2000 yield values, where possible. We chose Figs. 2 and 3) between 1975 and 2010. Rice and wheat have substantially
2000 as a well-studied baseline2,6 considered the end of a phase of the less area with growing yield gaps in 2000–2010 than in 1975–1985
green revolution1,25. We report results for the globe and eight geographical (Table 2). By contrast, both maize and soybean have growing yield
regions. Detailed results are presented in Supplementary Information for gaps in more than 37% of area for all periods studied (Table 2 and
countries whose production exceeds 1% of 2000 production. Supplementary Table 4).
Globally averaged yield gaps have increased for barley, maize,
Results rapeseed, rice, soybean and wheat from 1975 to 2010, with no significant
Yield gaps are dynamic change for cassava, oil palm, sorghum or sugar cane (Table 3 and Sup-
Consistent with a history of growth in actual yields12,26,27, attainable plementary Fig. 2). Only maize, rapeseed and soybean have seen grow-
yields have increased from 1975 to 2010 over a majority of areas (Table 1 ing yield gaps in both the first and last decade of this interval, whereas
and Supplementary Fig. 1). Attainable yields increased over more than rice and wheat went from growth in yield gaps of nearly 2% per year in
94% of the 2000 harvested area between 1975 and 2010 for six crops 1975–1985 to no growth in the 2000–2010 interval (Table 3). Region-
(maize, rapeseed, rice, sorghum, soybean and wheat). Cassava and ally, there is heterogeneity in yield gap change, with substantial areas
barley had attainable yield growth over 85% and 69% of harvested area. showing significant decreases in yield gaps (Supplementary Table 12).
Two perennial crops, sugar cane and oil palm, experienced growth in Averaged relative yield gaps for 1975 and 2010 are shown in Fig. 1,
attainable yield over 56% and 7%, respectively, of the year 2000 har- revealing regions where yield gaps have closed (for example, eastern
vested area (Table 1 and Supplementary Table 10). The extent of areas Asia, Brazil, Australia) and widened (for example, sub-Saharan Africa,
undergoing growth in attainable yield has changed dramatically across eastern Russia.)

Nature Food | Volume 5 | February 2024 | 125–135 126

Article https://doi.org/10.1038/s43016-023-00913-8

Table 2 | Percentage of harvested area with growth or a

decrease in yield gaps

Area with significant growth Area with significant

in yield gap decrease in yield gap
Entire First and last Entire First and last
period decades period decades
1975– 1975– 2000– 1975– 1975– 2000–
2010 1985 2010 2010 1985 2010

Barley 36% 21% 1% 4% 1% 1% 1975

Cassava 48% 0% 46% 13% 15% 3%
Maize 71% 41% 55% 16% 14% 14%
Oil palm 10% 0% 0% 7% 0% 4%
Rapeseed 52% 29% 5% 5% 3% 1%
Rice 56% 58% 15% 24% 9% 24%
Sorghum 62% 28% 5% 7% 8% 2%
Soybean 61% 44% 37% 15% 10% 17% Relative yield gap (%)

Sugar cane 28% 6% 4% 22% 3% 13%

2010 0 10 20 30 40 50 60 70 80 90 100

Wheat 62% 36% 5% 8% 1% 4% Fig. 1 | Average relative yield gaps for ten major crops in 1975 and 2010. Crops
Areas where yield gaps increase or decrease over the indicated period with 95% confidence include barley, cassava, maize, oil palm, rapeseed, rice, sorghum, soybean, sugar
intervals. Area calculation is relative to 2000 for the 1975–2010 and 2000–2010 intervals, and cane and rice. Relative yield gap (shown as the percentage of the attainable yield
relative to 1975 for the 1975–1985 interval.
achieved) in each grid cell is calculated as an area-weighted average across the
crops and is shown on the top 98% of the growing area.
Table 3 | Average annual percentage change in globally
averaged yield gaps
have a significant rate of closure over 23% of area, representing 27%
Average annual trend in yield gap
of 2000 rice production, and are on track to close by 2030 for 18%
Entire period First and last decades of 2010 area (Fig. 2 and Table 2). Soybean yield gaps are closing over
1975–2010 1975–1985 2000–2010 16% of area (Table 2). Yield gaps for maize, in contrast, are growing
significantly over 56% of area and closing over 13% while only 3% of
Barley 0.9% 1.7% a
0.3%a
maize area is trending towards closure by 2030 (Fig. 2 and Table 2).
Cassava 0.6%a −1.9%a 5.6% Yield gaps are widening in more areas than they are closing for
Maize 1.2% 1.1% 1.6% multiple crops (Table 2).
Oil palm −0.4%a −0.7%a −2.6%a Rates of gap closure are particularly acute in some countries and
regions. For example, 11%, 51% and 34% of rice production in Indonesia,
Rapeseed 1.5% 2.1% 1.5%
Bangladesh and Vietnam (the world’s 3rd, 4th and 5th largest produc-
Rice 0.8% 2.1% −0.4%a ers), respectively, are on a trajectory to close yield gaps by 2030 (Sup-
Sorghum 0.4%a −0.2%a 0.8%a plementary Information). In China, the world’s largest wheat producer,
11% of harvested area has yield gaps closing by 2030. In Thailand, 86%
Soybean 0.8% 1.1% 1.4%
of sugar cane production trends to yield gap closure by 2030.
Sugar cane 0.1%a 0.6%a −1.3%a
Wheat 1.2% 1.9% 0.1%a Yield gap typologies reveal trends relevant to food security
Reported change is per-year linear change relative to 2000 yield gaps for the 1975–2010 and Linear yield gap trends cannot by themselves differentiate underly-
2000–2010 intervals, and relative to 1975 for the 1975–1985 interval. Confidence intervals and ing causes of change, for example, stagnation in yield versus growth
regional results are shown in Supplementary Table 12. Means are calculated after rejecting
in attainable yield. We introduce a three-element typology of yield
outliers (some realizations have yield gap ∼0, leading to infinite relative growth rates).
Outliers are defined as points outside of the interval mean + 4s.d. Results with fixed-area gap trajectories (Fig. 3), including ‘steady growth’ in which yield
counterfactuals are shown in Supplementary Table 3b. aNot significant (95% confidence gaps grow while attainable yield (‘ceiling’) and actual yield (‘floor’)
intervals encompass zero). increase, ‘stalled floor’ in which attainable yield grows but actual
yield stagnates and ‘ceiling pressure’ in which yield gaps close and/
Yield gap closure correlates with yield stagnation or attainable yield stagnates.
Agricultural census units where the yield gap is closing have a greater Yield gap trends are experiencing steady growth over half the har-
likelihood of future yield stagnation than census units where the yield vested area for cassava (65.2%), maize (60.3%) and sorghum (51.2%). By
gap is static or widening. For all crops studied except oil palm, signifi- contrast, rice has the lowest percentage of harvested area experiencing
cant yield gap closure from 1986 to 2000 increases the likelihood of steady growth (12%) and 84% is experiencing ceiling pressure (Figs. 3
yield stagnation from 2000 to 2012 by factors ranging from 1.78 (sugar and 4, Table 4 and Supplementary Table 13). The crop with the second
cane) to 3.75 (soybean) (Supplementary Table 5). Here we define yield greatest area experiencing ceiling pressure is wheat at nearly 56%. All
stagnation following Grassini28 as a plateau after a period of growth. crops have some area with stalled floor, ranging from 3.9% (rice) to
This increase in stagnation likelihood is not merely a ‘reversion to the 18.7% (cassava.)
mean’ effect associated with high yield growth rates: limiting analysis
to regions in the top yield-growth quartile leads to qualitatively similar Discussion
results (Supplementary Information). These results inform a long-running debate on potential for future
We evaluate trends in yield gap over the period 1998–2012 to growth in yields. In general, we find that both actual yields and attain-
identify areas at risk of future yield stagnation. For rice, yield gaps able yields have continually improved over many decades8,12,26,27.

Nature Food | Volume 5 | February 2024 | 125–135 127

Article https://doi.org/10.1038/s43016-023-00913-8

Maize Rice

Widen
No trend
Closure by 2030
Closure by 2050
Closure 2050
onwards Wheat Soybean

Fig. 2 | Time to closure of yield gaps based on linear extrapolation of trends yield gap closure occurs within 95th percentile confidence intervals. ‘Widen’
from circa 2000 to circa 2010 for maize, rice, wheat and soybean. Yield gap indicates that yield ceiling and linear trend are significantly diverging (that is,
closure time is defined as the crossing point of the linear trend of attainable yield the crossing point is before 2010). See Supplementary Fig. 3 for other crops.
and the linear trend of actual yield relative to 2010. ‘No trend’ indicates that no

However, the nature of how they have grown—and the evolution of yield ceilings, notably investment in breeding technologies and improved
gaps—sheds light on contrasting claims regarding likely trajectories for agronomic practices24,23. Closing yield gaps in China have been noted
future yield growth. One view is that historical growth in crop yields is along with calls for increased investment in management technology
due to ‘one-time innovations’28: a ceiling is being approached, and meet- and advanced cultivars34.
ing demand beyond 2030 will require novel technological advances12. The ‘stalled floor’ category is characterized by a significant
This could be the situation for rice: over 3 decades, the ratio of har- decrease in growth rate of actual yield coupled with an increase in
vested area with yield gaps growing versus shrinking flipped from 6:1 the yield gap. There can be multiple reasons for this, ranging from
to 1:2 (Supplementary Table 4) with 84% of rice area now experiencing economic shocks to adoption of environmental policies to lack of
‘ceiling pressure’ (Supplementary Table 13). A contrasting view has been investment in agronomics. Practices in western Europe intended to
that economic incentives will lead to continuous future improvement29 provide environmental benefits may lead to stalling yields because
and that increases in agricultural research and development (R&D) will such practices (for example, reduction of fertilizer inputs, adoption of
lead to further increases in crop yield30,31. Maize trends are consistent agro-ecological systems, banning of chemical pesticides) do not aim
with this view, with increases in production in areas with diverging yield solely to maximize yields but rather to balance trade-offs with environ-
gaps increasing over the decades (Supplementary Table 4) and 60% of mental impacts35. A related example is legislation in Italy limiting the
maize area experiencing ‘steady growth’ (Supplementary Table 13). uptake of genetically modified strains of maize that can resist pests
The present approach represents a possible method for relating yield has caused yields to decrease36 (Supplementary Fig. 5). Agronomic
outcomes to investment in agricultural R&D and extension that will be decisions can also lead to decreases in yield growth. As yield is reported
critical to achieving food security and climate goals32. in units of tonnes per cultivated hectare, this can be associated with
Regions experiencing ‘ceiling pressure’ have closing yield gaps increased productivity in conjunction with increasing multicropping,
and are at risk of a pronounced decrease in future yield growth (if not particularly prevalent for rice37, but can also come about because a crop
outright stagnation) in the absence of investment to raise attainable is less favoured, such as sorghum in the United States being displaced
yields. While this result is intuitive, we have quantified the effect and by maize on the most productive farmland23.
shown that measures based on yield gap closure rates are more reliable Yield gaps evolve with change in either yield ceiling or floor, and
predictors of future stagnation than measures based on local yield ‘steady growth’ reflects ongoing increase in both, characterized by
time series or size of yield gap. By calculating yield gap closure over an consistent growth in attainable yield, associated with investment in
equivalent time interval, we identify risk of future stagnation for multi- agricultural R&D, and growth in actual yield, associated with diffusion
ple regional crops, particularly rice in Asia, with Vietnam, Bangladesh of improved management practices and results of R&D. An example of
and China at heightened risk with 75%, 59% and 34%, respectively, of this category is sorghum in India (Supplementary Fig. 5), where a strong
production in areas where yields are approaching the ceiling (Sup- programme of private–public partnership led to the development of
plementary Fig. 4 and Supplementary Information). Wheat yields in private plant breeding and uptake of improved hybrids23. Maize in
many European countries also have much harvested area undergoing many countries (Fig. 4) is undergoing this steady growth. In Mexico,
‘ceiling pressure’ (Supplementary Fig. 4 and Supplementary Informa- this trend is evident (Supplementary Figs. 4 and 5) in spite of relatively
tion), consistent with the yield stagnation observed by other authors33. large yield gaps associated with farmers growing lower-yielding white
Policies to address ‘ceiling pressure’ include those that increase yield corn preferred by consumers38.

Nature Food | Volume 5 | February 2024 | 125–135 128

Article https://doi.org/10.1038/s43016-023-00913-8

a with a fixed-area counterfactual, indicating that production has shifted

Steady growth to areas with lower yield ceilings (Supplementary Table 3) with oil palm,
rapeseed and sugar cane as exceptions. Similarly, yield gaps for all crops
except rapeseed have decreased compared with the fixed-area coun-
terfactual suggesting that increases in cultivated area tend to occur
in regions with higher yield attainment. This is consistent with Jevon’s
paradox of intensification leading to increased landcover change.
Recent yield gap trends (circa 2000 to 2010) are significantly
Yield

different for maize and soybean than for wheat and rice. Maize and
soybean, which respectively deliver 24% and 52% of calories directly
or indirectly as food39, have significantly increasing yield gaps. By
contrast, wheat and rice, which deliver over 78% and 86%, respectively,
of their calories as food39, have seen no significant change in yield gaps.
Attainable yield This result is probably due to increased investment for maize and soy-
Yield gap
Yield bean, correlated with net production value increases over this time
period of 289% and 292% compared with increases of 106% and 87% for
wheat and rice, respectively27. Thus, ‘steady growth’, with its attendant
b Stalled floor
potential for future yield improvement12, is occurring for crops that
largely do not feed people (directly). An optimistic reading of this is
that commensurate levels of investment are possible for crops that are
more critical to food security, and a growing body of research suggests
that R&D investment in crops critical to food security can improve
yields31,40,41 and will be required to meet future food demand12. Recent
research, however, shows that the temporal relationship between R&D
expenditure and yield improvements exhibits longer lag times and
Yield

more uncertainty than previously estimated, especially for developed

countries42. Such lags increase the urgency of making R&D investments.
The quantile regression approach, based on a large census-based
empirical yield dataset and global biophysical dataset, complements
other possible methods for calculating yield gaps over time. Experi-
Attainable yield
Yield gap mental approaches to calculating yield potential can be highly accurate
Yield for specific locations23, but difficult to generalize across space and
time owing to cost and the difficulty of representing a larger region.
Running process-based models at point locations to calculate poten-
c
Ceiling pressure tial yield can incorporate detailed information about soils, weather
and management and be more representative of a larger region8,43–46.
However, results from such models can depend on the details of how the
models are parameterized47. Unlike experimental and process-based
approaches, the global quantile regression approach used here is sensi-
tive to the dynamic, real-world changes in cultivar and management as
it is based on census and survey data at administrative units around the
Yield

world. However, the present approach also has limitations owing to the
scale and type of information that can be used in a global model. For
example, we may overestimate yield gaps in Africa because of limited
soil rooting depth48. Variables such as slope, soil organic carbon and
pH are significant in explaining yield quantiles, but using these model
Attainable yield terms to predict results of extensification requires detailed assump-
Yield gap
Yield
tions about the soil properties of land available for expansion. These
results do not by themselves speak to the constituent components of
1970 1975 1980 1985 1990 1995 2000 2005 2010 yield gap, a topic that can be usefully addressed with process-based
Year approaches18. The focus of our study is temporal trends in attaina-
Fig. 3 | Typologies of yield gap closure. a, ‘Steady growth’ category, an ble yields—application of similar methods using cross-validation in
archetype of which is the situation when attainable yield (‘ceiling’) and actual physical space and climate space would result in attainable yield sur-
yield (‘floor’) benefit from agronomic investment in new technologies and faces more appropriate for the assessment of production gaps. While
increased uptake of management practices. b, ‘Stalled floor’ category; yield gap is there may be crop and climate combinations that do not experience
increasing because ‘best in class’ management practices for maximizing yield are advanced management leading to conservative predictions of attain-
not widely adopted for reasons that could include economic barriers, selection able yield, the trends will still be valid. Global approaches can show
of lower-yielding higher-quality cultivars or adoption of environmental policies. sensitivity to choice of weather variables45,49 although we find our
c, ‘Ceiling pressure’ category, in which small yield gaps indicate a need for results to be robust with regard to selection of climatic datasets (Sup-
breeding and ‘new agronomy’24 to improve yield ceilings. Some archetypal
plementary Information). The present method is explicit with regard to
examples are shown in Supplementary Fig. 5.
choice of weather variables and climate dependencies (Tables 5 and 6).
The present approach, with its global scope and basis on empiri-
Yield gaps can vary with year-to-year on-field practices, but also cal yield data, emphasizes policy relevance over agronomic precision.
with structural changes, such as lower-productivity land coming into Indeed, top-down frameworks such as the one used here can lead to
cultivation. Yield ceilings for most crops have decreased compared instances at a local scale of predictions of yield ceiling below current

Nature Food | Volume 5 | February 2024 | 125–135 129

Article https://doi.org/10.1038/s43016-023-00913-8

Steady growth
Stalled floor
Maize Ceiling pressure Rice

Wheat
Soybean

Fig. 4 | Maps of typologies of yield gap change for maize, rice, wheat and soybean. Typologies are as defined in the text and illustrated in Fig. 3. Maps for other crops
are shown in Supplementary Fig. 4.

Table 4 | Global allocation of circa 2000 harvested area into Table 5 | Biophysical variables used in the construction of
three typologies of yield gap evolution the yield attainment model

Steady growth Stalled floor Ceiling pressure Variable Definition Source

Area Area Area GDD Growing degree days, Tbase = 0°C for WorldClim V2.1
all crops (ref. 63)
Barley 40.6% 6.4% 53%
MAP Mean annual precipitation WorldClim V2.1
Cassava 65.2% 18.7% 16.1% (ref. 63)
Maize 60.3% 14.6% 25.1% PCI Precipitation concentration index WorldClim V2.1
Oil palm 39.7% 10.7% 49.6% (ref. 63), Oliver64

Rapeseed 42.2% 11% 46.8% IRR Fraction of area equipped for irrigation Portmann et al.68,
Siebert et al.69
Rice 12% 3.9% 84.2%
AWC Available water capacity in the upper ISRIC65
Sorghum 51.2% 9.1% 39.6% 30 cm of soil
Soybean 48% 11.2% 40.8% SOC Soil organic carbon in the upper 30 cm ISRIC65
of soil
Sugar cane 28.8% 18% 53.2%
PH The pH of the upper 30 cm of soil ISRIC65
Wheat 28.8% 15.5% 55.7%
SLOPE30 Proportion of area with slope >30° Harmonized World
46
Soils Database67
production . By using census-based yields for the empirical com-
parison, this method sidesteps the need to untangle exploitable and VF Vernalization factor, binary used for WorldClim V2.1
wheat only; 1 if −8 ≤ Tavg ≤ 5 °C, where (ref. 63)
agronomic potential8,11,19–21. The global nature of the dataset provides Tavg is the average temperature of the
a framework for sustainability studies. Moreover, while the approach coldest month
is global, it has predictive power at the local scale as evidenced by the t Year
skill it shows to predict likelihood of yield stagnation. The inclusion
of irrigated fraction is another important benefit, allowing for the
exploration of scenarios with changes in irrigation; until now, this has thus, yield gap calculations are only indirectly sensitive to change in
been a notable drawback of analogue approaches8. This approach com- yield due to climate. However, future climate change could impact
plements other global approaches such as the Global Agro-Ecological projections from this analysis owing to factors beyond the scope of this
Zones50, finding similar quantitative results at a global scale for a dif- model such as yield losses due to increased pathogen risk51 or increased
ferent set of crops (Supplementary Fig. 6) while providing trends in yields due to adaptations of cultivars and cropping calendars52. Despite
attainable yield. this, a recent econometric analysis argued that historical adaptation
Recent climate change does not alter the conclusions of this analy- to climate change is negligible53 and actual yield losses from increased
sis. Because our method fits model parameters anew every year, climate pathogen risk are not well quantified and increased primarily in regions
change impacts on yield will be mirrored in changes to attainable yield; with yield gains from a changing climate51. Furthermore, future climate

Nature Food | Volume 5 | February 2024 | 125–135 130

Article https://doi.org/10.1038/s43016-023-00913-8

Table 6 | Terms included in each selected model can be obtained from a snapshot, and is critical to the definition of the
typology categories ‘ceiling pressure’, ‘steady growth’ and ‘stalled floor’,
Barley Y∼1+MAP+GDD^2+PCI+IRR+IRR*MAP+IRR*PCI+SLOPE30 which can help to determine types of intervention and where to target
Cassava Y∼1+GDD+MAP^2+GDD^2+IRR+IRR*MAP efforts to increase production. Continued growth of the attainable
yields identified here will require ongoing investment in agricultural
Maize Y∼1+GDD+MAP+GDD^2+PCI+IRR+IRR*MAP+SOC+SLOPE30
+PH technologies. Assuring this ‘room to grow’ for future crop yields is
critical to fulfilling the promise of the green revolution and providing
Oil palm Y∼1+GDD+MAP+GDD*MAP+GDD^2+PCI+IRR+IRR*MAP+SOC
food security for future generations.
Rapeseed Y∼1+GDD+MAP+GDD*MAP+IRR+IRR*MAP+IRR*PCI+SLOPE30
Rice Y∼1+GDD+MAP+GDD*MAP+MAP^2+GDD^2+IRR+IRR*MAP+MA Methods
P*PCI+IRR*PCI+SOC+PH Crop data
Sorghum Y∼1+GDD+MAP+IRR+IRR*MAP+MAP*PCI+IRR*PCI+SLOPE30 Tables of annual yield and harvested area data for barley, cassava,
maize, oil palm, rapeseed, rice, sorghum, soybean, sugar cane and
Soybean Y∼1+MAP+MAP^2+PCI+IRR+IRR*MAP+IRR*PCI+PH
wheat were compiled following the methods of Ray et al.56. In general
Sugar cane Y∼1+MAP+MAP^2+GDD^2+IRR+IRR*MAP+PH
terms, the data compilation method relies on obtaining and reconciling
Wheat Y∼1+GDD+GDD*MAP+MAP^2+IRR+IRR*MAP+MAP*PCI+VF*GD data for crop yield and area from a variety of public sources. Data are
D+VF+SLOPE30+PH reconciled across administrative units (that is, if the sum of reported
state-level production exceeds reported national-level production,
change will shift the likelihood of extreme conditions, which may also the state-level data are reduced by a factor to assure agreement at the
drive extreme yield losses54. While extreme climate change would lead coarsest levels). There is a gap-filling procedure that is required when
to the functional form of the yield surfaces being poorly matched in there is a gap in a data series at a subnational level. In this case, the miss-
later years, there are two reasons to discount this here. One is that ing data are filled with the last 5 year average data available, so that it
impacts on yield due to climate change over the time period studied scales with the data at the higher administrative level while retaining the
are typically smaller than growth due to evolution in cultivars and subnational patterns for a crop. Full details are available in Ray et al.56.
management26,55–57. A second reason is more concrete: we repeated We note that yield data are reported in terms of tonnes per har-
analysis with yearly weather data instead of a fixed climatology and vested hectare; the same parcel of land can have multiple harvests in
found that the latter led to lower temporal cross-validation errors a year. A data quality check for each crop–year combination rejected
(Supplementary Information). In short, while future climate change points with harvested area greater than 300% (that is, the multicrop-
may move the global system into a climate space that is poorly captured ping index for the entire region exceeds 3) or yield values greater than
by this model, the historical analysis presented here appears robust to 2 s.d. above the area-weighted 97.5th percentile yield value for each
climate change experienced to date. crop–year.
Identifying yield gaps is not the same as prescribing how to close Data were developed using the detailed process described in sec-
yield gaps. While this analysis looks to identify yield gaps and catego- tion 1.2 of Supplementary Text 1 in Ray et al.56, which in turn is based on
rize regimes of yield gap evolution, it does not offer prescriptions the approach of Monfreda et al.62. All data are from public sources and
regarding how those gaps should be addressed. The very concept of can be replicated by a reader using the methods and sources provided.
closing yield gaps is not value-neutral and can be problematic58. In If those sources are no longer available, the second author will provide
addition to the difficulties of addressing the myriad socio-economic that data. The maize, rice, wheat and soybean yield data are available in
factors that keep farmers from improving production, small yield gaps Supplementary Table 8 in Ray et al.56. Requests for the actual gridded
imply reduced potential for future growth20,58. In some contexts, for maps can be sent to the first author of Ray et al.56.
example, in which smallholder subsistence agriculture is prevalent,
significant investment in closure of yield gaps in primary crops may Biophysical data
not be appropriate. This might be the case in which the investments in Climate data. We reprocessed global datasets of monthly average
inputs required to increase yields lead to high levels of debt for farmers, temperature and monthly precipitation at 5 arcmin resolution from
especially where availability of inputs may fluctuate year to year, leav- the WorldClim V2.1 (ref. 63) to calculate grids of growing degree days
ing farmers open to risk of over-investment in specific crops or crop (GDD) with base temperature (Tbase) = 0 °C, mean annual precipitation
varieties. Singular focus on yields of staple crops may also occur at the (MAP), precipitation concentration index64 (PCI) and a binary vernaliza-
expense of diversified crop production, with negative consequences tion factor (VF) that takes on the value 1 if the coldest monthly winter
for local nutrition and food security. In the context of transnational temperature is less than or equal to 8 °C and zero otherwise. These
land acquisitions, closing yield gaps can increase production while resulting grids (all at 5 arcmin resolution) were incorporated into data
putting local food security at risk59. By contrast, Zhang et al.60 report tables as described below.
that intense agricultural outreach to smallholder farmers in Quzhou,
China, helped close yield gaps and increase farmer incomes. Yield gaps Soils data. We obtained 30 arcsec grids of available water capacity, pH
may also persist as reflections of consumer preferences (for example, (PH) and soil organic carbon at various soil depths from the SoilGrid-
white corn in Mexico38 or high-protein low-yield wheat61). s1km project65,66 (downloaded from www.isric.org on 27 June 2017).
Since 2000, the growth rates in yield gap experienced by maize, After individual depth layers were aggregated to 5 arcmin resolution,
soybean and rapeseed over the green revolution have continued, while soil properties in the top 30 cm were obtained using a trapezoidal
yield gaps have significantly closed for rice and stagnated for wheat. integration following Hengl et al.65. We included topographical data
The discrepancy in attainable yield growth between crops reinforces by downloading 5 arcmin grids with percentage of 100 m × 100 m sub-
calls for increased agricultural research investment in crops critical to pixels with average slopes in the intervals below 10°, between 10° and
food security19,30,31, and suggests region and crop combinations where 30°, and above 30° from Harmonized World Soil Database v 1.2 (ref. 67).
investment should be targeted. The method introduced here for iden- These resulting grids (all at 5 arcmin resolution) were incorporated
tifying and analysing yield gaps should be viewed as a complement to into data tables as described below.
computational approaches, particularly those that upscale local and
regional results from process-based models43,44. The temporal nature Irrigated area. To determine a time series of irrigated area, we scaled
of the yield gap analysis presented here provides more insights than the fraction of irrigated area as the maximum proportion of the crop

Nature Food | Volume 5 | February 2024 | 125–135 131

Article https://doi.org/10.1038/s43016-023-00913-8

growing area irrigated in each grid cell (IRR) from Mueller et al.6 (which ∑i τ| yi −qi |ai
yi > qi
⎧
was based on MIRCA2000 (ref. 68)) using ratios of area equipped for ⎪ ∑i ai
LFτ,year = (1)
irrigation from Siebert et al.69. We use linear extrapolation of the area ⎨ ∑i (1−τ)| yi −qi |ai
⎪ yi ≤ qi
equipped for irrigation to extrapolate beyond 2005, constraining the ⎩ ∑i ai

result such that IRR is between 0 and 1 (inclusive). These resulting

crop-specific grids (all at 5 arcmin resolution) were incorporated into where LF⍺,year is the loss function, τ is the quantile (here 0.95), yi is the set
data tables as described below. of yield values at each location i for ‘year’, qi is the set of model predic-
tions at each location i and ⍺i is the harvested area of the ith location in
Data tables. For each combination of administrative unit, crop and ‘year’ (corresponding to the administrative units used to make up the
year with both yield and area data, we compiled data tables from the data tables). These loss functions are summed up for each of the eight
gridded yield and area data. The yield and area data for the tables were half-decadal intervals. The ‘child’ model with the lowest sum of loss
calculated with an area-weighted average (over the crop-specific har- functions then becomes the parent model, and the process continues
vested area). Similarly, a value for each biophysical parameter in Table 5 until a model with the lowest loss function is selected as the best quan-
was calculated via an area-weighted average over the crop-specific tile regression model. We extended one generation past the lowest
harvested area. All datasets are available at 5 arcmin resolution with loss function in each case to assure there was not a local minimum in
the exception of the CRU data that were downscaled from 10 arcmin that generation.
to 5 arcmin. Thus, a consistent model for each crop was built with the
smoothed yield datasets using the cross-validation-in-time pro-
Dataset selection cedure discussed above. We then removed time terms from that
We carried out substantial portions of the analysis with three dif- model and determined model coefficients for each year based on
ferent datasets: a climatology from WorldClim V2.1 (Fick et al.63), a the annual data.
climatology from CRU V4.05 (Harris et al.70) and annual climate from We determined model coefficients for each year with a regularized
CRU V4.05 (Harris et al.70). We selected the WorldClim climatology quantile regression whose loss function was modified to assure that
because it led to the lowest temporal cross-validation error, sug- the yield surface encompassed the intended proportion of harvested
gesting it is most appropriate for a study focused on the interpreta- area. In other words, equation (1) places a weak constraint on the total
tion of time trends. We compared calculations for yield gap trends harvested area above and below the regression surface, so we added a
based on the use of a climatology to annual data to confirm that the term, λ, to enforce this. The regularized quantile regression is shown
conclusions presented in the paper are independent of this choice in equation (2).
(Supplementary Table 9).
∑i τ| yi −qi |ai
⎧ yi > qi
∑i ai
⎪
Model construction LFτ,year = λ +
⎨ ∑i (1−τ)| yi −qi |ai
Quantile regression models were built to predict the 95th quantiles ⎪ yi ≤ q i (2)
∑i ai
of yield (Y) as a linear function of several biophysical input variables ⎩
at each year. A multistep process was used to assure that model | ∑j a j |
λ = s.d.(y)∑i ai || − τ||
selection artefacts did not introduce spurious time trends. In a first ∑a

step, we adopted a stepwise approach for selecting a parsimoni-

ous model for each crop species. Our starting point was a model i is summed over all values, and j is summed over values for which q > y.
including the biophysical input variables shown in Table 5 with terms All results presented in the paper are based on averages of
selected on a physical basis (equation (3) or (4)). For each crop, the individual-year version of the model. Single-year results are
the parameters of this model were estimated at the 95th quantile area-weighted averages over a ±2 year window.
from the 5 year time series using the ‘quantreg’ package (version 5.33) Equation (3) represents all crops other than wheat.
in R (version 3.4.0) implementing the method described by Koenker71.
2 2
We carried out this procedure with a non-overlapping windowed Y ∼ GDD + MAP + GDD + MAP + GDD × MAP
dataset smoothed to eight half-decadal intervals via an average + PCI + PCI × MAP + IRR + IRR × MAP (3)
over ±2 years. As a step for removing terms, non-significant terms
+ IRR × PCI + AWC + SOC + Slope 30 + PH
(whose 95% confidence intervals overlap with zero) were sequentially
removed (starting with the term whose confidence interval most
centrally overlapped 0) until there were no non-significant terms left. Equation (4) represents only wheat.
The irrigation fraction and irrigation fraction–precipitation cross
2 2
term were kept in the model regardless of confidence intervals. Then, Y ∼ GDD + MAP + GDD + MAP + GDD × MAP + PCI + PCI
we further simplified the model using an iterative cross-validation ×MAP + IRR + IRR × MAP + IRR × PCI + AWC (4)
procedure: we generated a series of simplified ‘child’ models by
+ SOC + Slope 30 + PH + VF + VF × GDD + VF × MAP
removing each term (excepting linear and quadratic time terms,
the irrigation term and the precipitation–irrigation cross term).
We sequentially removed each layer of the smoothed time series as
a testing dataset, fitted the model on the remaining seven layers and Selected quantile regression models for 95th percentile yield
predicted the 95th yield quantile for the removed data. The predic- Table 6 represents the terms included in each crop-specific selected
tions were assessed by calculating a quantile-regression-specific loss model.
function71 based on the difference between the regression model
prediction and the yield data for the omitted year. We follow Mein- Model coefficients
shausen and Ridgeway72 and Koenker71 to compute the loss func- Model coefficients are given in the Supplementary Information
tion, LF⍺,year, although we modify equation 3 from Meinshausen and and Supplementary Table 14. The calculations are carried out using
Ridgeway72 to allow area weighting, make the summation explicit and z-scores, so the model coefficients are unitless. Supplementary Table 14
add a subscript ‘year’ to denote the central year of each smoothed contains the normalization factors relating the z-scores to the variables
half decade: in physical units.

Nature Food | Volume 5 | February 2024 | 125–135 132

Article https://doi.org/10.1038/s43016-023-00913-8

Model outputs Data availability

Overview. Model predictions and their confidence intervals were All weather, soil and irrigation data used in this study are publicly
derived with a bootstrap method. Using 1,000 random samples (with available and sourced in refs. 63,65–70 of Methods. Crop yield and area
replacement) of the crop and biophysical data, we generated 1,000 data are derived from publicly available sources (agricultural census
realizations of the coefficients of the quantile regression model for and survey reports as identified in ref. 56) as described and further
each crop. These 1,000 realizations were used to determine the confi- referenced in Methods. All data generated in the current study (Figs. 1,
dence intervals. We used 95th percentile confidence intervals unless 2 and 4) as well as annual potential yield surfaces for ten crops over the
noted otherwise. period 1973–2012 can be downloaded from https://doi.org/10.5281/
zenodo.10234041. All data inputs to the study, as well as all results, are
Calculation of time to yield gap closure. Calculation of time to yield available upon request from the corresponding author. The authors
gap closure is based on a linear regression of annual yield gaps over the commit to full and timely cooperation with any validation studies.
interval 1998–2012 (to assess data circa 2000–circa 2010.) We perform
a linear regression on each of the 1,000 time series of yield gap at each Code availability
political unit. If the 2.5th percentile and 97.5th percentile linear slopes The code was developed in the Matlab and R programming lan-
have the same sign, the time to yield gap closure is considered signifi- guages and can be downloaded from https://doi.org/10.5281/
cant. For each of the 1,000 realizations, the zero-crossing year of the zenodo.10234041.
average linear fit is at 1998 −x0/x1, where x0 is the y-value at 1998 and
x1 is the slope. Thus, yield gap closure time tclose relative to 2010 is cal- References
x
culated as tclose = − x0 − 12 for each realization. If, for a political unit, the 1. Evenson, R. E. et al. Assessing the impact of the green revolution,
1
time to yield gap closure is considered significant, the median closure 1960 to 2000. Science 300, 758–762 (2003).
time is used for analysis. 2. Foley, J. A. et al. Solutions for a cultivated planet. Nature 478,
The calculation used to construct Table 4 uses the same procedure, 337–342 (2011).
but over the time period 1986–2000. 3. Alexandratos, N. & Bruinsma, J. World Agriculture Towards
2030/2050: the 2012 Revision (FAO, 2012).
Calculation of year-specific data. All data are computed as a ±2 year 4. Springmann, M. et al. Options for keeping the food system within
average around the year presented, with the an area-weighted aver- environmental limits. Nature 519–525 (2018).
age carried out over five individual-year calculations of the specific 5. West, P. C. et al. Leverage points for improving global food
quantity. As an example, to determine the circa 2010 global average security and the environment. Science 345, 325–328 (2014).
yield gap, the attainable yield surfaces are calculated for 2008, 2009, 6. Mueller, N. D. et al. Closing yield gaps through nutrient and water
2010, 2011 and 2012; then, the actual yield surfaces are subtracted management. Nature 490, 254–257 (2012).
from these for each year, and the five resulting maps of yield gap are 7. Burney, J. A., Davis, S. J. & Lobell, D. B. Greenhouse gas mitigation
averaged together with weights according to the reported harvested by agricultural intensification. Proc. Natl Acad. Sci. USA 107,
area in each year. 12052–12057 (2010).
8. van Ittersum, M. K. et al. Yield gap analysis with local to global
Calculation of stagnation probabilities. To identify yield stagnation, relevance—a review. Field Crops Res. 143, 4–17 (2013).
we carried out a piecewise linear regression over 27 years of data, from 9. Mueller, N. D. & Binder, S. Closing yield gaps: consequences for
1986 to 2012, with a single discontinuity in slope at 2000. For each crop, the global food supply, environmental quality & food security.
at each political unit, we consider stagnation to occur in the following Daedalus 144, 45–56 (2015).
case: 95th percentile confidence limits around the 1986–2000 slope 10. Transforming Our World: The 2030 Agenda for Sustainable
are both positive, and lower confidence limits around the 2000–2012 Development (UN, DESA, 2016).
slope are negative. This definition of stagnation follows Grassini14. 11. Lobell, D. B., Cassman, K. G. & Field, C. B. Crop yield gaps: their
These test years were chosen so that there are 15 years in the time series importance, magnitudes, and causes. Annu. Rev. Environ. Resour.
leading to 2000, allowing us to draw conclusions from trends in yield 34, 179–204 (2009).
gap from the 15 year time series from 1998 to 2012. 12. Cassman, K. G. Ecological intensification of cereal production
systems: yield potential, soil quality, and precision agriculture.
Calculation of impact of yield gap closure on likelihood of stagna- Proc. Natl Acad. Sci. USA 96, 5952–5959 (1999).
tion. To assess the impact that yield gap closure trends on the likelihood 13. Fischer, R. A. Definitions and determination of crop yield,
of stagnation, we assessed linear yield gap closure trends over a 15 year yield gaps, and of rates of change. Field Crops Res. 182,
time series from 1986 to 2000 (Supplementary Table 7). We found that 9–18 (2015).
linear yield gap closure trends that will be closing within 30 years are 14. Licker, R. et al. Mind the gap: how do climate and agricultural
associated with a doubling of the probability that the yield series from management explain the ‘yield gap’ of croplands around the
2000 to 2012 will then stagnate (using the definition of stagnation world? Global Ecol. Biogeogr. 19, 769–782 (2010).
in a previous paragraph). To assure that this result is not an artefact 15. Egli, D. B. & Hatfield, J. L. Yield gaps and yield relationships in
of quick rise in yield making a finding of stagnation more likely, we central U.S. soybean production systems. Agron. J. 106, 560–567
compared stagnation likelihood across political units that were in the (2014).
third quartile of yield growth rates and had yield gaps that were on a 16. Hatfield, J. L. & Beres, B. L. Yield gaps in wheat: path to enhancing
trajectory to close within 30 years, and political units that were in the productivity. Front. Plant Sci. 10, 1603 (2019).
third quartile of yield growth rates. We repeated this analysis with a 17. Evans, L. T. & Fischer, R. A. Yield potential: its definition,
75th percentile confidence interval to test the sensitivity of the result measurement, and significance. Crop Sci. 39, 1544–1551 (1999).
to the number of census units identified as undergoing yield stagnation 18. van Dijk, M. et al. Reducing the maize yield gap in Ethiopia:
(Supplementary Table 8). decomposition and policy simulation. Agric. Syst. 183, 102828
(2020).
Reporting summary 19. Beddow, J., Hurley, T., Pardey, P. & Alston, J. Rethinking Yield Gaps
Further information on research design is available in the Nature Port- Staff Papers (College of Food, Agricultural and Natural Resource
folio Reporting Summary linked to this article. Sciences, University of Minnesota, 2015).

Nature Food | Volume 5 | February 2024 | 125–135 133

Article https://doi.org/10.1038/s43016-023-00913-8

20. Cassman, K. G., Dobermann, A., Walters, D. T. & Yang, H. 42. Baldos, U. L. C., Viens, F. G., Hertel, T. W. & Fuglie, K. O. R&D
Meeting cereal demand while protecting natural resources and spending, knowledge capital, and agricultural productivity growth:
improving environmental quality. Annu. Rev. Environ. Resour. 28, a Bayesian approach. Am. J. Agric. Econ. 101, 291–310 (2019).
315–358 (2003). 43. van Wart, J., Kersebaum, K. C., Peng, S., Milner, M. & Cassman, K. G.
21. van Dijk, M. et al. Disentangling agronomic and economic yield Estimating crop yield potential at regional to national scales. Field
gaps: an integrated framework and application. Agric. Syst. 154, Crops Res. 143, 34–43 (2013).
90–99 (2017). 44. van Bussel, L. G. J. et al. From field to atlas: upscaling of
22. van Oort, P. A. J. et al. Can yield gap analysis be used location-specific yield gap estimates. Field Crops Res. 177,
to inform R&D prioritisation? Glob. Food Secur. 12, 98–108 (2015).
109–118 (2017). 45. Grassini, P. et al. How good is good enough? Data requirements
23. Fischer, T., Byerlee, D. & Edmeades, G. Crop Yields and Global for reliable crop yield simulations and yield-gap analysis. Field
Food Security: Will Yield Increase Continue to Feed the World? Crops Res. 177, 49–63 (2015).
(Australian Centre for International Agricultural Research, 2014). 46. Rattalino Edreira, J. I. et al. Spatial frameworks for robust
24. Fischer, R. A. & Connor, D. J. Issues for cropping and agricultural estimation of yield gaps. Nat. Food 2, 773–779 (2021).
science in the next 20 years. Field Crops Res. 222, 121–142 (2018). 47. Rosenzweig, C. et al. Assessing agricultural risks of climate
25. Gollin, D., Morris, M. & Byerlee, D. Technology adoption in change in the 21st century in a global gridded crop model
intensive post-green revolution systems. Am. J. Agric. Econ. 87, intercomparison. Proc. Natl Acad. Sci. USA 111, 3268–3273 (2014).
1310–1316 (2005). 48. Guilpart, N. et al. Rooting for food security in sub-Saharan Africa.
26. Ray, D. K., Ramankutty, N., Mueller, N. D., West, P. C. & Foley, J. A. Environ. Res. Lett. 12, 114036–114037 (2017).
Recent patterns of crop yield growth and stagnation. Nat. Commun. 49. Anderson, W., You, L., Wood, S., Wood-Sichra, U. & Wu, W. An
3, 1293 (2012). analysis of methodological and spatial differences in global
27. FAOSTAT (FAO); https://www.fao.org/faostat/ Accessed cropping systems models and maps. Glob. Ecol. Biogeogr. 24,
17 Oct 2022. 180–191 (2015).
28. Grassini, P., Eskridge, K. M. & Cassman, K. G. Distinguishing 50. Fischer, G. et al. Global Agro Ecological Zones version 4 (GAEZ v4).
between yield advances and yield plateaus in historical crop GAEZ Data Portal https://gaez.fao.org/ (2021).
production trends. Nat. Commun. 4, 2918 (2013). 51. Chaloner, T. M., Gurr, S. J. & Bebber, D. P. Plant pathogen infection
29. Hertel, T. W., Burke, M. B. & Lobell, D. B. The poverty implications risk tracks global crop yields under climate change. Nat. Clim.
of climate-induced crop yield changes by 2030. Glob. Environ. Change 11, 710–715 (2021).
Change 20, 577–585 (2010). 52. Minoli, S., Jägermeyr, J., Asseng, S., Urfels, A. & Müller, C. Global
30. Alston, J. M., Beddow, J. M. & Pardey, P. G. Agricultural research, crop yields can be lifted by timely adaptation of growing periods
productivity, and food prices in the long run. Science 325, to climate change. Nat. Commun. 13, 7079 (2022).
1209–1210 (2009). 53. Wing, I. S., De Cian, E. & Mistry, M. N. Global vulnerability of crop
31. Alston, J. M. A Meta-Analysis of Rates of Return to Agricultural R&D. yields to climate change. J. Environ. Econ. Manag. 109, 102462
Research Report 113 (International Food Policy Research Institute, (2021).
2000). 54. Lesk, C. et al. Stronger temperature–moisture couplings
32. Fuglie, K., Ray, S., Baldos, U. L. C. & Hertel, T. W. The R&D cost of exacerbate the impact of climate warming on global crop yields.
climate mitigation in agriculture. Appl. Econ. Perspect. Policy 44, Nat. Food 2, 683–691 (2021).
1955–1974 (2022). 55. Lobell, D. B. & Field, C. B. Global scale climate–crop yield
33. Brisson, N. et al. Why are wheat yields stagnating in Europe? A relationships and the impacts of recent warming. Environ. Res.
comprehensive data analysis for France. Field Crops Res. 119, Lett. 2, 14002 (2007).
201–212 (2010). 56. Ray, D. K. et al. Climate change has likely already affected global
34. Chen, Y., Wang, P., Zhang, Z., Tao, F. & Wei, X. Rice yield food production. PLoS ONE 14, 1–18 (2019).
development and the shrinking yield gaps in China, 1981–2008. 57. Aggarwal, P., Vyas, S., Thornton, P., Campbell, B. M. & Kropff, M.
Reg. Environ. Change (2017) https://doi.org/10.1007/s10113-017- Importance of considering technology growth in impact
1168-7 (2017). assessments of climate change on agriculture. Glob. Food Secur.
35. Schauberger, B. et al. Yield trends, variability and stagnation 23, 41–48 (2019).
analysis of major crops in France over more than a century. Sci. 58. Sumberg, J. Mind the (yield) gap(s). Food Secur. 4, 509–518 (2012).
Rep. 8, 16865 (2018). 59. Müller, M. F. et al. Impact of transnational land acquisitions on
36. Mariani, L., Ferrero, A. & Cola, G. The evolution of cereal yields in local food security and dietary diversity. Proc. Natl Acad. Sci. USA
Italy over the last 150 years: the peculiar case of rice. Agron. J. 113, 118, 1–9 (2021).
3372–3383 (2021). 60. Zhang, W. et al. Closing yield gaps in China by empowering
37. Waha, K. et al. Multiple cropping systems of the world and the smallholder farmers. Nature 537, 671–674 (2016).
potential for increasing cropping intensity. Glob. Environ. Change 61. Sourour, A., Afef, O., Salah, B., Mounir, R. & Mongi, B. Y. Correlation
64, 102131 (2020). between agronomical and quality traits in durum wheat (Triticum
38. White Maize: A Traditional Food Grain in Developing Countries durum Desf.) germplasm in semi arid environment. Adv. Plants
(Centro Internacional de Mejoramiento de Maíz y Trigo (CIMMYT) Agric. Res. 8, 612–615 (2018).
and FAO, 1997). 62. Monfreda, C., Ramankutty, N. & Foley, J. A. Farming the planet: 2.
39. Cassidy, E. S., West, P. C., Gerber, J. S. & Foley, J. A. Redefining Geographic distribution of crop areas, yields, physiological types,
agricultural yields: from tonnes to people nourished per hectare. and net primary production in the year 2000. Glob. Biogeochem.
Environ. Res. Lett. 8, 34015 (2013). Cycles https://doi.org/10.1029/2007GB002947 (2008).
40. Yadav, O. P. et al. Genetic improvement of maize in India: 63. Fick, S. E. & Hijmans, R. J. WorldClim 2: new 1-km spatial
retrospect and prospects. Agric. Res. 4, 325–338 (2015). resolution climate surfaces for global land areas. Int. J. Climatol.
41. Yu, Y., Huang, Y. & Zhang, W. Changes in rice yields in China since 37, 4302–4315 (2017).
1980 associated with cultivar improvement, climate and crop 64. Oliver, J. E. Monthly precipitation distribution: a comparative
management. Field Crops Res. 136, 65–75 (2012). index. Prof. Geogr. 32, 300–309 (2010).

Nature Food | Volume 5 | February 2024 | 125–135 134

Article https://doi.org/10.1038/s43016-023-00913-8

65. Hengl, T. et al. SoilGrids250m: global gridded soil information methods with input from the other authors; D.K.R. contributed the
based on machine learning. PLoS ONE 12, e0169748 (2017). crop yield and area dataset. J.S.G., N.D.M. and P.C.W. outlined an earlier
66. Batjes, N. H. ISRIC-WISE Derived Soil Properties on a 5 by version of this work. All authors contributed to the writing of the paper.
5 Arc-minutes Global Grid (Version 1.0) (ISRIC—World Soil
Information, 2006). Competing interests
67. Nachtergaele, F. et al. Harmonized World Soil Database (Version 1.1); The authors declare no competing interests.
(FAO, 2007) https://www.fao.org/3/aq361e/aq361e.pdf Accessed
20 Aug 2007. Additional information
68. Portmann, F. T., Siebert, S., Döll, P. & Doell, P. MIRCA2000— Supplementary information The online version
global monthly irrigated and rainfed crop areas around the contains supplementary material available at
year 2000: a new high-resolution data set for agricultural and https://doi.org/10.1038/s43016-023-00913-8.
hydrological modeling. Glob. Biogeochem. Cycles https://doi.
org/10.1029/2008GB003435 (2010). Correspondence and requests for materials should be addressed to
69. Siebert, S., Kummu, M., Porkka, M. & Döll, P. A global data set of James S. Gerber.
the extent of irrigated land from 1900 to 2005. Hydrol. Earth Syst.
Sci. 19, 1521–1545 (2015). Peer review information Nature Food thanks Jerry Hatfield,
70. Harris, I., Osborn, T. J., Jones, P. & Lister, D. Version 4 of the CRU TS Martin van Ittersum and the other, anonymous, reviewer(s) for
monthly high-resolution gridded multivariate climate dataset. Sci. their contribution to the peer review of this work.
Data 7, 1–18 (2020).
71. Koenker, R. Quantile Regression (Cambridge Univ. Press, 2005). Reprints and permissions information is available at
72. Meinshausen, N. & Ridgeway, G. Quantile regression forests. www.nature.com/reprints.
J. Mach. Learn. Res. 7, 983–999 (2006).
Publisher’s note Springer Nature remains neutral with
Acknowledgements regard to jurisdictional claims in published maps and
This work has been funded by a variety of sources; data collection institutional affiliations.
efforts as well as early versions of the analysis were funded by
the Gordon and Betty Moore Foundation. J.S.G., D.K.R. and P.C.W. Open Access This article is licensed under a Creative Commons
received support from USAID grant number AID-OAA-L-14-00006, Attribution 4.0 International License, which permits use, sharing,
The World Bank award number 7192822 and the Institute on the adaptation, distribution and reproduction in any medium or format,
Environment. J.S.G. and P.C.W. received support from the Belmont as long as you give appropriate credit to the original author(s) and the
Forum/FACCE-JPI-funded ‘DEVIL’ project (NE/M021327/1). D.M. source, provide a link to the Creative Commons license, and indicate
received support from Project CLAND (ANR 16-CONV-0003) and if changes were made. The images or other third party material in this
Metaprogramme CLIMAE (INRAE). E.E.B. received support from article are included in the article’s Creative Commons license, unless
Biological Integration Institutes Grant NSF-DBI-2021898. J.J. received indicated otherwise in a credit line to the material. If material is not
support from the Minnesota Agricultural Experiment Station. The included in the article’s Creative Commons license and your intended
funders had no role in the study design, data collection and analysis, use is not permitted by statutory regulation or exceeds the permitted
decision to publish or preparation of the paper. use, you will need to obtain permission directly from the copyright
holder. To view a copy of this license, visit http://creativecommons.
Author contributions org/licenses/by/4.0/.
J.S.G. conceived and designed the research project and performed
the analysis; J.S.G., D.M., E.E.B. and N.D.M. developed the quantitative © The Author(s) 2024

Nature Food | Volume 5 | February 2024 | 125–135 135

 nature portfolio | reporting summary
Corresponding author(s): James Gerber, [email protected]
Last updated by author(s): Oct 5, 2023

Reporting Summary
Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection No software was used for data collection.

Data analysis All codes were developed in the Matlab and R programming languages. All codes associated with novel analysis are available at https://
doi.org/10.5281/zenodo.10234041
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A description of any restrictions on data availability
- For clinical datasets or third party data, please ensure that the statement adheres to our policy
April 2023

All weather, soil, and irrigation data used in this study are publicly available and sourced in references 66 and 68-73 of the methods, Crop yield and area data are
derived from publicly available sources (agricultural census and survey reports as identified in Reference 58) as described and further referenced in the methods
section. All data generated in the current study (Figures 1,2,4) as well as annual potential yield surfaces for ten crops over the period 1973-2012) can be

1
downloaded from https://doi.org/10.5281/zenodo.10234041. All data inputs to the study, as well as all results, are available upon request from the corresponding
author. Authors commit to full and timely cooperation with any validation studies.

nature portfolio | reporting summary

Research involving human participants, their data, or biological material
Policy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation),
and sexual orientation and race, ethnicity and racism.
Reporting on sex and gender N/A

Reporting on race, ethnicity, or N/A

other socially relevant
groupings

Population characteristics N/A

Recruitment N/A

Ethics oversight N/A

Note that full information on the approval of the study protocol must also be provided in the manuscript.

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Ecological, evolutionary & environmental sciences study design

All studies must disclose on these points even when the disclosure is negative.

Study description Study of crop yield gaps based on compilations of publically available census data

Research sample Census and survey data for yield and area for 10 crops over the period 1973-2012

Sampling strategy No sampling was performed to obtain the dataset. Sub-sampling of the dataset was performed to assess confidence intervals.

Data collection Crop data was collected by the second author.

Timing and spatial scale Dataset is global, data are annual over period 1973-2012. Spatial scale depends on the number of administrative units within each
country's agricultural census.

Data exclusions No data were excluded.

Reproducibility We have an entire ensemble of census data; there is no experiment to be repeated.

Randomization Bootstrapping methods based on random sampling of the data were used to generate confidence intervals.

Blinding n/a

Did the study involve field work? Yes No

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.
April 2023

2
Materials & experimental systems Methods

nature portfolio | reporting summary

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Clinical data
Dual use research of concern
Plants

April 2023