A half-century of climate change in major agricultural regions: Trends, impacts, and surprises

Significance

The productivity of staple crops is a key factor shaping the affordability of food and the amount of land and other resources used in agriculture. We synthesize evidence on how the weather faced by these crops has changed and how these changes have affected productivity. Most cropping regions have experienced both rapid warming and atmospheric drying, with significant negative global yield impacts for three of the five crops. Models can largely reproduce these changes and impacts with two important exceptions—they overstate warming and drying in North America and understate drying in most other temperate regions. These insights can help to guide adaptation efforts and model improvements.

Abstract

Efforts to anticipate and adapt to future climate can benefit from historical experiences. We examine agroclimatic conditions over the past 50 y for five major crops around the world. Most regions experienced rapid warming relative to interannual variability, with 45% of summer and 32% of winter crop area warming by more than two SD (σ). Vapor pressure deficit (VPD), a key driver of plant water stress, also increased in most temperate regions but not in the tropics. Precipitation trends, while important in some locations, were generally below 1σ. Historical climate model simulations show that observed changes in crops’ climate would have been well predicted by models run with historical forcings, with two main surprises: i) models substantially overestimate the amount of warming and drying experienced by summer crops in North America, and ii) models underestimate the increase in VPD in most temperate cropping regions. Linking agroclimatic data to crop productivity, we estimate that climate trends have caused current global yields of wheat, maize, and barley to be 10, 4, and 13% lower than they would have otherwise been. These losses likely exceeded the benefits of CO₂ increases over the same period, whereas CO₂ benefits likely exceeded climate-related losses for soybean and rice. Aggregate global yield losses are very similar to what models would have predicted, with the two biases above largely offsetting each other. Climate model biases in reproducing VPD trends may partially explain the ineffectiveness of some adaptations predicted by modeling studies, such as farmer shifts to longer maturing varieties.

Will climate change compromise the ability of humanity to feed itself? This question is among the most longstanding and prominent concerns related to climate change, due in part to the recognition that food shortages have historically led to humanitarian disaster, environmental destruction, and domestic and international conflict. Publications that projected climate change impacts on global and regional agricultural productivity appeared more than 30 y ago (1) and continue to this day. Nearly as much attention has been paid to the related question of what types of adaptations will help to reduce harms or increase benefits created by the shifting climate (2, 3).

Identifying future risks and actions that can mitigate these risks is an important research endeavor, yet less appreciated is the value of examining changes and impacts that have already been experienced in farms around the world. Although no actions can be taken to avoid what has already happened, understanding past changes is important for many reasons. First, the detection of significant changes is by itself a key line of evidence, independent from models or process understanding, that climate change is likely to affect food production systems (3). Second, by comparing what has occurred to what models predicted, we can better understand which aspects of models are trustworthy and which need further improvement. Third, quantifying the damage or benefits of historical emissions can provide a basis for the design of financial mechanisms to compensate those who were most harmed (4). Fourth, understanding the relative speed with which specific climate stresses are changing, or certain crops or regions are being affected, can help to inform and prioritize breeding and agronomic adaptation efforts.

For at least 20 y studies have documented agroclimatic changes occurring in agricultural areas around the world. Early work identified wheat, barley, and maize as experiencing losses from warming over the latter decades of the 20th century when aggregated to the global scale, with minimal effects of rainfall changes (5). Subsequent work showed that while some countries and crops have benefited from warming, most cases appear to show net damages from climate changes (6, 7). At the same time, increases in CO₂ have generally benefitted crops by increasing photosynthetic capacity (in C₃ crops) and transpiration efficiency (8). In some cases, these benefits may outweigh the losses associated with climate.

The goal of the current paper is to revisit the question of how historical climate trends have affected agriculture across the world, with an emphasis on major production regions. We focus on the past 50 y, as most of the historical global temperature increase occurred during this period (9) and crop and weather data are sparser in prior decades. We also consider how the observed changes compare to what models would have predicted, to ask whether impacts have been faster than expected and to identify aspects of models most in need of improvement.

Results

Most Agricultural Regions Are Experiencing Rapid Warming.

For nearly all locations and growing seasons, farms around the world have experienced unprecedented amounts of warming (Figs. 1 and 2 and SI Appendix, Fig. S1). Our preferred measure of warming is the linear trend in growing season temperature, expressed in units of historical SD (σ) around the trend. These units facilitate comparison between locations and seasons with different historical variability, with a trend of 1σ indicating that the average current growing season is warmer than ~84% of seasons 50 y ago, and a trend of 2σ indicating that the average season now would have been extremely rare (~2%) without the trend [these percentages ignore changes in variability over time, which we find to be small (SI Appendix, Fig. S2)]. Trends of 2σ or greater are common, with 45% of all maize areas and 32% of all wheat areas experiencing trends of at least 2σ (Fig. 2).

Fig. 1.

Climate trends for maize-growing regions and seasons, 1974–2023. (A–C) Observed trends, expressed as multiple of historical SD, for average temperature, vapor pressure deficit (VPD), and precipitation. Values show the average of trends from TerraClimate and ERA5-Land datasets. (D–F) Median simulated trend from 28 general circulation models (GCMs) participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6). (G–I) Percent of model simulations whose trend is below the observed trend. For temperature and VPD, areas with dark green in (G and H) (e.g., much of United States) had observed trends that were less positive than every GCM, whereas areas with dark brown (e.g., in Europe or China) indicate observed trends that were more positive than every GCM. For precipitation, dark green in (I) shows areas with observed trends with more wetting than any GCM, whereas dark brown shows areas with observed trends with more drying than any GCM. In all panels, only grid cells with more than 30 ha of maize according to the CROPGRIDS dataset are shown. In many regions, other summer crops (soybean, rice) have similar growing seasons as maize.

Fig. 2.

Distribution of observed and modeled climate trends for 1974–2023, for (A–C) maize and (D–F) wheat regions and seasons. Each panel shows an area-weighted cumulative distribution of trends, with crop area from CROPGRIDS used to weigh the trends across grid cells. Values on the y-axis indicate the percent of global maize or wheat area that experienced trends below the value on the x-axis. Observed trends (Obs) are the average across TerraClimate and ERA5-Land, whereas modeled trends (Mod) are the median across 28 GCMs. Gray lines show the distribution for individual GCMs.

Rather than considering changes at individual locations, which more closely reflect the experience of individual farmers, one can also consider the experience at the national or regional scale by aggregating conditions each growing season based on maps of where specific crops are grown. These aggregate conditions are most relevant for understanding impacts on national crop productivity and will tend to be less variable than any individual location since weather is not perfectly correlated across a country. For example, dry conditions for a growing season in one area may be balanced by wet conditions in another. For the top five most widely grown crops in the world, we calculate aggregated temperatures using gridded maps of crop area and country-specific growing seasons for the top producing countries (Materials and Methods). Trends in these area-weighted temperature measures are often greater than 2σ, in part because of the lower interannual variability resulting from aggregation, with some cases (EU maize, wheat, and barley) even reaching trends of 5σ (Fig. 3A). Trends of this magnitude indicate that even the coolest growing season in the present day is warmer than the warmest season that would have occurred 50 y ago (SI Appendix, Fig. S3).

Fig. 3.

Observed and modeled trends for 1974–2023 in area-weighted (A) average temperatures, (B) VPD, and (C) precipitation for the top five producing countries of each crop. All trends are expressed as multiples of σ, the SD of detrended values. Points indicate the observed trend according to the two climate datasets, colored lines show simulated trends for individual GCMs, and the black line shows the median GCM trend. Numbers in parentheses indicate the global production rank of the country for each crop. Observed trends generally fell within the model distribution, with the notable exceptions of Tavg and VPD for maize and soybean in the United States and VPD for several crops in EU and China.

Despite the dramatic warming for most crops and countries, a clear and important exception has been the experience in the United States and Canada. Most locations growing maize and soybean in the United States have experienced some warming, but with magnitudes well below 1σ (Fig. 1A). Similarly, temperature trends for wheat have been modest in many locations, and even a slight cooling has been seen in wheat areas of the Northern Great Plains and central Canada (SI Appendix, Fig. S1A).

Most Temperate, But Few Tropical, Systems Are Experiencing Rapid Atmospheric Drying.

High temperatures are well known to commonly co-occur with periods of high VPD, a key measure of atmospheric aridity. One advance since the earliest studies of agroclimatic trends is an improved recognition that the negative correlation between agricultural outcomes and high temperature can often be explained by high VPD (10–13). That is, although heat can (and often does) have many direct physiological effects on crop productivity, the existence of a strong correlation does not necessarily mean that heat is directly causing harm. Rather it may be that high VPD is leading to conditions of water deficit, which can be the primary cause of yield loss. This observation does not absolve warming of an important role but shifts it to an indirect effect associated with the fact that high temperatures raise the saturation vapor pressure (SVP) of air, thus causing higher VPD. However, because VPD is also influenced by humidity levels, it is clearly important to evaluate changes in VPD on their own.

We find that VPD in most agricultural regions is rising rapidly (Fig. 1B and SI Appendix, Fig. S1B), consistent with other recent studies examining global VPD trends (14). As with warming, some of the strongest trends are observed in Europe and East Asia. However, trends throughout the tropics are more muted, with increases rarely exceeding 2σ. Overall, the magnitudes of change are lower for VPD than for temperature, with only ~17% of maize areas and ~22% of wheat areas having trends above 2σ (Fig. 2 B and E). As with temperature, when we aggregate VPD to the national scale, trends are substantially larger relative to interannual variability, with trends exceeding 4σ for several crops in the EU (Fig. 3B). Unlike with temperature, a significant amount of area has exhibited negative VPD trends over the past half-century, including in the United States and India. These differences underscore the importance of considering VPD alongside temperature when assessing climate change impacts.

In contrast to both temperature and VPD, observed trends in growing season precipitation are typically small relative to variability (Fig. 1C and SI Appendix, Fig. S1C). Areas with increases and decreases are roughly equal for both maize and wheat (Fig. 2 C and F) and very few locations have trends exceeding 1σ in magnitude in either climate dataset. Aspects of subseasonal rainfall distributions and rainfall intensity would require daily data and could show additional trends (15, 16), but these tend to be less important than seasonal totals for driving large scale shifts in national crop output (17, 18) and we therefore leave examination of these and other aspects of agroclimatic conditions to future work.

Observed Trends Present Two Main Surprises Relative to Model Simulations.

We now turn to the issue of whether any of the trends described above are surprising, and what lessons these surprises could hold for improving models of climate impacts and adaptations. To do this, we compare observations to the ensemble of historical simulations run by GCMs that participated in the Sixth Assessment Report of the Intergovernmental Panel of Climate Change (IPCC AR6). These models were run with historical changes in greenhouse gases and other natural and anthropogenic forcings such as aerosols and represent our best understanding of how the climate system responds to anthropogenic activities.

Importantly, there are several aspects of the observed trends that the models capture very well. The strong positive trends in temperature, the positive but relatively smaller trends in VPD, and the muted trends in precipitation are all well captured by the median model projection (Figs. 1 D–F and 2 and SI Appendix, Fig. S1 D–F). The models also correctly predict that VPD changes in the tropics are much smaller and sometimes negative, in contrast with the increases in temperate systems. Even when examining trends for individual crops and countries, we find that the observed trends generally lie within the distribution of model predictions (Fig. 3).

Climate models therefore appear to be useful guides to the expected changes that will occur in agricultural regions, albeit with a large range of modeled trends in many cases (Fig. 3). However, two important aspects of the historical experience lie completely outside the range of model-predicted trends. First and foremost is the case of the United States, a major global producer of maize, soybean, and wheat. The observed trends are below every single climate model projection for most maize and soybean locations in the United States for both temperature and VPD (Fig. 1 G and H). Observed trends are also at the lower end of projections for wheat, although some models predicted less warming than observed. The lack of significant warming in the United States is often described as the “warming hole” and many hypotheses have emerged to explain its existence (Discussion).

The second main surprise that we consider consequential is the significant underestimation of VPD increases by most models in temperate systems. For example, many maize locations outside the United States, including in the EU, China, Argentina, and much of Africa, have VPD trends that exceed even the highest trend in models (Fig. 1H). Similarly, wheat areas in the EU, China, and South America have commonly experienced VPD increases greater than any model projection (SI Appendix, Fig. S1H). This mismatch between observed and modeled VPD trends is illustrated by the large gap between the fraction of maize and wheat areas that have experienced more than a 1σ increase in VPD (49% and 57%, respectively), and the much smaller fraction predicted by models (30% and 14%) (Fig. 2 B and E).

Climate Change Has Significantly Slowed Global Yield Progress for Barley, Wheat, and Maize.

Moving beyond the changes in observed agroclimatic conditions, we consider how much these trends likely mattered for total crop output at national and global scales. To estimate this, we build simple regression models that relate aggregate national agroclimatic conditions to crop yield, with standard controls for the shifting baseline of yields (Materials and Methods). These models explain a significant amount of variation for most crops and countries, often reducing out-of-sample mean squared error (MSE) by 20% relative to a model without weather, and sometimes by more than 50% (SI Appendix, Fig. S4). We use these models to estimate the effect of trends for each crop-country combination (SI Appendix, Fig. S5) and then aggregate across countries using historical production to weigh each country. Overall, we find that trends have been most impactful for barley and wheat, with the mean estimate of aggregate global effects equivalent to roughly 12% to 14% of yield loss for barley and 8 to 12% for wheat (Fig. 4; ranges give mean estimated impacts for the two climate datasets). Maize yields have also been hurt by trends, with aggregate impacts of ~4%. Soybean impacts are more ambiguous, with significant differences between the two climate data sources. Rice impacts have been small and close to zero. Aggregating across all crops based on total calorie production, we estimate aggregate climate impacts on calorie production of −5% (Table 1).

Fig. 4.

Estimate impacts (%) of climate trends for 1974–2023 on global yields of each crop. Mean estimate for each climate dataset (bars) along with 5–95% CI based on bootstrap estimates.

Table 1.

Estimated net impacts of climate trends and [CO₂] increases from 1974 to 2023 for the five most widely grown crops in the world

Crop	Global harvested area (Mha)	Global Production		Climate trend impact (%)			CO2 impact (%)	Net impact (climate + CO2)
		Million t	Trillion kcal	TerraClim	ERA5	Average
Wheat	217	767	2,562	−11.9	−8.0	−10.0	+6.4	−3.6
Maize	199	1,158	4,122	−4.2	−4.2	−4.2	0	−4.2
Rice	164	772	2,162	−3.4	+1.5	−1.0	+6.4	+5.5
Soybean	127	352	1,179	−1.6	−8.3	−5.0	+9.6	+4.7
Barley	49	151	501	−13.7	−11.9	−12.5	+6.4	−6.4
Total	756	3,200	10,526	−6.1	−4.8	−5.4	+4.3	−1.2

Area and production values refer to the 2018–2022 average based on FAO data (19). Climate effects are the mean of estimates shown in Fig. 4, and CO₂ effects are the mean of estimates from a recent meta-analysis of experiments (8). Impacts in the row labeled Total are weighted averages using total calorie production as weights.

The impacts of climate trends on global productivity have been close to those predicted using the median climate model projection, with differences between the two not significantly different from zero in most cases (SI Appendix, Fig. S6). For individual countries (e.g., United States), we often find significant disagreements between impacts estimated with observed vs. modeled trends (SI Appendix, Fig. S5). However, at the global scale, regions with more favorable than expected climate trends were apparently balanced by regions with worse-than-expected trends.

The effects of these climate trends can be compared to the direct effects of the CO₂ increases that have accompanied, and indeed caused much of, the climate trends. The combined effect of CO₂ and climate trends comes closer to measuring the net impacts of all anthropogenic emissions on crop yields, although a complete view would also need to consider effects of air pollutants such as ozone, nitrous oxides, and particulate matter, and corresponding trends in solar radiation. Over the 50 y study period, atmospheric [CO₂] increased by 91 ppm (20). According to a recent synthesis of open-air field experiments, the average yield effect of +200 ppm [CO₂] is 14% for wheat and rice and 21% for legumes including soybean (8). Although barley was not reported separately, experiments often observe similar responses as for wheat (21). For crops with a C₄ photosynthetic pathway (including maize and sorghum) no significant yield enhancement was observed on average, although a 20% mean effect was observed across four studies conducted under water deficit conditions. Assuming a linear response over this range of [CO₂] would imply that CO₂ increases in the past 50 y have directly boosted yields of wheat, barley, and rice by 6.4%, and of soybean by 9.6%. Maize yield benefits were certainly much smaller, and based on the average response across all experiments we assume here a negligible benefit. Overall, the net effects of CO₂ and climate trends we estimate are negative for global barley (−6.4%), maize (−4.2%), and wheat (−3.6%) yields and positive for soybean (4.7%) and rice (5.5%) (Table 1). Aggregating across all crops, we estimate a net combined impact of climate and CO₂ trends of −1.2%.

Discussion

Key Knowledge Gaps Remain for Understanding Agroclimatic Trends.

Our examination of trends in cropping regions underscores both the dramatic shifts occurring for most farmers, as well as the overall skill of climate models in projecting these shifts when driven with the relevant climate forcings. However, some important exceptions to these general patterns were observed. First, farmers in North America continue to be mostly spared from the warming and increased VPD that their counterparts in other temperate regions are experiencing. The persistence of this warming hole has benefitted global grain production given the large role the United States plays in maize, soybean, and wheat production.

Although the term warming hole is now more than 20 y old (22), considerable debate remains about the extent to which it will continue in the coming decades. Proposed hypotheses over the years have included a shift in summertime atmospheric circulation and rainfall patterns associated with ocean response to greenhouse gas forcing (22), increased evapotranspiration (ET) from a more productive agricultural and partially irrigated land surface (23), and natural variability of the climate system, for instance that associated with the Interdecadal Pacific Oscillation (24). The persistence of the warming hole for multiple decades, the fact that it relates only to daytime but not nighttime temperature, and the fact that it extends from April to October combine to suggest that at least part of the warming hole is a response to greenhouse forcing and thus would continue into the future (25). Most climate models fail to capture the observed pattern of sea surface warming, which can explain why they fail to reproduce the warming hole (25). At the same time, recent work has proposed that additional mechanisms may drive future drying in the Midwest, in particular, a decline in storm activity driven by a reduced north–south temperature gradient, combined with an enhanced north–south humidity gradient that drives export of moisture from the region (26). Given the importance of the region for global food supply, resolving these uncertainties is a key need for any prognosis of future climate impacts.

Second, models have consistently understated the rate of VPD increase throughout temperate cropping systems outside of North America, most notably in Europe, China, and South America. The critical role of VPD in shaping agricultural outcomes, and the wide areas over which VPD increase has been underestimated, means that resolving these discrepancies are arguably no less important than resolving the lack of warming and VPD change in North America. Several studies have noted a widespread decline in relative humidity (RH) and increase in VPD across a range of observational datasets (14, 27, 28). Although climate models typically predict a decline in RH with elevated greenhouse gases, owing to a slower rate of warming in oceans than in land (29, 30), the magnitude of RH decline is far greater in observations than models (31, 32). Model discrepancies are most severe in arid locations and times of year and may relate to a decline in ET from land systems that is not captured by models (31, 33). As with the warming hole, this is an area of active research with important implications for global agricultural impacts.

Crop Models Based on Climate Projections Likely Overstate Some Adaptation Benefits.

Climate trends matter not only for the effects on existing cropping systems, but for the way they expand or contract the options available to farmers. Many studies have simulated the ability of farmers to adapt to a warming climate, often suggesting substantial benefits from relatively minor adjustments. Two key shifts that have been prominent since the earliest modeling studies are a change in sowing dates and adoption of varieties with longer times to maturity (1, 34). The latter is intended to avoid a shortening of the season caused by faster accumulation of thermal time under a warmer climate. In a recent global study, shifts to longer cultivars led to average adaptation gains of 10% by the end of the 21st century, accounting for most of the 12% gain from overall adaptations (34).

However, as has been noted by others, one would expect farmers to already be adopting longer varieties if this was an effective adaptation, given the amounts of warming already experienced. Instead, farmers are often observed to be growing varieties with similar or even lower thermal requirements than older varieties in many regions. For example, a recent study in the United States examined maize hybrids sold to farmers in each county since 2000 and found that hybrid maturities have shortened over the majority of the Midwest, which the authors note “contrasts with widespread expectations of hybrid maturity aligning with magnitude of warming” (35). In Europe, a study of maize hybrids since 1950 found that a lengthening of thermal time in the vegetative phase was offset by a shortening of the grain filling phase (36), and a study of phenology trends over the same period in a range of crop and wild species found a significant shortening of time between flowering and maturity in crops, in contrast to wild plants (37). Again, these results appear at odds with modeling studies suggesting that longer hybrids in the region should be a major source of adaptation benefits (34, 38).

In our view, multiple factors can explain the discrepancy between projected and observed changes in cultivar maturity ratings, including the geographies seed companies focus breeding efforts on, which shapes the characteristics of elite hybrids available to farmers, and farmer consideration of labor cycles and grain drying costs. We hypothesize that one key factor in many regions is that models are simply too optimistic about the amount of water available to support longer growing varieties. For example, the aforementioned global study (34) relied on four GCMs to provide inputs into the crop model simulations. All four GCMs exhibited VPD changes in our study well below the observed increases in EU and China for both winter and summer seasons (SI Appendix, Fig. S7).

Switching to shorter varieties is a well-known response to drought risk because it reduces the chance that crops will face drought during grain filling. More work is needed to understand how much GCM errors in aridity trends are affecting crop model estimates of adaptation gains. At a minimum, the observed increases in VPD are almost surely pushing cultivar decisions in an opposite direction from the lengthening predicted by models. In a similar vein, Siebert and Ewert (39) observed that farmers in Germany are increasingly favoring earlier oat varieties, which, in the authors’ words, “is surprising because the expected adaptation to the warming trend should be in the use of cultivars with larger thermal requirements but it is in agreement with advice of the extension services encouraging the use of early developing varieties to avoid yield losses due to late season drought stress.”

Global Impacts of Climate Trends Are Apparent for Some Crops.

Our estimates of yield impacts of the observed climate trends suggest that for several key crops—wheat, maize, and barley—yields are significantly less today than they would have been without climate trends. Compared against a backdrop of rapidly increasing crop yields over the past half-century, which range from 69 to 123% for the crops considered here, impacts of 5 % to 10% may seem trivial. However, it is important to remember that demand has also rapidly increased, and a 5% yield loss on the margin can have important ramifications for prices and food security. For example, estimates of global calorie supply and demand elasticities indicate that a 5% shock to total calorie supply leads to a 30% increase in prices and a decline in consumer surplus of ~$180 billion (40).

We consider our yield estimates to be conservative for several reasons. We do not assess impacts of extreme events beyond the degree to which they are correlated with mean growing season conditions. Although the impacts of some extremes such as frost events are likely declining, the impacts of others such as heavy spring rains are likely increasing in many regions (41). We are also conservative in that we examine only the top five commodity crops, and therefore do not count climate impacts on other crops for which recent supply shocks and price increases have been prominent, including cocoa, oranges, and coffee. Moreover, we consider only the effects of climate trends on yields of harvested cropland area, whereas additional production losses are likely from climate-driven declines in the total amount of harvested area per year in most regions (42).

At the same time, we recognize that other factors could cause our estimates to be overly negative. Since yield sensitivities were estimated using the entire 50-y time series, they represent the average sensitivity. Adaptations such as adoption of new cultivars or irrigation could be causing the sensitivity to decline over time. Although analysis of this effect is beyond the scope of this paper, we note that a recent analysis of subnational data in multiple countries found yield sensitivities to be fairly stable over time, with more examples of yield sensitivity rising over time than decreasing (43). It is also possible that, because we use a static map of crop area, shifts in where crops are grown within a country could have reduced exposure to adverse weather compared to what we estimate (44). However, subnational data again indicate that the aggregate effect of this shift on national yield impacts is likely small relative to that of climatic trends (43, 45).

The overall picture of the past half-century is that climate trends have led to a deterioration of growing conditions for many of the main grain-producing regions of the world. The net yield impacts of these trends and the associated CO₂ increases have been negative for some crops and positive for others. At the global scale, yield impacts have accrued roughly linearly over time (SI Appendix, Fig. S8), and therefore it would be reasonable to expect that negative impacts will continue to grow at roughly the same pace in the coming decades. Yet two key questions remain whether the United States will continue to experience benign trends, and whether the rapid VPD increases in other temperate regions (EU, China, Argentina) will continue. In both cases, the observed trends have been outside of the range of climate model projections, undermining confidence in our ability to refine impact estimates and adaptation strategies. Beyond resolving these uncertainties, extending the analyses here to outcomes for nongrain crops and to trends in subseasonal extremes are both deserving topics, especially as the quantity and quality of data to study these aspects improve.

Materials and Methods

Historical Climate Observations.

We include two datasets on historical climate, TerraClimate (TC) and ERA5-Land, both of which provide gridded estimates for the past 50 y and were accessed in Google Earth Engine (GEE) (46). TC is a monthly dataset for global terrestrial surfaces at a spatial resolution of approximately 5 km, spanning from 1958 to the present (47). It includes a range of variables essential for analyzing climatic trends, including minimum and maximum temperature (Tmin and Tmax), VPD, and precipitation (P). ERA5-Land is a reanalysis dataset derived from the European Centre for Medium-Range Weather forecasts (ECMWF) ERA5 hourly data (48). It aggregates these data into daily values and provides a spatial resolution of approximately 11 km from 1950 to present. We utilized daily values for 2 m Tmin and Tmax, dew point temperature (Td), and precipitation. VPD for ERA5-Land was calculated as the difference between SVP and actual vapor pressure (AVP):

$$\mathrm{VPD}=\mathrm{SVP}-\mathrm{AVP},$$ [1]

where SVP is defined based on average daily Tmax (in °C) as

$$\mathrm{SVP}=0.6108\times \exp \left(\frac{17.27\times T\max }{T\max +237.3}\right),$$ [2]

and AVP is similarly defined based on Td:

$$\mathrm{AVP}=0.6108\times \exp \left(\frac{17.27\times T_d}{T_d+237.3}\right).$$ [3]

The above equations refer to VPD during the hottest hour of the day. As an alternative, we also used the daily ERA5-Land outputs to calculate VPD on an hourly basis and then calculate the average VPD for daylight hours. Trends for these average daytime VPD estimates were nearly identical to those using only Tmax (SI Appendix, Fig. S9).

Climate Model Simulations.

The NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6) dataset, available in GEE, comprises globally downscaled climate scenarios derived from GCM simulations conducted under the CMIP6 (49). This dataset has a resolution of approximately 28 km and spans from 1950 to 2100. We used data covering both the historical (pre-2015) and SSP 585 (for 2015–2023) scenarios to create a complete time series covering the 1974–2023 study period. Among the 34 different models available in GEE, we selected models that provided information on temperature, RH, and precipitation, resulting in 29 models. The “GFDL-CM4” model includes data for two distinct configurations, resulting in a total of 30 GCM realizations. Following initial analysis, two models were excluded: TaiESM1 due to misalignment between historical and future simulations, and HadGEM3-GC31-MM due to an anomalous drop in average temperature around 1980. In total, 28 GCM realizations were used, with a full list of model names given in SI Appendix, Table S1.

From the temperature and RH bands, we calculated the average temperature and VPD using Eq. 1, with AVP calculated from RH (which is available for more models) rather than Td:

$$\text{AVP}=\text{SVP} \times \frac{\text{RH}}{100}.$$ [4]

Crop Calendars and Masks.

Calculating crop exposure to weather requires defining the time of year the crop is grown. For grid-cell level analysis (Figs. 1 and 2 and SI Appendix, Fig. S1), we used the global gridded crop calendar at 0.5° resolution for maize and wheat developed by the WorldCereal team (50). The dataset provides planting or Start of Season (SOS) and harvesting dates or End of Season (EOS) dates, expressed as day of the year. For winter wheat the SOS captures the end of dormancy rather than the planting date. Since our analysis focuses on the critical months for crop development, we considered a three-month window starting one month after the reported SOS. Climate trends were very similar when shifting or expanding this window by one month. Although we refer to these as “maize” and “wheat” following the WorldCereal definitions, we note that maize shares a similar season to other summer crops (e.g., soybean, rice) in many countries, while wheat shares a season with other winter cereals (e.g., barley, rye).

Calculating Aggregate Crop Exposures.

For calculating country-level crop exposure, we first spatially aggregated the monthly gridded climate datasets using crop-specific masks from the CROPGRIDS dataset (51). This dataset, with a spatial resolution of ~5 km, provides area information for specific crops with a reference year of 2020. We utilized the wheat, maize, rice, soybeans, and barley layers, and calculated for each crop and country a weighted average for each month and each climate variable (T, VPD, P) using the crop-specific areas in the CROPGRIDS dataset as weights. The aggregated national monthly time series were then temporally aggregated using crop-specific growing seasons, which were derived based on a combination of calendars from the United States Department of Agriculture Foreign Agriculture Service (52) and inspection of the explanatory power of regression models (see next section) for different seasons. The defined growing seasons did not generally span the entire period from planting to harvest, but refer to the key months of weather that are most determinative for yield. The calendars used for each crop and country are shown in SI Appendix, Fig. S10.

Calculating Local Agroclimatic Trends.

Using the crop-specific growing periods as described above, we calculated growing season averages for each year in the 1974–2023 period for Tmin, Tmax, average temperature ((Tmin + Tmax)/2), VPD and P from the two observational datasets and all CMIP6 climate models included in the study. Each variable was then regressed against year, with the slope of the regression defined as the trend and the residuals of the regression used to calculate the SD (σ) of interannual variability. The trend was normalized by dividing by σ then multiplied by the total number of years in the analysis. This approach provides a standard measure of agroclimatic trends over the 1974–2023 period, expressed as the total change (as a multiple of σ) over the 50-y period.

Agricultural Statistics.

We utilized production, yield, and area statistics for wheat, maize, rice, soybeans, and barley from FAO’s Statistical Database (19). While most countries were analyzed individually, we grouped the 27 European Union member countries into a single region (EU) for further analysis. Countries were sorted by total production from 2018 to 2022 to identify the leading global producers for each of the five crops. For some analyses (e.g., Fig. 3) we present only results for the top five producers of each crop, whereas to calculate yield impacts (Fig. 4 and SI Appendix, Fig. S4) we used for each crop the top countries that collectively account for at least 85% of global production (with EU counted as a single country). This resulted in 11 countries for wheat and barley, 10 for maize and rice, and 3 for soybean.

Calculating Yield Impacts.

To estimate the impact of agroclimatic conditions on annual yields, we first fit the following ordinary least squares regression model for each crop ($\mathrm{k}),$ and country ($\mathrm{c})$:

$${\mathrm{Y}}_{\mathrm{c},\mathrm{k},\mathrm{t}}={\beta }_0+{\beta }_1\bullet \mathrm{t}+{\beta }_2\bullet {\mathrm{t}}^2+{\sum }_{\mathrm{j}=1}^{\mathrm{N}}{\theta }_{\mathrm{j}}\bullet {\mathrm{W}}_{\mathrm{j},\mathrm{c},\mathrm{k},\mathrm{t}}+{ϵ}_{\mathrm{c},\mathrm{k},\mathrm{t}},$$ [5]

where:

• •${\mathrm{Y}}_{\mathrm{c},\mathrm{k},\mathrm{t}}$: log of observed annual yield for crop $\mathrm{k}$ in country $c$ in year $t$
• •$\mathrm{t}$: year
• •$t^2$: quadratic term for time
• •${\mathrm{W}}_{\mathrm{j},\mathrm{c},\mathrm{k},\mathrm{t}}$: weather variables for crop $\mathrm{k}$ in country $c$ in year $t$
• •$N$: total number of weather variables in the model
• •${ϵ}_{\mathrm{c},\mathrm{k},\mathrm{t}}$: residual error term

The models were fit using the log of annual yield data as the dependent variable and weather variables as independent variables, controlling for both linear and quadratic time trends. The quadratic time trend ensures that long-term nonlinear trends unrelated to weather, such as from genetic progress, are accounted for.

Three alternative specifications were used to define the components of W:

• 1.Temperature model: W = {Tavg, diurnal temperature range (DTR = Tmax − Tmin), P}
• 2.VPD model: W = {VPD, P}
• 3.Combined model: W = {Tavg, DTR, VPD, P}

The first model corresponds to the most common formulation in the literature, with weather defined by T and P. The second more explicitly focuses on VPD as a driver of yield variability. The third combines the two, with the potential downside of high collinearity between Tavg and VPD leading to instability in the regression.

To evaluate the performance of each model, we randomly split the 50-y time series into 10 groups of 5 y. A 10-fold cross validation was then performed where each group is successively left out of the training and the model trained on the other nine groups was used to predict the held-out years. In addition to the three specifications for weather above, we also fit a model without weather as reference. Model performance was then measured as the mean reduction in MSE when using a model with weather compared to the reference model without weather. As shown in SI Appendix, Fig. S4, models with weather tended to reduce MSE, often by more than 25%. Across all crops and countries, there was not a single weather specification that consistently outperformed the others. For the main results presented in the paper, we use results from the combined model that uses both T and VPD. As an alternative, we also present results where for each crop and country we use the specification that performs best out-of-sample, finding that aggregate impacts are very similar (SI Appendix, Fig. S11). We note that the VPD model often outperforms the temperature model by a considerable margin (e.g., in China wheat, maize, and soybean), illustrating the limitations of using temperature-only models in some regions.

We recognize that all three specifications represent simple linear formulations that rely on growing season averages, whereas the literature often points to using nonlinear specifications or phase-specific weather. However, we justify this choice on four grounds. First, VPD itself is a strongly nonlinear function of temperature, and so many studies using nonlinear measures of temperature could instead use linear measures of VPD. Second, because we defined the growing season partly based on which months perform best, we often used only the key months for yield prediction rather than the entire season (SI Appendix, Fig. S10). Third, for any individual country the range of temperatures was typically less than 5 °C, such that a linear approximation to the response curve is more reasonable than it would be if fitting a model to data from multiple countries. Fourth, a simple specification helps to avoid overfitting, which is common when using only 50 y of observations, as seen in the fact that the combined model often does worse out-of-sample than the other specifications (SI Appendix, Fig. S4).

We also recognize that a quadratic time trend may fail to adequately control for technological changes over time in some countries, particularly those with very nonlinear time trends. As an alternative we also calculated models that control for time trends using a Hamilton filter (53), which replaces the year terms in Eq. 5 with four terms representing the four previous years of outcomes (in our case yields) as a flexible way to control for time trends. This model gave similar but weaker results for weather effects, with a lower reduction in out-of-sample MSE than the quadratic time trend model (SI Appendix, Fig. S12), as well as smaller aggregate impacts of climate trends (SI Appendix, Fig. S13). For that reason, we treat Eq. 5 as our preferred specification but note that the qualitative results were similar when using the Hamilton filter.

To calculate the impact of agroclimatic trends on yields, we first sample 50 y (with replacement) and fit the model (Eq. 5) to estimate the coefficients (${\theta }_{\mathrm{j}}$). These coefficients are then multiplied by the 50-y trends ($\mathrm{\Delta }{W}_j)$ of those variables, and we then sum the impacts across the different variables.

$${\mathrm{Impact}}_{\mathrm{c},\mathrm{k}}={\sum }_{\mathrm{j}=1}^{\mathrm{N}}{\theta }_{\mathrm{j}}\bullet \mathrm{\Delta }{\mathrm{W}}_{\mathrm{j},\mathrm{c},\mathrm{k}},$$ [6]

where:

• •${\mathrm{I}\mathrm{m}\mathrm{p}\mathrm{a}\mathrm{c}\mathrm{t}}_{\mathrm{c},\mathrm{k}}$: estimated yield impact for crop $\mathrm{k}$ in country $c$
• •$N$: total number of weather variables in the model
• •${\theta }_{\mathrm{j}}$: the regression coefficients for the $j$-th weather variable
• •${\mathrm{\Delta }\mathrm{W}}_{\mathrm{j},\mathrm{c},\mathrm{k}}$: the 50-y trend for the $j$-th weather variable for crop $k$ in country $c$

To estimate impact in percentage points, we convert from log units to percentage:

$${\mathrm{PercImpact}}_{\mathrm{c},\mathrm{k}}=100\ast \left(e^{{\mathit{Impact}}_{c,k}}-1\right).$$ [7]

A bootstrap resampling approach was used to calculate uncertainties for the estimates in Eqs. 6 and 7, with each calculated for 500 different samples (with replacement) of the data. This resulted in a distribution of estimates for the impacts in each country. To estimate the total global impact for each crop, we aggregated the country-level impacts (of the countries collectively accounting for up to 85% of global production) using a production-weighted approach. Specifically, the total global impact for crop $k$ was calculated as

$$\mathrm{Global} {\mathrm{PercImpact}}_{\mathrm{k}}={\sum }_{c=1}^C\left(\frac{P_{c,k}}{\sum P_{c,k}}\bullet {\mathrm{PercImpact}}_{\mathrm{c},\mathrm{k}}\right),$$ [8]

where:

• •$\mathrm{Global} {\mathrm{PercImpact}}_{\mathrm{k}}$: total global impact for crop $k$
• •$P_{c,k}$: production of crop $k$ in country $c$ (average for 2018–2022)
• •$\sum P_{c,k}$: total production of crop $k$ across all countries
• •${\mathrm{PercImpact}}_{\mathrm{c},\mathrm{k}}$: estimated impact for crop $k$ in country $c$

A distribution of global impacts was obtained by calculating Eq. 8 500 times, each time using a different bootstrap estimate for each individual country.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

All data used in this study are publicly available. The data and code necessary to reproduce the analysis and figures are available at https://doi.org/10.5281/zenodo.15078285 (54). Previously published data were used for this work (47–52, 49).