Greater Sensitivity to Drought Accompanies Maize Yield Increase in the U.S. Midwest

See allHide authors and affiliations

Science  02 May 2014:
Vol. 344, Issue 6183, pp. 516-519
DOI: 10.1126/science.1251423

Predicting Responses to Drought

The U.S. Corn Belt accounts for a sizeable portion of the world's maize growth. Various influences have increased yields over the years. Lobell et al. (p. 516; see the Perspective by Ort and Long) now show that sensitivity to drought has been increasing as well. It seems that as plants have been bred for increased yield under ideal conditions, the plants become more sensitive to non-ideal conditions. A key factor may be the planting density. Although today's maize varieties are more robust to crowding and the farmer can get more plants in per field, this same crowding takes a toll when water resources are limited.


A key question for climate change adaptation is whether existing cropping systems can become less sensitive to climate variations. We use a field-level data set on maize and soybean yields in the central United States for 1995 through 2012 to examine changes in drought sensitivity. Although yields have increased in absolute value under all levels of stress for both crops, the sensitivity of maize yields to drought stress associated with high vapor pressure deficits has increased. The greater sensitivity has occurred despite cultivar improvements and increased carbon dioxide and reflects the agronomic trend toward higher sowing densities. The results suggest that agronomic changes tend to translate improved drought tolerance of plants to higher average yields but not to decreasing drought sensitivity of yields at the field scale.

Drought is currently one of the main constraints to crop production in rainfed systems throughout the world, including in the United States. As a result, much breeding and agronomic research has been designed, at least in part, to improve performance under drought conditions (1). However, success in experimental fields does not always easily or completely translate to yield progress in farmers’ fields because drought characteristics in farmers’ commercial fields can differ substantially from trial conditions (2) and because different drought scenarios can favor different genotypes or management practices (3, 4).

In the United States, which typically supplies 40% of global annual maize production and 35% of global soybean production (5), several factors could be affecting drought sensitivity in farmers’ fields. On the positive side, increased use of low or no-till systems has likely increased soil moisture in dry years, and increasing CO2 concentrations typically lead to higher plant water-use efficiencies, thereby reducing drought sensitivity (6, 7). At the same time, in maize, modern genetics has facilitated increased sowing density because individual plants are better able to develop ears under stress (8, 9), roots are able to penetrate deeper and access more water (10), and genetically engineered pest resistance has reduced root damage from soil insects (11). All of these have contributed to historical gains in biomass and yields, but increased density can be detrimental under drought conditions because of excessive stress exposure for individual plants (12). Thus, the net effect of recent genetic, agronomic, and environmental changes on drought sensitivity remains an open empirical question.

An obstacle to measuring progress in farmers’ fields has been lack of accurate field-level data on both environmental conditions and yield performance that span a range of drought conditions and time. Here, we use a data set of field-level records on yields and sowing dates collected by the U.S. Department of Agriculture (USDA) Risk Management Agency since 1995 (Fig. 1) and combine that with 800-m-resolution daily weather data sets (see supplementary materials). These data are used to identify and group “location-year” combinations that have similar drought stress levels and measure yield progress over time for different stress levels. The study was conducted for maize and soybean, the two major field crops in the region.

Fig. 1 Study area and yield distributions.

(A) Map of main study region, with individual points showing locations of fields for maize in the USDA data. (B) Summary statistics of field-level maize (green) and soybean (orange) yields over the study period. Horizontal bands indicate sample median, and bars show interquartile range (25th to 75th percentile). Bu/Ac, bushels per acre.

Tracking yield progress at different levels of stress requires a clear and accurate measure of stress experienced by a crop. In field trials, stress levels are often quantified as the average yield across all cultivars grown at a given site and season, sometimes referred to as an environment index (13, 14). A plot of a cultivar’s yield versus environment index is then used to measure the sensitivity of that cultivar’s yield to stress. To define the environment index, we use the USDA data to calibrate a statistical model relating yields to weather. To allow for potential nonlinearities, asymmetries, and interactions between different variables, we employ multivariate adaptive regression splines (15). As candidate predictors for the regression splines, we use four weather variables [minimum and maximum temperatures, precipitation, and daytime vapor pressure deficit (VPD)] averaged over five successive 30-day periods spanning from 30 days before sowing to 120 days after sowing, resulting in a total of 20 candidate weather predictors (16). VPD is a widely used measure of atmospheric water demand that depends on air temperature and humidity and has a strong influence on plant growth rates in these systems (17, 18). We also repeated the analysis using the Palmer Drought Severity Index (16).

Figures S1 and S2 summarize the regression model calibrated for the maize and soybean data, respectively. The maize model selected six variables, with the most influential being VPD in the third month after sowing, which is typically July for a field sown in early May. An increase in VPD has a negative and nonlinear effect, with high values especially damaging, both of which are consistent with physiological understanding and previous empirical and crop modeling studies (17, 19). Some of the effects of high VPD may be associated with direct effects on pollen and grain set (20), but this mechanism appears to play a minor role in this region relative to VPD effects on water stress and overall plant growth rates (17). For soybean, five variables were selected, with VPD again the most important, and with precipitation playing a more important role than in maize.

The regression models are then used to calculate an environment index for each field in each year, which is defined as the predicted yield assuming a constant level of technology (year is fixed at 2000 for all predictions). As seen for maize in Fig. 2A (or fig. S3 for soybean), most years have the full range of conditions experienced throughout the region, but the exact locations of low or high stress change each year.

Fig. 2 Maize yields in different environments over time.

(A) The locations of environment index quintiles for maize for each study year. Red indicates the lowest environment index values, indicating the worst yield conditions, and blue indicates the best conditions. (B) Average maize yields (Bu/Ac) for each environment index quintile by year. Dashed lines show best-fit linear regression for each quintile. (C) Same as in (B) but expressed as fraction of trend yield under good conditions (i.e., highest quintile).

Mean yields for each quintile of the environment index exhibit significantly positive trends (P < 0.05) over time in absolute yields for all stress quintiles (Fig. 2B). In an absolute sense, yields under drought conditions have steadily improved over time, as emphasized in previous analyses of improved stress tolerance in maize (8, 9). However, yields have been rising more slowly at the lowest quintile than other stress levels, even when expressed as a percentage of yields at the start of the study period (1995), so that yields in the most stressed quintile have been declining relative to other quintiles (Fig. 2C).

A possible confounding factor in the above analysis is the fact that some locations are more frequently exposed to stress than others (Fig. 2A), and therefore trend differences that arise from location-specific factors that are unrelated to stress could either obscure or enhance any trend differences due to changes in stress sensitivity. As a robustness check, we therefore repeat the analysis of Fig. 2B, but first removing all time trends for each county and for all variables. This “detrended” analysis therefore considers whether positive yield anomalies when conditions are anomalously good have been rising over time. Consistent with Fig. 2B, we find that absolute anomalies are indeed growing over time (fig. S4).

Conditions of high VPD 61 to 90 days after sowing are the most important driver of a low environment index in this region, and we find that yield trends for the highest quintile of VPD are significantly lower than for the lowest quintile (P = 0.04) (Fig. 3A). Importantly, this finding is robust to the exclusion of 2012, which was a particularly low-yielding and high-VPD year (fig. S5). In contrast, the second most important factor for maize yields, late sowing, exhibits yield trends that are actually higher than trends for the more favorable early sow dates (Fig. 3B), although these differences are not significant. Thus, the overall slower progress under low-yielding conditions is the net effect of significantly slower progress for high VPD relative to low VPD, countered by slightly better progress for late relative to early sowing. Similar to maize, yield trends in soybean are higher in absolute value under good conditions than in poor conditions (fig. S5). However, soybean yield trends do not exhibit any significant dependence on VPD or other environment index predictors.

Fig. 3 Effects of vapor pressure deficit (VPD) and sowing date on maize yields.

(A) Yield response curves for VPD from multivariate adaptive regression spline model (see fig. S1 for explanation). Colors at bottom of figure indicate ranges for each quintile of that variable. (Inset) Yield trends for 1995 to 2012 corresponding to each quintile of VPD. (B) Same as in (A) but for sowing date. Error bars, mean ± 1 SE. Time trends for the highest quintile of VPD are significantly lower than for the lowest quintile (P = 0.04). Response curves and yield trends for other predictors and for soybean are shown in figs. S1, S2, and S5.

Repeating the analysis for the Palmer drought index rather than VPD leads to qualitatively similar results, with yield trends lower for the lowest quintile of the Palmer index (fig. S6). However, the Palmer index is a poorer predictor of yields than VPD, and as a result trend differences between low and high values of the Palmer index are less statistically significant. This agrees with previous work showing that the Palmer index, although a common variable used to measure drought, is actually a relatively poor predictor of crop water stress as simulated by a crop model (21).

As an alternative test of whether sensitivity to VPD has indeed been increasing over time in this region for maize, a cross-sectional regression model between yield and VPD was fit separately for each year. The coefficient on VPD becomes increasingly negative over time, with a significant negative time slope even if excluding 2012, and even if including county-fixed effects in the cross-section regressions to control for omitted variables (Fig. 4A and fig. S7). All other predictors of maize yields showed insignificant time trends in cross-sectional regressions (fig. S7).

Fig. 4 Changes in vapor pressure deficit and its impacts.

(A) Estimates of maize yield sensitivity to VPD 61 to 90 days after sowing from a cross-sectional regression for each year in the study period, along with best-fit trend lines with (solid) or without (dashed) including 2012 for computing the trend. Red dots indicate sensitivity estimates from APSIM simulations with sowing densities corresponding to the start and end of the study period. (B) Average July VPD in the study region for historical and projected periods. Dots show individual year observations, gray line shows linear trend for 1995 to 2012, black line shows mean VPD projected using 29 climate models, blue shading indicates 25th to 75th percentile of model projections, and gray shading indicates 5th to 95th percentiles. (C) Estimated impact of mean VPD projections on average maize yields using either constant yield sensitivity of –27.5% per kPa or a linear increase in sensitivity at the historical rate of 7% per kPa per decade.

Although this study is focused on the core Corn Belt states, USDA data are also available for other states. When our analysis is repeated for a broader group of Corn Belt states, the slower trends for high VPD persists (fig. S8). We also consider separately the western Corn Belt, where seed companies have arguably focused more of their efforts to test and release drought-tolerant seeds in the past decade (22), and one might expect more progress for drought conditions. Figs. S9 and S10 show the analysis repeated for rainfed maize fields in Nebraska, South Dakota, and Kansas. Similar variables are identified as important for maize in this region, with VPD 61 to 90 days after sowing the most critical, and precipitation in that period also important. The disparity between yield trends in low- and high-stress conditions is even more evident in this region than in the eastern states, with significantly slower yield progress under high VPD conditions.

One likely explanation for the increased sensitivity to VPD is the continuing trend toward denser sowing of maize crops. In Illinois, for example, average plant populations have gone from slightly under 24,000 plants per acre in 1995 to 30,000 by 2012 (a 25% increase over our study period) (23, 24). Similar trends have occurred in other Corn Belt states (23). To evaluate this mechanism more directly, we simulated maize yields using the Agricultural Production Systems Simulator (APSIM) model for a representative site under sowing densities of 24,000 and 35,000 plants per acre, or 6.0 and 7.5 plants per m2, respectively. Simulations under higher density exhibited higher average yields, but significantly higher sensitivity to VPD than simulations for lower density (fig. S11). The estimated yield response to high VPD 61 to 90 days after sowing in the simulations (–15% and –28% per kPa, respectively) agreed with the cross-sectional analysis of the USDA data (Fig. 4A). The importance of density is also suggested by the fact that soybeans, which show weaker evidence of increased drought sensitivity, have exhibited relatively constant sowing densities since 1995 in this region (24, 25).

Our results agree with the general notion that as farmers become more adept at removing all nonwater constraints to crop production, the sensitivity to drought generally increases (26). Given the dominant role of temperature (via VPD) in driving water stress in this region, our results are also consistent with the finding that heat sensitivity of maize yields in Indiana increased in recent decades (27). One implication is that climate change effects may be more severe than predicted by models that assume current crop genetics and management. Climate model projections indicate that July VPD for this region will become more severe, with an expected increase in average VPD of roughly 20% over the next 50 years (Fig. 4B), driven both by higher temperatures and reduced relative humidity. At current VPD sensitivity, these VPD trends would reduce yields by about 15% over the next 50 years. If maize yields continue to become increasingly sensitive to VPD, then yield losses from VPD trends could be as much as 30% (Fig. 4C).

Overall, we find no evidence to support the notion that farmers’ yields are becoming less sensitive to drought in the main maize- and soybean-growing states. Instead, we find evidence that drought sensitivity in maize, in particular sensitivity to high VPD, has steadily increased over the past 18 years. Whether the recent push by seed companies to develop and market drought-tolerant seeds will reverse this trend remains to be seen. It is clear that cultivar changes are not the only relevant factor for changes in drought sensitivity, because agronomic trends—often facilitated by the cultivar changes—can be as or more important. It is also clear that new, field-level data sets on farm performance can complement field-based experiments in efforts to understand changes in drought sensitivity in this important production region.

Supplementary Materials

Materials and Methods

Figs. S1 to S12

Database S1

References (28, 29)

References and Notes

  1. Materials and methods are available as supplementary materials on Science Online.
  2. Acknowledgments: We thank G. McLean for assistance with APSIM simulations and acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison, and the Working Group on Coupled Modelling of the World Climate Research Programme (WCRP) for their roles in making available the WCRP CMIP5 multimodel data set. This work was supported by NSF grant SES-0962625 and National Oceanic and Atmospheric Administration grant NA11OAR4310095. B.B.L. was supported by a Research Services Agreement from USDA’s Risk Management Agency, and G.L.H. by grant LP100100495 from the Australian Research Council. Data used in this study are available as supplementary materials on Science Online.
View Abstract

Stay Connected to Science

Navigate This Article