Simulating US agriculture in a modern DustBowl droughtMichael Glotter1†* and Joshua Elliott2,3†*

Drought-induced agricultural loss is one of the most costlyimpacts of extreme weather1–3, and without mitigation, climatechange is likely to increase the severity and frequencyof future droughts4,5. The Dust Bowl of the 1930s was thedriest and hottest for agriculture in modern US history.Improvements in farming practices have increased productivity,but yields today are still tightly linked to climate variation6 andthe impacts of a 1930s-type drought on current and future agri-cultural systems remain unclear. Simulations of biophysicalprocess and empirical models suggest that Dust-Bowl-typedroughts today would have unprecedented consequences, withyield losses ∼50% larger than the severe drought of 2012.Damages at these extremes are highly sensitive to temperature,worsening by ∼25% with each degree centigrade of warming.We find that high temperatures can be more damaging thanrainfall deficit, and, without adaptation, warmer mid-centurytemperatures with even average precipitation could lead tomaize losses equivalent to the Dust Bowl drought. Warmertemperatures alongside consecutive droughts could make upto 85% of rain-fed maize at risk of changes that may persistfor decades. Understanding the interactions of weatherextremes and a changing agricultural system is therefore criticalto effectively respond to, and minimize, the impacts of the nextextreme drought event.

The Dust Bowl of the 1930s is the most severe period of large-scale droughts in recent US history. Three distinct droughts definethe Dust Bowl era (1930–31, 1933–34 and 1936), accounting forthree of the six driest (for maize, soy and wheat) and hottest (formaize and soy) US growing seasons since 1901 (Fig. 1 andSupplementary Fig. 1). Persistent drought conditions resulted indouble-digit percentage loss in Great Plains wheat yields each yearfrom 1933 to 1939, with losses as high as 32% in 19337.Production declines led to lasting effects8,9 (such as populationshifts to urban counties10), and have motivated many to investigatethe impacts of the Dust Bowl11. Previous studies that focus on sitesin the Great Plains have confirmed the severity of a modern dustbowl for wheat7,12,13, but do not address the full effects on majorcommodity crops. Authors have recently called for furtherinvestigation of extreme weather events14 and food security11.

In this work we use the Dust Bowl as a case study to estimate theUS agricultural impacts of an extreme drought under current andfuture socioeconomic and environmental conditions. We use a par-allel version of the Decision Support System for AgrotechnologyTransfer (pDSSAT) biophysical process crop model to evaluatemaize, soybean and wheat for the contiguous United States usingobserved weather from the 2000s (a representative decade) andthe 1930s (a hot decade with severely reduced precipitation) (seeFig. 1 and Methods). We build on previous studies7,12,13 by includingthe upper Midwest, where most maize and soybean is grown

(Supplementary Fig. 2). We characterize the vulnerabilities of agricultureto Dust-Bowl-like weather, a critical step to better understandsystem-wide sensitivities and prepare for future extremes.

Simulations of crop yields driven by Dust Bowl weather indicatesevere loss to US agriculture from extreme droughts (Fig. 2, blue).Estimated maize and soy yields under 2012 technology are lowerin the median yielding year of the 1930s decade than in thelowest yielding year of the 2000s decade. Despite major advance-ments in farming practices from the 1930s to present day (mostnotably the significant increase in irrigation in the Great Plainsstates), simulations of the 1936 drought still result in losses of∼40% for maize and soy and ∼30% for wheat (Fig. 2, compare1936 blue dot with 2000s median). We test maize simulationsagainst an empirical model based on growing season temperatureand precipitation (following common practice in refs 6,15,16; seeMethods and Supplementary Fig. 3), and find similar results(Supplementary Fig. 4). Estimates from both the high-resolutionprocess-based and empirical simulations conclude that a 1936-typedrought would be ∼50% worse than the 2012 drought, when lossesreached ∼26–28%.

The impacts of extreme drought are highly sensitive to timingand location. The 1934 drought is commonly considered the mostsevere of the Dust Bowl era17, but yield simulations suggest 1936to be most damaging. Well-timed upper-Midwest precipitation in

0

5

10

15

20

25

His

togr

am d

ensi

ty

2 3 4 5Precipitation (mm day−1)

1901−20121930−19402000−2010

1993

1936

2012

1988

, 193

0, 19

76, 1

933,

1913

Average

Figure 1 | Histogram of historical JJA (June, July, August) average PGF(Princeton University Global Meteorological Forcing Dataset, Phase 2;ref. 41) precipitation weighted by present-day US maize production. Yearsat the tails of the distribution are identified in the text (ordered by rainfallamounts within bin, with the lowest amount at the bottom). Precipitationfrom the 1930s is shown in red and the 2000s (a more typical decade) inblue. For reference, the historical mean is marked by a dotted black line.See Supplementary Fig. 1 for all crops.

1Department of the Geophysical Sciences, University of Chicago, 5734 S Ellis Avenue, Chicago, Illinois 60637, USA. 2NASA Goddard Institute for Space Studies,2880 Broadway, New York, New York 10025, USA. 3Computation Institute, University of Chicago, 5735 S Ellis Avenue, Chicago, Illinois 60637, USA.†These authors contributed equally to this work. *e-mail: glotter@uchicago.edu; jelliott@ci.uchicago.edu

LETTERSPUBLISHED: 12 DECEMBER 2016 | VOLUME: 3 | ARTICLE NUMBER: 16193

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants 1

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

1934 (especially in July during the critical period around anthesiswhen crops are sensitive to drought) helped many maize and soyfarmers avoid the most significant damages, but offered no reliefto wheat farmers in neighbouring states further south and west(Supplementary Fig. 5). Previous studies of the Dust Bowl that arelimited to the Great Plains13 or select sites12 are unable to capturethis heterogeneity.

Warmer temperatures decrease drought-stressed maize and soyyields yet further (Fig. 2). We consider three temperature shiftsbased on observed and modelled warming under RepresentativeConcentration Pathway 8.5 between 1920–1949 and 1980–2009(ΔT1), 2010–2039 (ΔT2) and 2040–2069 (ΔT3). (See Methods andSupplementary Table 1 for average temperature shifts. It shouldbe noted that precipitation patterns could change as well, but thischange remains a strong point of disagreement amongst climatemodels.) Higher temperatures decrease maize and soy yieldsconsistently across all simulation years. Damage to maize under1936 drought conditions responds linearly to increasing tempera-ture, approximately doubling (from ∼40 to 80% losses) with 4 degreesof additional warming. Wheat yields, however, change little inhigher temperature scenarios. Wheat is less spatially concentratedthan maize and soy (Supplementary Fig. 2), and temperaturechanges have differing effects across regions: warmer temperaturesdecrease spring and southern-grown winter wheat but increasenorthern-grown winter wheat. Competing effects cancel whenaggregating to national yields.

Effects at warmer temperatures are not unique to the extremedroughts of the 1930s. We uniformly adjust temperature (in 1 °Cadditive intervals) and precipitation (in 10% multiplicative intervals)inputs separately to isolate the effects of each on simulated yields.Resulting yields are shown in coloured contours in Fig. 3, withestimated yields for 1936, the 2000s median and 2012 marked forcomparison. For the 1936 drought (blue dot labelled 1936), tempera-ture and precipitation anomalies impact yields similarly: temperatureanomalies alone lead to yield loss of ∼24%, and precipitationanomalies alone ∼19%. For an equivalent 1936 drought in mid-century (blue dot labelled 1936 +ΔT3), however, temperature anomaliesdominate and alone lead to yield loss of ∼68%. By mid-century,yields in a typical year (green dot labelled 2000s median +ΔT3) aresimilar to those of 1936 (blue dot labelled 1936). Higher temperaturesimpact yields through water stress (by means of increased evapotran-spiration) and physiological effects (for example, from acceleratedgrowing seasons, as described in Methods)18.

Advancements in management practices and higher CO2 con-centrations raise baseline yields, but do little to reduce relative loss

from extreme droughts (Fig. 2, red). Simulating yields using 2035management and CO2 conditions (as defined in Methods) haslittle impact on wheat (wheat growing seasons are often restrictedby rotations with spring crops), but increases maize and soy yieldestimates by ∼5–10%. For maize, management advancements takeadvantage of longer growing seasons. For soy, a C3 crop, yieldstake advantage of elevated CO2 concentrations. However, despiteabsolute yield gains in maize and soy, using 2035 conditions hasonly minimal relative impact. In the most damaging drought of1936, using 2035 conditions reduces relative losses by only ∼2–4%.

Advancements in farm management and policy (for example,irrigation and the formation of the Soil Conservation Service)have reduced the risk of damage from droughts, but some cropsand regions may still be vulnerable to changes (such as crop switch-ing, investment in capital or farm abandonment) with lasting socio-economic and environmental costs. To illustrate the extent ofproduction that may be vulnerable, we identify areas whereaverage rain-fed maize productivity loss across five consecutiveseasons falls below a critical threshold, estimated to be between 40and 60% of expected yields (see Methods). Under weatherconditions in the 1930s and a 40–60% loss threshold, we estimateonly 0.2–8% of rain-fed maize production vulnerable to long-termchange from consecutive droughts (Supplementary Fig. 6, left). Inwarming scenarios, however, even an optimistic threshold implieswidespread risk: under 1930s +ΔT3 weather, we estimate 23–85%of production vulnerable (Supplementary Fig. 6, right).

Irrigation is often considered a natural response to persistent dryconditions, but opportunities to switch fields from rain-fed toirrigated may be limited in vulnerable years. Present-day waterdemand is already depleting aquifers19, and we estimate that a1930s drought (even without any assumptions of warming) couldincrease water demand on currently irrigated areas yet further by10–20% (Supplementary Fig. 7). We recommend studies of coupledagricultural and hydrological models to better understand irrigationlimitations in persistent drought conditions (for example, ref. 20).

Other possible responses to long-term vulnerabilities fromconsecutive droughts are limited. Abandonment is one option,but present-day insurance and farm protection programmes(many of which were inspired by conditions in the 1930s) wouldlikely be sufficient to keep most land in production. (Farmerscould take some land out of production by participating indepression-era Farm Service Agency programmes, such as theConservation Reserve Program, which pays farmers to convert aportion of their land to forest/wetland or leave it fallow.) Anotherpossible adaptation strategy includes switching to generally lower

2000s 1930s ΔT1 ΔT2 ΔT3

Climate inputs

2000s 1930s ΔT1 ΔT2 ΔT3

Climate inputs

2000s 1930s ΔT1 ΔT2 ΔT3

1930s + 1930s +1930s +Climate inputs

0.2

0.4

0.6

0.8

1.0

1.2

1.4Fr

actio

n of

200

0s m

edia

n

1

2

3

4

Yield (tonne ha−1)

4

6

8

10

12

14

2012

1936 1930

1934

2012: tech. + CO2

2035: tech. + CO2

2012

1936

1934 1930

1

2

3

1936 1934

Figure 2 | Box plots of US modelled maize (left), soy (centre) and wheat (right) yields under 2012 (blue) and 2035 (red) technology (tech.) and CO2

concentrations. We drive crop models with 1930s and 2000s PGF climate, and shift temperature inputs to 1980–2009 (ΔT1), 2010–2039 (ΔT2), and2040–2069 (ΔT3) averages. Boxes represent the median and 25th–75th percentile for national yields, and circles represent yields outside this range.For comparison, the grey dotted lines mark the modelled yields from the 2012 drought.

LETTERS NATURE PLANTS

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants2

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

yielding21 but more drought-resistant varieties, or to more drought-resistant crops (for example, sorghum). However, crop switchingrequires significant capital investment and may be discouraged bygovernment-backed insurance programmes22. Finally, it shouldnot be assumed that seed distributors will develop new geneticlines that reduce negative impacts at high temperatures.Researchers have long attempted to breed for heat stress tolerance,but with little success to date23,24. Work towards these goals isongoing, along with other possible adaptation strategies such asadjustments to row width and seeding rates.

We use the 1930s as an analogue for present-day extremedroughts, but it is possible that the Dust Bowl droughts were abnor-mally severe even by modern standards. Recent studies suggest thatfeedbacks from crop failure and dust storms contributed to DustBowl conditions25,26 (see Methods for review), and managementand climate conditions that have changed considerably from the1930s to present day almost certainly means that a future dustbowl would be different than the past. We expect that improvedmanagement and oversight (such as the Soil Conservation Service,founded in response to the Dust Bowl) have reduced the risk ofextreme dust storms, but this should be evaluated alongsideexpectations of warming from anthropogenic activity. Indeed,climate change has been shown to significantly increase the risk ofmulti-decadal drought27, which may lead to droughts that are evenmore severe than those of the 1930s28. In some locations, climatechange could also lead to increased dust production29. We evaluatelarge climate model ensembles from the CESM LENS experiment toestimate extreme temperature and precipitation event probabilitiesin the recent past and future (see Methods and grey histograms inFig. 3 for ensemble projections). By mid-century, about 83% of

maize seasons are expected to experience average summer high temp-eratures hotter than the summer of 1936 (Fig. 3, compare the dark greytemperature histogram with the blue dot labelled 1936).

Changes in management practices from the 1930s to the presentday make modern US agriculture more robust in many ways butmore vulnerable in others. Improvements in management andtechnology have eliminated many non-climatic yield stresses anddramatically increased yield, but have also made agriculturetightly linked to changes in weather6 and potentially more vulner-able to extreme drought. In fact, it has been shown that morerecent droughts have had a larger impact on production thandroughts of the past30. Production is also more spatially concen-trated today than in the 1930s (Supplementary Fig. 2), whichcould further amplify system vulnerabilities. We show that observedrelative maize yield loss was similar in 1934/1936 and 1988/2012(Supplementary Fig. 8), but significantly less severe than simulatedyield loss for a 1930s drought under present or near-future con-ditions. (Note that this does not account for failed acres, whichwere ∼2–2.5 times higher than typical in 1934 and 1936, but only∼1.4 times higher than typical in 1988 and 2012. Differencesare explained by changes in cultivated area and expanded use ofirrigation in the most arid parts of the Great Plains.)

The Dust Bowl of the 1930s is best known for relentless dust stormsand devastating effects to agriculture and society. A present-day multi-year drought of the same severity as the 1930s would no doubt createsignificantly different (but not necessarily less damaging) impacts.Mitigation and adaptation measures—including efforts to reducefuture emissions—will therefore be critical to minimize losses indrought years and moderate long-term vulnerabilities.

MethodsYield simulations and model inputs. We simulate yields using a parallelizedversion of the commonly used Decision Support System for AgrotechnologyTransfer (DSSAT) crop modelling framework31,32. DSSAT is a biophysical processcrop model that has been used for more than 20 years by researchers across theglobe. The current version of DSSAT can simulate more than 40 different crops, andincludes representations for crop responses to weather, management (for example,fertilization, irrigation, planting/harvesting dates) and soil properties. We estimatemaize, wheat and soy yields throughout the contiguous United States at 5 arcminuteresolution ((1/12)° or ∼10 km) using DSSATmodules CERES-Maize (DSSAT V4.5),CERES-Wheat (DSSAT V4.6) and CROPGRO-Soybean (DSSAT V4.5), respectively.We use the parallelized System for Integrating Impacts Models and Sectors (pSIMS;ref. 33), which itself uses the Swift parallel scripting language34, to run DSSATconcurrently on several clusters in a highly parallelized process (henceforth,pDSSAT). We calibrate pDSSAT to match county- and state-level observed plantingand crop development dates (for example, flowering, harvest) from the USDepartment of Agriculture National Agricultural Statistics Service (NASS) cropprogress report35. The resulting simulations capture both yield variability andtechnological trends with high accuracy (Supplementary Fig. 8). See ref. 36 for amore detailed validation of pDSSAT, including under drought conditions. (Note thatpDSSAT historical calibration simulations, as shown in Supplementary Figs 3 and 8,are driven by the Daymet climate dataset37. Simulations in the remainder of themanuscript use the PGF or GSWP climate sources are described below.)

For this exercise, we use fixed planting and crop development dates from 2012 tobetter represent management practices under drought conditions, and useprojections of technological trends in 2035 for exploratory purposes (see below fordetails). We do not attempt to model 1930s management, when yields were aboutfive times lower than modern agriculture (Supplementary Fig. 8). We model rain-fedand irrigated yields at all locations, then aggregate to state and national levels basedon harvested hectares (c. 2000) from the Monthly Irrigated and Rainfed Crop Areas(MIRCA2000) dataset38. All modelled yields in this manuscript are thereforecombinations of both rain-fed and irrigated estimated production. We use anirrigation scheme where soil moisture is automatically returned back to 100% whenit falls below an 80% threshold. We define soil characteristics using the 1 km resolutionGlobal Soil Dataset for Earth System Modeling aggregated to 5 arcminutes39.

We drive pDSSAT with climate inputs (minimum and maximum temperature,precipitation and solar radiation) from 1930 to 1940 (henceforth, 1930s) and2000–2010 (henceforth, 2000s) from two sources: the Global Soil WetnessProject-Phase 3 (GSWP40), and the Princeton University Global MeteorologicalForcing Dataset-Phase 2 (PGF, updated from ref. 41). GSWP dynamicallydownscales the 1901–2010 twentieth Century Reanalysis (20CR42) and PGFstatistically downscales the 1901–2012 NCEP-NCAR Reanalysis; both then use CRU(Climate Research Unit43, which interpolates station observations to a 0.5° global

1936

+ΔT 1

+ΔT 2

+ΔT 320

12

10100%0% 000%0% −1

0100%0%

−2220%20

%

−330%0%

−44440%40

%40

%

−550%0%

−660%0%

−770%0%

2000

s med

ian

+ΔT 1

+ΔT 2

+ΔT 3

Temperature shift (°C)

−50

−40

−30

−20

−10

0

10

20

−1 0 1 2 3 4 5 6 7

30

40

50

Prec

ipita

tion

mul

tiplie

r (%

)

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

Yield change (%)

Historical

Near-future

Near-futureClimate model temperature projections

Clim

ate model precipitation

Historical

Figure 3 | National maize yield response surface to changes intemperature and precipitation. Yield simulations are driven with 2012technology and 2002 (the median yielding year in the 2000s decade) PGFweather, with weather inputs uniformly shifted for temperature in 1 °Cadditive intervals (from −1 to +7 °C) and for precipitation in 10%multiplicative intervals (from −50 to +50%). Contours identify resultingyield changes in 10% intervals. Circles mark yield estimates usingtemperature and precipitation for years 2012 (grey), 1936 (blue) and the2000s median (green), including estimates using ΔT1−3 temperature shifts.Bordering histograms show temperature and precipitation projections fromclimate models for the historical (1920–2005, light grey) and near-future(2037–2068, dark grey) periods, illustrating the likelihood of differentclimate conditions represented in the contour plot.

NATURE PLANTS LETTERS

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants 3

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

grid) to bias-correct temperature, but use different sources to bias-correctprecipitation. We use nearest-neighbour interpolation to downscale GSWP and PGFweather yet further from 0.5° to 5 arcminute resolution for crop model simulations.GSWP and PGF produce similar estimates of climate, but GSWP is biased slightlywarm and wet compared with CRU. Unless otherwise noted, we use PGF to driveyields, and use GSWP only as a test of robustness in Supplementary Information.

We examine the impacts of the droughts in the 1930s with and without changesin temperature. We use four global climate models (CESM1-CAM5, GISS-E2-R,MIROC5, BCC-CSM1-1) from the CMIP5 (Coupled Model IntercomparisonProject phase 5) multi-model ensemble archive to estimate changes in monthlymean minimum and maximum temperatures between climate model outputs from1920 to 1949 and several future periods: 1980–2009 (ΔT1), 2010–2039 (ΔT2) and2040–2069 (ΔT3). (We select global climate models from the CMIP5 archive thatboth meet the requirements of our methodology and are representative of the rangeof projected warming throughout the contiguous United States.) For future climateprojections, we use Representative Concentration Pathway 8.5. For each time period,we calculate the average change in each month across all four global climate modelsand apply the resulting shift to 1930s daily temperature from GSWP and PGFsources (equation 1):

Ti,t+Δtobs = Ti,t

obs + ΔTi,mmod (1)

where t indicates the day of the year, Δt the shift to a future time period, Tobs the1930s GSWP/PGF daily temperature data, and ΔTi,m

mod the average modelledtemperature shift for location i, month m. Resulting shifts are shown inSupplementary Table 1. Simulations in this manuscript are not projections ofdrought for a specific year, rather should be considered projections of drought underdifferent assumptions of temperature distributions. We do not shift precipitationvariables, and therefore assume that the droughts in the 1930s represent the tail-endof US precipitation distributions. We use 2000s weather as an example of arepresentative decade (Fig. 1 and Supplementary Fig. 1), and compare 1930s yieldsto 2000s median yields for benchmarking purposes.

Estimating future farming practices and technology. We estimate managementand technology practices in year 2035 to evaluate the ability of advancements infarming practices to mitigate yield loss from extreme drought. At each county, weextrapolate to 2035 linear trends in 1980–2012 planting dates (from NASS) andcrop-specific growth parameters (from calibrated pDSSAT simulations) for maize(Supplementary Fig. 9), soy (Supplementary Fig. 10) and wheat (SupplementaryFig. 11). We then shift 2035 extrapolations with 2012 anomalies from trend to betterrepresent management practices under drought conditions. (For example,early-season drought helps dry out soils faster after the winter thaw, giving farmersearlier opportunity in the field to operate heavy machinery and plant seeds.) We alsomodel effects from changing CO2, assuming a business-as-usual emissions scenariowith a 2035 CO2 concentration of 468 ppm. Estimates of 2035 management andtechnology are not intended to fully capture all adaptive responses to climate change;future estimates instead provide insight into crop responses to extreme droughtunder some advancements in crop management. For maize, changes in plantingdates and crop growth parameters are estimated to account for ∼26% of yieldincrease throughout the historical period44.

Maize. The CERES-Maize model specifies crop development primarily throughcumulated growing degree days (GDD)—growing season cumulated temperatureabove a defined baseline:

cumulated GDDi,t =∑t

n=0

Ti,nmin + Ti,n

max

2− Tbase (2)

where t indicates day of the growing season, n = 0 the start date of the GDDcalculation, and T minimum/maximum/baseline daily temperature(in degrees centigrade).

We characterize historical maize development through two parameters:p1 (GDD from emergence to flowering) and p5 (GDD from silking to maturity). Theresults from the pDSSAT model calibrations indicate an increase in historical p5 innearly all US maize-growing regions (Supplementary Fig. 9, top right), extending thegrain filling period of maize growth and thus increasing yields. Earlier planting dates(Supplementary Fig. 9, top left) offset this longer grain filling period, allowing forrelatively constant harvest dates that are important for economic profitability44.

We extrapolate historical trends in planting dates and crop development indices,p1 and p5, to estimate future maize management and technology (SupplementaryFig. 9, bottom). We constrain planting date slopes to be zero or negative, and 2035planting to occur no earlier than March 1. We finally restrict 2035 estimates of p1 to80–500 GDD, and p5 to 500–1,500 GDD.

Soy. Changes in temperature have also significantly impacted historical soy yields45.For maize, crop development is well-defined by GDD metrics. For soy, however,crop development is also strongly linked to day length46: CROPGRO-Soybeanspecifies crop development by photothermal indices, a combination of cumulatedtemperature and day length. Crop cultivars are then classified into 13 maturity

groups, ranging from cold weather/long-day varieties to warm weather/short-daytropical varieties. We use integer cultivar identification numbers to map linearlyacross these standard 13 maturity groups from 1 (coldest) to 161 (warmest). Cultivaridentification numbers have increased slightly over the historical period(corresponding with longer growing seasons) and are balanced by earlier plantingdates (Supplementary Fig. 10, top).

We extrapolate historical trends in planting dates and cultivar identificationnumbers to estimate future soy management and technology (SupplementaryFig. 10). Like maize, we constrain slopes in planting date trends to be zero ornegative, and 2035 estimates to occur no earlier than April 1. We do not constrainextrapolations in cultivar identification numbers.

Wheat. The CERES-Wheat model specifies crop development similar to CERES-Maize through cumulated growing degree days. (Differences in crop growing seasonsand crop traits however do require variations in maize and wheat GDD calculations.For example, CERES measures GDD using a base temperature of 1 °C for wheat and8 °C for maize.) US wheat consists of both winter and spring varieties: winter wheatis planted in autumn and harvested in spring, and spring wheat is planted in springand harvested in summer. (US wheat is dominated by the winter variety, accountingfor approximately two-thirds of total production.)

We characterize wheat management and technological trends through additivechanges in planting date and multiplicative changes (scale factors) in growthparameters p1 (flowering to terminal spikelet), p2 (terminal spikelet to end of leafgrowth), p3 (end of leaf growth to end of spike growth) and p5 (grain fillingduration). Decreases in p1, p2 and p3 follow decreases in p1 for maize, but wheatplanting dates trend later in the season and changes in grain filling duration vary byregion (Supplementary Fig. 11). We constrain 2035 winter wheat planting dates tooccur no later than November 15, and constrain slopes of growth parameter scalefactors to ±1% per year. These conditions are constraining only for selectoutlier locations.

Note that trends for wheat are largely driven by considerations for other crops.For example, where wheat is grown in a rotation with maize, trends towardlonger-season maize will drive later wheat planting. In fact, using trends of theparameters as shown in Supplementary Fig. 11 actually leads to a slight overallreduction in simulated yields by 2035, relative to 2012 technology (Fig. 2).

Empirical model. We develop a simple empirical model to validate and evaluateresults from pDSSAT simulations. We estimate maize yields using production-weighted national maximum temperature and precipitation, averaged over June, Julyand August (JJA). We follow common practice6,15,16 and define a functional formbased on both linear and quadratic terms of temperature and precipitation variables(equation 3). We calibrate model parameters by linearly regressing temperature andprecipitation predictor variables on detrended national NASS survey yields47 for1970–1999. We exclude from this regression year 1993, when anomalously highrainfall (Fig. 1) led to crop failure. To test the empirical model, we then extrapolate to2012 and compare against observed yields. We develop this empirical approach forexploratory purposes only; we do not make recommendations about the bestpredictor variables to estimate yields, nor compare the usefulness of empirical versusprocess methods for agricultural assessments. For these reasons, we model maizeyields only, the most sensitive of analysed crops to changes in weather:

Y = a1 + a2P + a3T + a4P2 + a5T

2 (3)

where Y is annual US maize yield, ai values are coefficients calibrated during1970–1999, and P and T are production-weighted JJA (June, July, August) averageprecipitation and temperature, respectively.

Production-weighted JJA precipitation and maximum temperature can reliablyestimate interannual changes in yield (Supplementary Fig. 3). The empirical modelcorrectly identifies drought-induced yield loss in 1983, 1988 and 2012, butconsistently underestimates the magnitude of loss. As expected, the empirical modelfails to identify yield loss in 1974 and 1993, when early season frost and flooding/lateplanting (respectively) lowered yields48,49. The empirical model does not considerthese effects. For in-sample years, the fraction of NASS variability that can bedescribed by the PGF empirical model (R2) is only ∼0.35, but rises to ∼0.73 whenexcluding years 1974 and 1993. For out-of-sample years, R2 is ∼0.84.

We detrend both NASS and pDSSAT simulations in Supplementary Fig. 3 toevaluate empirical simulations. For NASS yields, we detrend using a trendline foryears 1970–1999 (extrapolated to 2012) to both retain consistency with the empiricalsimulation extrapolation and to avoid effects from using an extreme 2012 endpoint.Similarly, for pDSSAT yields, we detrend using a trendline for 1980–1999(extrapolated to 2012). Because yield anomalies can be sensitive to detrendingmethodology, we recommend comparing yields with trends, as shown inSupplementary Fig. 8 (top) for NASS and pDSSAT simulations.

Vulnerability index. Consecutive droughts such as what occurred in the 1930s mayhave long-term effects on US farmers and the larger agrosystem. We define aland-use change vulnerability index to evaluate the extent to which sustained yieldloss from several near-consecutive droughts could initiate lasting changes in farmingpractices. We consider a 5-year rolling average of rain-fed yields at each location, and

LETTERS NATURE PLANTS

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants4

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

define vulnerable as any area where the 5-year mean falls below some threshold,defined here as between 40 and 60% of typical rain-fed yields (that is, where averagelosses exceed 40–60%). (In reality, many factors—including local conditions, debtburden, government programmes and insurance levels—could impact the actualvulnerability threshold, and vulnerability would likely vary by farm, even within thesame county.) For this exercise we focus on the 5-year mean simulated yields fromthe peak of the 1930s droughts in years 1933–1937, and compare to mediansimulated yields from the 2000s climate.

We estimate vulnerability under two scenarios: using 1930s weather and 2012technology and CO2, and using 1930s + ΔT3 weather and 2035 technology and CO2.We assume that yields <1 tonne ha–1 are left unharvested (failed crop), although thishas only small impact on vulnerability calculations. We develop a 40–60%vulnerability threshold through expert solicitation, but do not consider it a definitivebreaking point for local agriculture. Instead, we use the vulnerability index as ameans to evaluate sensitivities in the agricultural system that may be exposed bypersistent drought conditions (Supplementary Fig. 6).

Irrigation is often considered a natural response to dry conditions, so we evaluatevulnerabilities alongside potential changes to irrigation demand (Supplementary Fig. 7).Low precipitation in drought conditions means irrigated fields require additional water,and this additional demand could affect opportunities to convert fields from rain-fed toirrigated during vulnerable years. We simulate irrigated yields using 1930s and 2000sweather (without warming) and 2012 technology and CO2, and calculate changes inirrigation demand from the 2000s median to the 1933–1937 mean. We calculate waterdemand changes for currently irrigated maize and soy fields only. We do not estimatechanges in irrigation demand if currently rain-fed fields are suddenly irrigated.

Temperature and precipitation sensitivity analysis. In addition to the 1930s yieldsimulations described in the sections above, we adjust temperature and precipitationinputs separately to isolate the effects of each on simulated yields. We drive maizesimulations using 2012 technology and with a variety of climate combinations: weuse 2002 as the baseline weather (the median yielding year from the PGF simulationsin the 2000s) and shift temperatures uniformly from −1 to 7 °C (in single degreeintervals) and precipitation from −50 to +50% (in 10% intervals). Results for the99 simulations (9 temperature × 11 precipitation) are shown in Fig. 3. Although weadjust temperature and precipitation drivers separately in this analysis, it should benoted that temperature and precipitation are correlated, and precipitation is shownto significantly affect temperature in drought conditions50.

The estimated yields under extreme drought in Fig. 2 and those represented inFig. 3 are similar despite significant differences in simulation methodology. Wemark Fig. 3 with dots for 2002 and 1936 JJA temperature and precipitationanomalies, with and without ΔT1−3 shifts in temperature (as defined inSupplementary Table 1). For simulations in Fig. 3, we adjust both temperature andprecipitation using shifts that are homogeneous in space and time. For simulationsin the remainder of the manuscript, we adjust temperature using shifts from GCMoutput that are heterogeneous in space and time, and use precipitation observationsdirectly from the 1930s droughts. Unprecedented yield losses are robust to thedifferent methodologies for estimating extremes.

To provide some context for the T–P dimensions in Fig. 3, we add histograms forhistorical and near-future temperature and precipitation extracted from theCommunity Earth System Model (CESM) Large Ensemble (LENS) Project51. For thehistorical period, the LENS ensemble includes 30 simulations (differing only byinitial conditions) from 1920–2005, and we use the full sample (2,580 total years,labelled ‘historical’ in Fig. 3). For the near-future period, the LENS ensemble includes30 simulations from 2006–2100 using Representative Concentration Pathway 8.5,and we use years in the middle third of the simulations (950 total years, labelled‘near-future’ in Fig. 3). (Near-future histograms therefore nominally represent years2037–2068, following methodology from the 1930s +ΔT3 simulations.) Here,simulations are global and use a 1 degree spatial resolution version of the fullycoupled CESM1(CAM5) model.

Causes of the Dust Bowl. We provide a brief review of potential causes of the DustBowl. Many hypotheses exist that attribute the Dust Bowl to different physical andhuman drivers, and here we only describe select examples. (For more thoroughreview, see ref. 11.)

Historically, changes in tropical Pacific sea surface temperatures (SSTs) canexplain most prolonged dry episodes in the Great Plaines and Southwest US over thehistorical period52. In the 1930s, cool tropical Pacific SSTs (caused by a La Niñaevent in the early 1930s) and warm North Atlantic SSTs initiated dry conditions inthe western United States53. However, SST drivers alone cannot fully explain theextent and location of 1930s precipitation reductions54,55.

Recent work suggests that feedback from dust storms and crop failures also helpsexplain the hot and dry climate that characterized the Dust Bowl years. From 1870 tothe early 1930s, US settlers converted ∼30% of the Great Plains from (drought- anderosion-resistant) grasslands to row crops, including drought-sensitive wheat56.To make matters worse, the Homestead Act created a landscape of small acreagefarms that left little incentive for farmers to practice sustainable erosion control andsoil conservation8. Both factors led to widespread dust storms, giving the era itswell-recognized name. From a physical standpoint, atmospheric dust accumulationobstructed local incoming shortwave radiation, resulting in subsidence and adiabaticheating (for thermodynamic equilibrium). Subsidence, in turn, inhibited convection

and reduced precipitation25,26,55. Reduced precipitation (alongside poor soilmanagement) led to crop failure. Finally, crop failure hindered the transport of soilmoisture to the atmosphere by means of evapotranspiration and resulted inheightened surface temperatures. Only climate model simulations that include bothdust and crop abnormalities, in addition to unique SST patterns, are able toreproduce Dust Bowl conditions26. Simulations that do not include dustabnormalities are unable to reproduce precipitation observations, and simulationsthat do not include crop failure are unable to reproduce temperature observations25.

Data availability. The data that support the findings of this study are available fromthe corresponding author on request.

Received 24 May 2016; accepted 10 November 2016;published 12 December 2016

References1. National Mitigation Strategy – Partnerships for Building Safer Communities

(Federal Emergency Management Agency, 1995).2. Smith, A. B. & Katz, R. W. US billion-dollar weather and climate disasters:

data sources, trends, accuracy and biases. Nat. Hazards 67, 387–410 (2013).3. Billion-dollar weather and climate disasters. NOAA National Centers for

Environmental Information (NCEI) https://www.ncdc.noaa.gov/billions/ (2015).4. Sheffield, J. & Wood, E. F. Projected changes in drought occurrence under future

global warming from multi-model, multi-scenario, IPCC AR4 simulations.Clim. Dyn. 31, 79–105 (2008).

5. Wehner, M. et al. Projections of future drought in the continental United Statesand Mexico. J. Hydrometeorol. 12, 1359–1377 (2011).

6. Ray, D. K., Gerber, J. S., MacDonald, G. K. & West, P. C. Climate variationexplains a third of global crop yield variability. Nat. Commun. 6, 5989 (2015).

7. Warrick, R. A. The possible impacts on wheat production of a recurrence of the1930s drought in the US Great Plains. Clim. Change 6, 5–26 (1984).

8. Hansen, Z. K. & Libecap, G. D. Small farms, externalities, and the Dust Bowl ofthe 1930’s. J. Polit. Econ. 112, 665–694 (2004).

9. Hornbeck, R. The enduring impact of the AmericanDust Bowl: short and long-runadjustments to environmental catastrophe. Am. Econ. Rev. 102, 1477–1507 (2012).

10. Parton, W. J., Gutmann, M. P. & Ojima, D. Long-term trends in population,farm income, and crop production in the Great Plains. Bioscience 57,737–747 (2007).

11. McLeman, R. A. et al. What we learned from the Dust Bowl: lessons in science,policy, and adaptation. Popul. Environ. 35, 417–440 (2014).

12. Rosenberg, N. J. & Crosson, P. R. The MINK Project: a new methodology foridentifying regional influences of, and responses to, increasing atmospheric CO2and climate change. Environ. Conserv. 18, 313–322 (1991).

13. Rosenzweig, C. & Hillel, D. The Dust Bowl of the 1930s: analog of greenhouseeffect in the Great Plains? J. Environ. Qual. 22, 9–22 (1993).

14. Extreme Weather and Resilience of the Global Food System (The Global FoodSecurity programme, 2015).

15. Lobell, D. B., Schlenker, W. & Costa-Roberts, J. Climate trends and global cropproduction since 1980. Science 333, 616–620 (2011).

16. Urban, D., Roberts, M. J., Schlenker, W. & Lobell, D. B. Projected temperaturechanges indicate significant increase in interannual variability of US maizeyields. Clim. Change 112, 525–533 (2012).

17. Cook, B. I., Seager, R. & Smerdon, J. E. The worst North American drought yearof the last millennium: 1934. Geophys. Res. Lett. 41, 7298–7305 (2014).

18. Hatfield, J. L. et al. Climate impacts on agriculture: implications for cropproduction. Agronomy J. 103, 351–370 (2011).

19. Scanlon, B. R. et al. Groundwater depletion and sustainability of irrigation inthe US High Plains and Central Valley. Proc. Natl Acad. Sci. USA 109,9320–9325 (2012).

20. Elliott, J. et al. Constraints and potentials of future irrigation water availability onagricultural production under climate change. Proc. Natl Acad. Sci. USA 111,3239–3244 (2014).

21. Blum, A. Drought resistance, water-use efficiency, and yield potential—are theycompatible, dissonant, or mutually exclusive?Crop Pasture Sci. 56, 1159–1168 (2005).

22. Lewandrowski, J. & Brazee, R. Farm programs and climate change. Clim. Change23, 1–20 (1993).

23. Wahid, A., Gelani, S., Ashraf, M. & Foolad, M. R. Heat tolerance in plants:an overview. Environ. Exp. Bot. 61, 199–223 (2007).

24. Bita, C. E. & Gerats, T. Plant tolerance to high temperature in a changingenvironment: scientific fundamentals and production of heat stress-tolerantcrops. Front. Plant Sci. 4, 273 (2013).

25. Cook, B. I., Miller, R. L. & Seager, R. Amplification of the North American‘Dust bowl’ drought through human-induced land degradation. Proc. Natl Acad.Sci. USA 106, 4997–5001 (2009).

26. Cook, B. I., Seager, R. & Miller, R. L. Atmospheric circulation anomalies duringtwo persistent North American droughts: 1932–1939 and 1948–1957.Clim. Dyn. 36, 2339–2355 (2011).

27. Cook, B. I., Ault, T. R. & Smerdon, J. E. Unprecedented 21st century drought riskin the American Southwest and Central Plains. Sci. Adv. 1, e1400082 (2015).

NATURE PLANTS LETTERS

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants 5

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

28. Ault, T. R., Cole, J. E., Overpeck, J. T., Pederson, G. T. & Meko, D. M. Assessingthe risk of persistent drought using climate model simulations and paleoclimatedata. J. Climate 27, 7529–7549 (2014).

29. Munson, S. M., Belnap, J. & Okin, G. S. Responses of wind erosion toclimate-induced vegetation changes on the Colorado Plateau. Proc. Natl Acad.Sci. USA 108, 3854–3859 (2011).

30. Lesk, C., Rowhani, P. & Ramankutty, N. Influence of extreme weather disasterson global crop production. Nature 529, 84–87 (2016).

31. Jones, J. et al. TheDSSAT cropping systemmodel. Eur. J. Agron. 18, 235–265 (2003).32. Hoogenboom, G. et al. Decision Support System and Agrotechnology Transfer

(DSSAT) v.4.6 (DSSAT Foundation, 2015); http://dssat.net33. Elliott, J. et al. The parallel system for integrating impact models and sectors

(pSIMS). Environ. Model. Softw. 62, 509–516 (2014).34. Zhao, Y. et al. in 2007 IEEE Congress on Services 199–206 (IEEE, 2007).35. Crop Progress (NASS, 1995–2013); https://usda.mannlib.cornell.edu/

MannUsda/viewDocumentInfo.do?documentID=104836. Elliott, J. et al. Predicting agricultural impacts of large-scale drought:

2012 and the case for better modeling. Preprint at http://dx.doi.org/10.2139/ssrn.2222269 (2013).

37. Thornton, P. E. et al. Daymet: Daily Surface Weather Data on a 1-km grid forNorth America, Version 2. (Oak Ridge National Laboratory Distributed ActiveArchive Center, 2014); https://daac.ornl.gov/cgi-bin/dsviewer.pl?ds_id=1219

38. Portmann, F. T., Siebert, S. & Döll, P. MIRCA2000—Global monthly irrigatedand rainfed crop areas around the year 2000: a new high-resolution data set foragricultural and hydrological modeling. Glob. Biogeochem. Cycles http://dx.doi.org/10.1029/2008GB003435 (2010).

39. Shangguan, W., Dai, Y., Duan, Q., Liu, B. & Yuan, H. A global soil data set forearth system modeling. J. Adv. Model. Earth Sys. 6, 249–263 (2014).

40. Kim, H. Global Soil Wetness Project Phase 3 (2014); http://hydro.iis.u-tokyo.ac.jp/GSWP3/

41. Sheffield, J., Goteti, G. & Wood, E. F. Development of a 50-year high-resolutionglobal dataset of meteorological forcings for land surface modeling. J. Climate19, 3088–3111 (2006).

42. Compo, G. P. et al. The twentieth century reanalysis project. Q. J. R. Meteorol.Soc. 137, 1–28 (2011).

43. Harris, I., Jones, P., Osborn, T. & Lister, D. Updated high-resolution grids ofmonthly climatic observations—the CRU TS3.10 dataset. Intern. J. Climatol. 34,623–642 (2014).

44. Sacks, W. J. & Kucharik, C. J. Crop management and phenology trends in the USCorn Belt: Impacts on yields, evapotranspiration and energy balance. Agric. For.Meteorol. 151, 882–894 (2011).

45. Mourtzinis, S. et al. Climate-induced reduction in US-wide soybean yieldsunderpinned by region-and in-season-specific responses. Nat. Plants 1,14026 (2015).

46. Setiyono, T. et al. Understanding and modeling the effect of temperature anddaylength on soybean phenology under high-yield conditions. Field Crops Res.100, 257–271 (2007).

47. USDA National Agricultural Statistics Service (accessed 8 February 2016);http://www.nass.usda.gov/

48. Good, D. Corn and soybean production—some unfinished business. Universityof Illinois Extension (19 September 2013); http://web.extension.illinois.edu/cefj/news/news29698.html

49. Rosenzweig, C., Tubiello, F. N., Goldberg, R., Mills, E. & Bloomfield, J.Increased crop damage in the US from excess precipitation under climatechange. Glob. Environ. Change. 12, 197–202 (2002).

50. Livneh, B. & Hoerling, M. P. The physics of drought in the US central GreatPlains. J. Climate 29, 6783–6804 (2016).

51. Kay, J. et al. The Community Earth System Model (CESM) large ensembleproject: a community resource for studying climate change in the presence ofinternal climate variability. Bull. Am. Meteorol. Soc. 96, 1333–1349 (2015).

52. Seager, R., Kushnir, Y., Herweijer, C., Naik, N. & Velez, J. Modeling of tropicalforcing of persistent droughts and pluvials over western North America:1856–2000. J. Climate 18, 4065–4088 (2005).

53. Schubert, S. D., Suarez, M. J., Pegion, P. J., Koster, R. D. & Bacmeister, J. T.On the cause of the 1930s Dust Bowl. Science 303, 1855–1859 (2004).

54. Seager, R. et al. Would advance knowledge of 1930s SSTs have allowedprediction of the Dust Bowl drought?. J. Climate 21, 3261–3281 (2008).

55. Cook, B. I., Miller, R. L. & Seager, R. Dust and sea surface temperature forcing ofthe 1930s ‘Dust Bowl’ drought. Geophys. Res. Lett. 35 (2008).

56. Cunfer, G.On the Great Plains: Agriculture and Environment (Texas A&MUniv.Press, 2005).

AcknowledgementsThis research was performed as part of the Center for Robust Decision-making on Climateand Energy Policy (RDCEP) at the University of Chicago. RDCEP is funded by a grant fromNSF (no. SES-0951576) through the Decision Making Under Uncertainty program. M.G.acknowledges support of an NSF Graduate Fellowship (no. DGE-1144082). We thankC. Müller, A. Ruane and J. Winter—as well as the AgMIP (Agricultural ModelIntercomparison and Improvement Project) community—for valuable insight informulating the ideas for this research. We acknowledge the World Climate ResearchProgramme’s Working Group on Coupled Modelling, and we thank the climate modellinggroups for producing and making available their model output. For CMIP, the USDepartment of Energy’s Program for Climate Model Diagnosis and Intercomparisonprovides coordinating support and software infrastructure development in partnershipwith the Global Organization for Earth System Science Portals. Computing for this projectwas facilitated using the Swift parallel scripting language (NSF grant OCI-1148443), andcompleted in part with resources provided by the University of Chicago ResearchComputing Center.

Author contributionsM.G. and J.E. contributed equally to this work. Both authors designed and performed theexperiments, analysed the data, discussed the results, and wrote the paper.

Additional informationSupplementary information is available for this paper.

Reprints and permissions information is available at www.nature.com/reprints.

Correspondence and requests for materials should be addressed to M.G. and J.E.

How to cite this article: Glotter, M. & Elliott, J. Simulating US agriculture in a modern DustBowl drought. Nat. Plants 3, 16193 (2016).

Competing interestsThe authors declare no competing financial interests.

LETTERS NATURE PLANTS

NATURE PLANTS 3, 16193 (2016) | DOI: 10.1038/nplants.2016.193 | www.nature.com/natureplants6

© 2016 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.