- Access to this full-text is provided by IOP Publishing.
- Learn more
Download available
Content available from Environmental Research Letters
This content is subject to copyright. Terms and conditions apply.
Environ. Res. Lett. 14 (2019)114008 https://doi.org/10.1088/1748-9326/ab422b
LETTER
Mid-20th century warming hole boosts US maize yields
Trevor F Partridge
1,4
, Jonathan M Winter
1,2
, Lin Liu
3
, Anthony D Kendall
3
, Bruno Basso
3
and
David W Hyndman
3
1
Department of Earth Sciences, Dartmouth College, Hanover, NH, United States of America
2
Department of Geography, Dartmouth College, Hanover, NH, United States of America
3
Department of Earth and Environmental Sciences, Michigan State University, East Lansing, MI, United States of America
4
Author to whom any correspondence should be addressed.
E-mail: trevor.f.partridge.gr@dartmouth.edu
Keywords: climate impacts, warming hole, agriculture, machine learning
Supplementary material for this article is available online
Abstract
The Corn Belt of the United States, one of the most agriculturally productive regions in the world,
experienced a globally anomalous decrease in annual temperatures and a concurrent increase in
precipitation during the mid- to late-20th century. Here, we quantify the impact of this ‘warming
hole’on maize yields by developing alternative, no warming hole, climate scenarios that are used to
drive both statistical and process-based crop models. We show that the warming hole increased maize
yields by 5%–10% per year, with lower temperatures responsible for 62% of the simulated yield
increase and greater precipitation responsible for the rest. The observed cooling and wetting associated
with the warming hole produced increased yields through two complementary mechanisms: slower
crop development which resulted in prolonged time to maturity, and lower drought stress. Our results
underscore the relative lack of climate change impacts on central US maize production to date, and the
potential compounded challenge that a collapse of the warming hole and climate change would create
for farmers across the Corn Belt.
Introduction
The remarkable increase in the United States’agricul-
tural productivity during the 20th century (Dimitri
et al 2005)occurred largely against the backdrop of a
long-term global climate anomaly. During the second
half of the twentieth century, growing season tempera-
tures throughout most of the central US experienced
an anomalously cool period associated with a
phenomenon commonly referred to as the US warm-
ing hole (Hartmann et al 2013)as well as up to a 35%
increase in summer precipitation (Alter et al 2018).
This cooling and wetting may have created a more
agriculturally favorable climate over the central US as
the 20th century progressed. Given the global impor-
tance of US agriculture, the uncertainty in how future
climate change may impact agricultural productivity,
and the likelihood that the cooling trends associated
with US ‘warming hole’will disappear (Kumar et al
2013, Meehl et al 2015), understanding the historic
relationship between yield and climate variability is
critical to current and future adaptation strategies.
Here, we quantify the effects of the US warming hole
on maize, which is one of the most important crops
grown in the central US (i.e. the Corn Belt).
The US warming hole is characterized by an
abrupt shift toward cooler annually averaged tempera-
tures in 1958, followed by a prolonged cool period
over much of the central or southeastern US, depend-
ing on season (Rogers 2013, Mascioli et al 2017,
Partridge et al 2018). The exact timing and spatial
extent of the warming hole varies with season, shifting
from the southeastern US in winter and spring to the
central US in summer and autumn. From 1980 until
present, minimum temperatures throughout the US
have trended upward, however there has been no sig-
nificant trend in maximum temperatures within the
warming hole region (USGCRP 2017). The mechan-
isms responsible for the observed regime shift and
resultant lack of warming trend also vary by season
and region. Studies focusing on the southeastern US
suggest that the observed cooling trend is likely a result
OPEN ACCESS
RECEIVED
23 May 2019
REVISED
23 August 2019
ACCEPTED FOR PUBLICATION
6 September 2019
PUBLISHED
22 October 2019
Original content from this
work may be used under
the terms of the Creative
Commons Attribution 3.0
licence.
Any further distribution of
this work must maintain
attribution to the
author(s)and the title of
the work, journal citation
and DOI.
© 2019 The Author(s). Published by IOP Publishing Ltd
of internal variability in the climate system, such as
large-scale oscillations in the Pacific and Atlantic
oceans; (Meehl et al 2012, Partridge et al 2018),
increasing aerosol emissions during the early and mid-
twentieth century (Leibensperger et al 2012, Mascioli
et al 2017), and regional reforestation (Misra et al
2012, Ellenburg et al 2016). In the central US, agri-
cultural practices have been linked to the observed
summertime cooling trends. Increased irrigation
(Lobell and Bonfils 2008, Pei et al 2016, Deines et al
2017)and agricultural intensification (Mueller et al
2015, Alter et al 2018)have both been shown to have
remarkable influence on regional temperature and
precipitation patterns. In this paper, we focus on the
growing season of maize in the US Corn Belt, and thus
use the term warming hole in reference to the sum-
mertime cooling pattern observed over the central US
(figure 1).
Between 1930 and 2011, the yields within the aver-
age maize producing county in the US increased from
3.02 to 12.43 t ha
−1
, with annual growth rates peaking
at 3%−5% during the 1960s (Kucharik and Raman-
kutty 2005). Although increased fertilizer applications,
improved cultivar genetics, and innovations in agri-
cultural technology are responsible for much of the
yield improvement, decreased interannual climate
variability (Thompson 1986, Baker et al 1993)and
prolonged growing seasons with reduced heat
extremes since 1980 (Butler et al 2018)have also con-
tributed to the impressive maize yield gains.
Here, we investigate the relationship between maize
yields and climate through a multi-model approach,
combining results from a process-based and statistical
crop model. Process-based crop models can capturethe
physiological response of crops to variations in climate
and management conditions because they explicitly
simulate plant growth and development throughout
the growing season. Process-based models have been
extensively validated globally (Basso et al 2016), how-
ever, they require site-specific input on soil, climate,
genotypes and management which can often be limited,
making them more uncertain (Lobell and Asseng 2017).
Statistical crop models are computationally inexpen-
sive, require little to no field-based calibration data, and
provide reasonable estimates of yield variability (Lobell
and Burke 2010,Jeonget al 2016, Lobell and Asseng
2017). However, they lack physiological mechanisms
and thus cannot provide the mechanistic insights
offered by process models. We apply two separate mod-
els to simulate long-term rainfed and irrigated maize
yields: the System Approach to Land Use Sustainability
(SALUS)model (Basso et al 2006), and an empirical
machine learning algorithm calibrated to county-level
annual yields. We then statistically remove the warming
hole from the climate record, producing counterfactual
climate scenarios to drive the crop models. Finally,
we analyze the scenario outputs and quantify the net
yield benefit provided by the warming hole, as well
as the proportional benefits provided individually by
the reduced temperatures and increased precipitation
within the warming hole.
Methods
Aflowchart of the methodology described below is
included in the supporting information as figure S1 is
available online at stacks.iop.org/ERL/14/114008/
mmedia.
Data
We acquire historical 1930–2011 county level yield,
production, planting date, and plant growth stage data
for maize from the US Department of Agriculture
National Agricultural Statistical Service (USDA
NASS). Climate data is derived from daily gridded
Figure 1. Growing season temperature trends and average annual maize production. (a)County level trend in growing season average
temperature from 1930 to 2011 over cultivated land. Counties with no cultivated land are masked in grey. Growing season varies by
state based on USDA NASS planting data (approximately early May to September for Corn Belt).(b)Average annual 1990 to 2011
maize production in metric tons by county. Domain of focus (−82°Wto−102°Wand 37°Nto47°N)is identified by the black box.
2
Environ. Res. Lett. 14 (2019)114008
values of maximum temperature (Tmax), minimum
temperature (Tmin), daily precipitation (Pr), and
incoming shortwave radiation from the Livneh hydro-
meteorological dataset (Livneh et al 2013,2015). The
Livneh hydrometeorological dataset contains meteor-
ological observations from ∼20 000 weather stations
across the US and flux estimates gridded to a 1/16°
scale from 1915 to 2015 for the continental US and
parts of Canada and Mexico. Within Livneh, short-
wave radiation is estimated using the MTCLIM algo-
rithm in the Variable Infiltration Capacity model
(Maurer et al 2002). We note that our analysis ends in
2011 because Livneh shortwave radiation data are only
available for 1915–2011. We approximate county
irrigation by the fraction of average historical
1930–2005 area equipped for irrigation (AEI, Siebert
et al 2014)relative to a county’s cultivated area
(cultivated area calculation explained below).
Gridded Tmax, Tmin, and Pr values are aggre-
gated to the county level for the random forest. To
reduce the influence of non-representative grid cells
on county aggregates, such as cells covering mountai-
nous or metropolitan areas, we overlay the gridded cli-
mate data with a cultivated area mask developed using
the 2017 Cultivated Layer dataset from USDA-NASS
(Boryan et al 2012). The 2017 National Cultivated
Layer uses national cropland data layers from 2013 to
2017 to estimate cultivated area at 30 m binary resolu-
tion across the United States. We interpolate the culti-
vated layer to a 1/16°grid, and aggregate climate data
to the county level using grid cells corresponding with
cultivated land. As total land area devoted to agri-
culture has decreased by more than 400 000 km
2
dur-
ing the 20th century (Brown et al 2005; World
Bank 2019)we argue that using recent cultivated area
estimates are likely conservative.
Removing the warming hole from the climate record
We develop counterfactual (no warming hole)scenar-
ios using two techniques: a delta-change method
(hereafter Delta)and a quantile mapping method
(hereafter QMap). Both methods are commonly used
to bias correct global climate model data for use in
regional assessments (Maraun and Widmann 2018,
Winter et al 2019). The QMap approach produces a
nearly identical shift in the mean value as the Delta
approach, but also adjusts the variance of the post-
1958 distribution to pre-1958 values (figure S2). Thus,
including both methods allows for a direct compar-
ison of how a mean shift in climate affects maize yields
relative to a shift in both the mean and variance.
The Delta approach effectively adjusts observa-
tions by some change factor or delta, calculated as the
difference between a target and reference distribution,
additively for temperature and multiplicatively for
precipitation. We remove the observed 1958 regime
shift associated with the onset of the warming hole
(Rogers 2013, Partridge et al 2018)by adjusting daily
values of Tmax, Tmin, Pr, and SW by the difference or
ratio between the pre-1958 (1915–1957)and post-
1958 (1959–2011)long-term monthly means. This
approach preserves interannual variability of the post-
1958 data but adjusts the mean to pre-1958 values
(figure S2).
The QMap scenario is constructed in four steps
(Brocca et al 2011). First, we divide Tmax, Tmin, Pr,
and SW into pre- and post-1958 distributions. Second,
we find the difference between the pre- and post-1958
values for each post-1958 percentile. Third, we fita
polynomial function to the post-1958 values and the
calculated differences between the pre- and post-1958
values by percentile. Finally, we adjust the post-1958
values using the calculated differences from the poly-
nomial function. For both methods, we mandate that
all adjusted values lie within the bounds of historical
records. Any values (less)greater than the maximum
(minimum)historical value are set to the maximum
(minimum)historical value and any negative values of
Pr and SW are set to 0.
In addition to the Delta and QMap scenarios, we
create three additional scenarios to address the sensi-
tivity of the model outcomes to the time period inclu-
ded. To account for the inclusion of the Dust Bowl, an
extraordinarily warm time period, in the reference dis-
tribution, we create a separate counterfactual climate
scenario with the Dust Bowl years (1930–1939)
removed from the reference period. Additionally, we
test the sensitivity of onset of the warming hole by
creating two counterfactual scenarios with break-
points at 1953 and 1963, 5 years earlier and later than
the observed warming hole onset, respectively.
Model framework
In this study, we use the well-validated SALUS model
(Basso et al 2006)and the Treebagger random forest
algorithm in Matlab (Breiman 2001)to simulate long-
term rainfed and irrigated maize yield. Both models
are calibrated to county-level reported yields from
1930–2011 (random forest)and 1956–2011 (SALUS).
A detailed description of the SALUS model can be
found in the supplemental text S1. SALUS simulations
do not extend earlier than 1956 because this would
require assumptions based on limited to no cultivar
and management data. Our analysis spans the Corn
Belt, from −82°Wto−102°W and 37 °Nto47°N. To
ensure that this analysis includes predominantly
agricultural counties, any county with less than 30
years of yield data or fewer than 10 000 harvested acres
on average is excluded, resulting in 794 counties. For
both models, we assume static planting dates within
each state, set to the 1979–2011 average statewide
median reported planting date. The length of the
growing season is also static for the random forest, as
described below. However, maize harvest dates in the
SALUS model vary dynamically by state and year to
capture the maize growth cycle at maturity.
3
Environ. Res. Lett. 14 (2019)114008
For SALUS validation, we simulate both rainfed
(counties with AEI fraction <5%; n=590)and irri-
gated (counties with AEI fraction >60%; n=4, all in
Nebraska)maize yields. Automatic irrigation is trig-
gered when plant available water content drops below
50% of maximum soil water holding capacity. After
triggering, water is then added to the soil until plant
available water content reaches 70% of maximum soil
water holding capacity. We account for some of the
technological changes in the simulated 56 years by
varying management practices. The crop genetic coef-
ficients are obtained using an inverse calibration pro-
cedure to match state-wide average reported yields to
yields from maize cultivar and management practices
used in the 56 year simulated period. Specifically, we
vary maize cultivars from low yielding maize cultivars
in 1950s to high yielding cultivars in recent decades,
planting densities from 4 to 10 plants m
−2
, and nitro-
gen fertilizer application rates from less than 10 to
120 kg ha
−1
. We extract the dominant soil informa-
tion for each map unit polygon in the Soil Survey Geo-
graphic Database as soil input for each studied county
(USDA/NRCS 2018)and ran a unique SALUS simula-
tion for each cell of the 1/16°climate grid. For each
simulation year and grid cell, we calculate the AEI-
weighted average yield of the simulated rainfed and
irrigated yields. We then average the weighted annual
yield across the grid cells in each county.
SALUS allows us to quantify the mechanistic
effects of the warming hole on altered yields. Here we
use both time to maturity, calculated directly by
SALUS, as well as a drought stress metric that is the
average of a daily drought stress factor over the 20-day
period from vegetative-tasseling to early grain filling,
as defined by SALUS. The drought stress factor is
computed as
=- /
F
WW1
,
drought supply demand
where
W
supply is the total water available to the plant and
W
demand
is the potential evaporative demand, calcu-
lated using the Ritchie equations (Ritchie 1972). Thus,
values of
=
F
1
drought
indicates complete drought
stress, and
=
F
0
drought
indicates no drought stress.
Random forest is a non-parametric supervised
ensemble-based machine learning algorithm consist-
ing of a large number of individual decision trees
trained on random subsets of the input data and pre-
dictor variables (Breiman 2001). Random Forest algo-
rithms have been shown to consistently outperform
multiple linear regression methods in applications
including climate and agriculture (Jeong et al 2016).In
this study, we train a random forest of 700 trees on
1930–2011 county yield data for the 794 counties
using nine predictor variables; three non-climate and
six climate. For each decision tree, we mandate the leaf
size, or the number of observations per leaf (each split
in the decision tree produces two leaves), to be more
than 10 to avoid overfitting. The non-climate variables
include: (1)year, (2)average county latitude, and (3)
the average historical fraction of cultivated area that is
equipped for irrigation (AEIFrac). The six climate
variables are: (1)growing season growing degree days
(GDD, see supplemental text S2),(2)growing season
killing degree days (KDD; see supplemental text S2;
Butler and Huybers 2015),(3)total growing season
precipitation (Pr
gs
),(4)average temperature during a
critical period of the growing season (Tavg
Cr
),(5)
maximum Tmax during a critical period of the grow-
ing season (Txx
Cr
), and (6)total Pr during a critical
period of the growing season (Pr
Cr
). These predictor
variables are chosen from a full suite of climate metrics
by highest permuted predictor importance.
For the random forest model, the growing season
is defined as the period from the planting date (descri-
bed above)to the last day of the week corresponding
with at least 50% of a state’s corn reaching maturity
based on average 1979–2011 NASS survey data
(USDA/NASS 2018). We identify a county-specific
critical period within the growing season following a
similar approach as Teasdale and Cavigelli (2017).We
correlate 1930–2011 county yield timeseries with
weekly aggregates of Tavg and Pr for all combinations
of consecutive weeks from the first week of a county’s
growing season to the last week of the growing season,
for each year in the dataset. A county-specific critical
period is defined as the period of weeks that returns
the highest absolute correlation between yield and
either Tavg or Pr. We then define the critical period for
the model as the average of those from the 794 coun-
ties; here weeks 12–17 post planting for Tavg
Cr
and
Txx
Cr
and 9–12 weeks post planting for Pr. A graphical
representation of the correlation matrix identifying a
county’s critical period is shown in figure S3.
To evaluate both models’accuracy at simulating
historical maize yields, we use the root mean square
error and the mean absolute percentage error as well as
the coefficient of determination (r
2
), the p-value, the
slope, and the intercept of a linear regression model
between simulated and observed yield. For evaluating
the random forest, we use out-of-bag predictions that
produce unbiased metrics without needing to divide
the observations into training and test sets (Brei-
man 1996). Correlation plots of the observed versus
simulated yields are shown in figure S4.
Results
Both SALUS and the random forest simulate maize
yields that closely reproduce the reported yields (figure
S4). Table 1lists the model performance evaluation
metrics. The high evaluation metrics for the random
forest are due to our decision to train the model on
strongly-trending annual yield data while including
year as a predictor variable. Training the random
forest model with year as the only predictor variable
illustrates the significance of this decision (figure
S4(c)). However, importantly, aspects of interannual
climate variability (e.g. large drought events)are still
captured by this simplified model, and we argue that
4
Environ. Res. Lett. 14 (2019)114008
this approach is analogous to the common approach
of training a model on detrended yield data (e.g. Lobell
and Field 2007)given that the relationship between
year and yield within the full random forest model is
nearly linear (figure S5). In addition, using yields that
are not detrended allows for the direct comparison of
results between the random forest and SALUS crop
model.
SALUS and the random forest model results agree
that maize yields benefited from the warming hole.
Figure 2illustrates the effect of removing the warming
hole on Tmax and the corresponding impact on pre-
dicted maize yield for Green County, IL. Average
growing season Tmax is approximately 1.5 °C higher
under the no warming-hole scenario in Greene
County and the corresponding predicted yield decrea-
ses by an average of 26%.
Without the warming hole, over fifty percent of
counties in the Corn Belt would have experienced
average annual maize yields at least 5.2%–9.8% below
historical values, depending on the climate scenario
and model (figure 3). We report the yield difference as
the average annual percent difference between the pre-
dicted yield under the historical climate and a given
counterfactual climate scenario. Thus, a positive yield
difference represents a yield benefit from the observed
warming hole. The largest yield benefits occur under
the QMap scenario, with median differences of 9.8%
for the random forest and 7.8% for SALUS. Counties
in the Dakotas and western Minnesota exhibit the lar-
gest yield differences of up to 20% for SALUS simula-
tions and 24% according to the random forest,
suggesting that this area may have benefited the most
from the warming hole. The models forced with the
Delta scenario indicate that fifty percent of counties
would have experienced at least a 5.2% (random for-
est)and 7.2% (SALUS)reduction in yield without the
warming hole (figure S6). The QMap and Delta sce-
narios have a similar shift in mean climate, so the dif-
ference in median yield response between the two
scenarios is effectively the yield response due to a
change in climate variability alone. Surprisingly, the
random forest is more sensitive than SALUS to the
increased variability within a growing season as
imposed under the QMap scenario even though it uses
primarily on growing season aggregate climate data.
These results are qualitatively robust to the breakpoint
used in creating the counterfactual scenarios and the
exclusion of the Dust Bowl in the reference period.
Perhaps unsurprisingly, removing the Dust Bowl
from the reference period reduces the median yield
benefit to 4.6% for the random forest forced by the
QMap scenario, however the vast majority (>96%)of
counties still exhibit increased yields relative to the
Table 1. Model evaluation metrics for SALUS and the random forest models. Metrics include root mean square
error (RMSE)and mean annual percent error (MAPE).
Observed versus predicted linear regression
RMSE (tha
−1
)MAPE (%)Intercept (tha
−1
)Slope r
2
p-value
SALUS 1.44 19.3 1.22 0.81 0.67 <0.01
Random forest 0.71 15.9 −0.18 1.03 0.94 <0.01
Figure 2. Effect of removing the warming hole on Tmax and corresponding impact on yield.(a)Time series of maximum temperature
for Greene County IL, under historical climate and no-warming hole counterfactual scenario developed with quantile mapping.
(b)Time series of reported yield for Greene County, IL and predicted yield using the random forest algorithm under historical climate
and counterfactual no warming hole climate developed with quantile mapping. Predicted yield under historical climate (black line)is
estimated using out-of-bag (OOB)prediction.
5
Environ. Res. Lett. 14 (2019)114008
counterfactual scenario with the Dust Bowl removed
(not shown).
In addition to forcing the models with the full
counterfactual climate scenarios, we also explore the
isolated effects of temperature and precipitation on
maize yields (figure 3(c)). Crop models cannot attri-
bute the simulated yield benefit to one particular fac-
tor, given the complex nature of the soil-plant-
atmosphere interactions and feedbacks, so we quantify
the relative importance of temperature and precipita-
tion using the two additional counterfactual scenarios
derived from the QMap scenario. QMap
T
and QMap
Pr
represent historical observations of shortwave radia-
tion with either Tmin and Tmax or precipitation
adjusted by quantile mapping, respectively. The ran-
dom forest suggests that the yield difference under the
QMap scenario is driven almost entirely by the temp-
erature difference between the historical climate and
the QMap climate. For the random forest, there was
no statistical difference in maize yield when precipita-
tion alone was adjusted under the QMap scenario
(p=0.64). These results suggest that the maize yields
in the Corn Belt have benefited little from the increas-
ing summer precipitation in the region, with the possi-
ble exception of the more arid and non-irrigated
northwest corner of the domain (figures S7 and S8).
However, statistical models struggle to capture to
importance of precipitation on plant growth and con-
sequently yield (Lobell and Burke 2010, Ortiz-Bobea
et al 2019).
By contrast, SALUS results suggest that potential
yield gains as a result of the warming hole come from
both lower temperatures and increased precipitation.
SALUS results indicate that 62% of the median yield
benefit resulted from the lower temperatures asso-
ciated with the warming hole, while the remaining
benefit was a result from increased precipitation
(38%). We identify two potential causes for the differ-
ence in the importance of precipitation between the
random forest and SALUS. First, the random forest
was trained using yield observations from the whole
corn belt, and therefore has a limited ability to simu-
late geographical variations in the sensitivity of maize
to precipitation. Second SALUS is able to simulate the
impacts of daily precipitation on crop growth, while
the random forest aggregates precipitation.
To explore potential mechanisms responsible for the
yield reductions under the no-warming hole climate sce-
narios, we examine differences in simulated drought
stress and maturity time using the SALUS model. Under
Figure 3. Difference in historical maize yield between three climate scenarios and two modeling approaches. (a)Difference in
predicted maize yield from the SALUS crop model between the historical climate and alternative climate scenario with no warming
developed using a quantile mapping method (QMap).(b)Same but using the random forest statistical crop model. (c)Violin plots of
the yield difference for al counties under each climate scenario. Each climate scenario removes the warming hole using a different
statistical method. QMap: temperature and precipitation adjusted by quantile mapping approach; Delta: temperature and
precipitation adjusted by a Delta-change approach; QMap
T
: Only temperature adjusted by quantile mapping approach; QMap
Pr
:
Only precipitation adjusted by quantile mapping approach.
6
Environ. Res. Lett. 14 (2019)114008
both no warming hole scenarios, drought stress increases,
and crops mature earlier, leading to lower yields. Maize in
the average county reaches maturity 2.7 d or 2.9 d earlier
than the historical observed climate and experiences an
52% or 49% increase in drought stress for the QMap or
the Delta scenarios, respectively (figure 4).Thelargestdif-
ference in crop maturity time occurs in the northern
counties of Wisconsin where cooler late spring and early
autumn temperatures otherwise slow plant development.
Interestingly, the northern counties of Minnesota and the
Dakotas do not exhibit the same reduction in time to
maturity, likely due to later planting dates in this region.
There is a longitudinal gradient in the reduction of
drought stress between the historical and QMap scenar-
ios. Counties in the eastern half of the domain appear to
have substantially lower drought stress levels as a result of
the warming hole and concurrent precipitation trends.
Random forest algorithms are often referred to as
‘black boxes’, as it can be difficult to interpret the rela-
tionships formed within the ensemble of decision
trees. Accumulated local effect (ALE)plots provide a
way to visualize the relationship between the response
variable (yield)and each predictor variable in isolation
(Apley 2016). Figure 5shows ALE plots for climate
variables used in the random forest as well as the total
distribution of that variable within the domain. The
plots are centered at zero, which represents the mean
model prediction across all values of the variable.
Values below zero have a net negative effect on yield
(decreased simulated yield)and values above zero are
beneficial (increased simulated yield). ALE plots for
the non-climatic variables are shown in figure S5.
GDD, the hottest daily maximum temperature during
the critical period (Txx
cr
),12–17 weeks after planting,
and KDD (Butler and Huybers 2015)have a similar
effect on yield (see methods for description of random
forest and variable calculation). The ALE plot of aver-
age temperature during the critical period (Tavg
cr
)
suggests that days with average temperature above
22 °C can be detrimental to maize yields. However,
this threshold should be interpreted with caution as
Tavg is a statistical construct of daily Tmax and Tmin,
which are more likely to physically affect plant growth.
The ALE plots of Pr give insight into why forcing the
random forest with QMap
pr
had little effect on simu-
lated yield (figure 3(c)). Precipitation summed over
the growing season (Pr
gs
)and during the critical per-
iod (Pr
cr
; weeks 9–12 post planting)exhibit relatively
flat curves, and the corresponding precipitation dis-
tributions illustrate the general abundance of water in
the region. The more arid northwestern portion of the
domain did benefit from the historical increase in
summer precipitation, so a similar analysis focused
only on this region might find a stronger relationship
between Pr and maize yields (figure S7).
Conclusion
Understanding the dynamic relationship between
agricultural productivity and regional climate is of
critical importance given the challenges imposed by
population growth, shifting diets, and climate change.
Here we find that the 20th century trend in US maize
yield was significantly aided by anomalously cool
temperatures and increased growing season precipita-
tion. These results expand upon those of Butler et al
(2018), who use a statistical crop model to show that
increased minimum temperatures and a reduction in
maximum temperature extremes since 1980 have been
beneficial to maize yields. We find that this agricultu-
rally pleasant climate extends back to 1958 and also
quantify the importance of a simultaneous increase in
growing season precipitation. We further identify both
prolonged maturity time and reduced water stress as
the physiological mechanisms responsible for the yield
benefit(figure 4). It is unlikely that the pleasant climate
Figure 4. Impacts of warming hole on time to maturity and drought stress. (a)Difference in time to maturity of SALUS simulated
maize yield between historical and QMap climate scenario. (b)Difference in drought stress between historical climate and QMap
scenario. Positive values of time to maturity indicate later maturity, and negative drought stress values indicate reduced drought stress,
in the presence of the warming hole.
7
Environ. Res. Lett. 14 (2019)114008
observed during the latter half of the 20th century will
substantially persist into the future. Although max-
imum growing season temperatures in the Corn Belt
continue to increase at a slower rate than minimum
temperatures, possibly due to agricultural expansion
(Mueller et al 2015, Alter et al 2018, Nikiel and
Eltahir 2019), annual temperatures are projected to
increase by 2.3 °C by mid-century (RCP 4.5; Vose et al
2017). Summer precipitation in the region is projected
to decrease by approximately 10% while extreme
events will likely increase (Easterling et al 2017).
The dramatic improvements in technology and
management practices across US Corn Belt produced
remarkable increases in maize yield and productivity.
However, we show that a contemporaneous climate
anomaly referred to as the US ‘warming hole’, which
translated to more advantageous weather in the
region, also contributed to yield increases. We find
that the warming hole resulted in an increased median
annual maize yield of 5%–10%, primarily in response
to decreased growing season temperatures and
decreased heat extremes. SALUS driven with counter-
factual scenarios isolating the increase in precipitation
and decrease in temperature indicate that cooler
temperatures were responsible for 62% (and thus
increased precipitation was repsonsible for 38%)of
the simulated yield increase across most of the Corn
Belt. As the central US prepares to adapt to substantial
projected increases in temperatures by the end of the
21st century, it is essential to understand the extent
that the US warming hole has historically benefitted
maize yields. Our results underscore the relative lack
of climate change impacts on central US maize pro-
duction to date, and the potential compounded chal-
lenge that a collapse of the warming hole and climate
change would create for farmers across the Corn Belt.
Acknowledgments
This study was funded by the United States Depart-
ment of Agriculture National Institute of Food and
Agriculture (grants 2015-68007-23133 and 2018-
67003-27406), National Science Foundation (BCS
184018), and Nelson A. Rockefeller Center at Dart-
mouth College. This paper makes use of agronomic
data provided by the United States Department of
Agriculture National Agricultural Statistics Service
and the Livneh hydrometeorological climate dataset
provided by NOAA/OAR/ESRL PSD, Boulder, Col-
orado, USA, from their Web site at https://esrl.noaa.
gov/psd/.
Figure 5. Accumulated local effects (ALE)plots of climate variables from the random forest model. Cumulative density functions
show the distribution of each variable. Positive values of ALE indicate a yield benefit relative to the mean. Note the distinct vertical axes
scales between the bottom and top rows. ALE plots and histograms are truncated at the 1% and 99% quantiles to remove outliers.
8
Environ. Res. Lett. 14 (2019)114008
Author contributions
TFP, JMW, ADK and DWH designed the study and
developed the random forest model. TFP and JMW
developed the climate scenarios. LL and BB ran the
SALUS model simulations and analyzed drivers of
yield changes. All authors contributed to writing the
manuscript.
Competing interests statement
We have no conflicts of interest to declare and none of
the material has been published or is under considera-
tion elsewhere.
ORCID iDs
Trevor F Partridge https://orcid.org/0000-0003-
1589-4783
Jonathan M Winter https://orcid.org/0000-0003-
1261-4774
Lin Liu https://orcid.org/0000-0001-9078-9516
Anthony D Kendall https://orcid.org/0000-0003-
3914-9964
Bruno Basso https://orcid.org/0000-0003-
2090-4616
David W Hyndman https://orcid.org/0000-0001-
9078-9516
References
Alter R E, Douglas H C, Winter J M and Eltahir E A B 2018
Twentieth century regional climate change during the
summer in the central united states attributed to agricultural
intensification Geophys. Res. Lett. 45 1586–94
Apley D W 2016 Visualizing the effects of predictor variables in
black box supervised learning models arXiv:1612.08468
Baker D G, Ruschy D L and Skaggs R H 1993 Agriculture and the
recent ‘Benign Climate’in Minnesota Bull. Am. Meteorol. Soc.
74 1035–40
Basso B, Dumont B, Cammarano D, Pezzuolo A, Marinello F and
Sartori L 2016 Environmental and economic benefits of
variable rate nitrogen fertilization in a nitrate vulnerable zone
Sci. Total Environ. 545–546 227–35
Basso B, Ritchie J T, Grace P R and Sartori L 2006 Simulation of
tillage systems impact on soil biophysical properties using the
SALUS model Ital. J. Agronomy 1677
Boryan C, Yang Z and Di L 2012 Deriving 2011 cultivated land cover
data sets using usda national agricultural statistics service
historic cropland data layers 2012 IEEE Int. Geoscience and
Remote Sensing Symp. (Piscataway, NJ: IEEE)pp 6297–300
Breiman L 1996 Out-Of-Bag Estimation Technical Report (Berkeley,
CA: Department of Statistics, University of California,
Berkeley)(https://www.stat.berkeley.edu/~breiman/
OOBestimation.pdf)
Breiman L 2001 Random forests Mach. Learn. 45 5–32
Brocca L, Hasenauer S, Lacava T, Melone F, Moramarco T,
Wagner W and Bittelli M 2011 Soil moisture estimation
through ASCAT and AMSR-E sensors: an intercomparison
and validation study across Europe Remote Sens. Environ. 115
3390–408
Brown D G, Johnson K M, Loveland T R and Theobald D M 2005
Rural land-use trends in the conterminours United Sates,
1950–2000 Ecol. Appl. 15 1851–63
Butler E E and Huybers P 2015 Variations in the sensitivity of US
maize yield to extreme temperatures by region and growth
phase Environ. Res. Lett. 10 034009
Butler E E, Mueller N D and Huybers P 2018 Peculiarly pleasant
weather for US maize PNAS 115 11935–40
Deines J, Kendall A and Hyndman D W 2017 Annual irrigation
dynamics in the US Northern High Plains derived from
Landsat satellite data Geo. Res. Lett. 44 9350–60
Dimitri C, Effland A and Conklin N 2005 The 20th Century
Transformation of US Agriculture and Farm Policy (Economic
Information Bulletin No. 3)(Washington, DC: USDA Eco-
nomic Research Service)(www.ers.usda.gov)
Easterling D R, Arnold J R, Knut son T, Kunkel K E, LeGra nde A N,
Leung L R and Wehner M F 2017 Precipit ation Change in the
United States. Climate Science Special Report: Fourth
National Climate Assessment ed D J Wueebbles et al vol 1
(Washington , DC: US Global Change R esearch Program)
pp 207–30
Ellenburg W L, McNider R T, Cruise J F and Christy J R 2016
Towards an understanding of the twentieth-century cooling
trend in the Southeastern United States: biogeophysical
impacts of land-use change Earth Interact. 20 1–31
Hartmann D J, Klein Tank A M G, Rusticucci M, Alexander L V,
Brönnimann S, Charabi Y A R and Zhai P 2013 Observations:
Atmosphere and Surface Climate Change 2013: The Physical
Science Basis. Contribution of Working Group I to the Fifth
Assessment Report of the Intergovernmental Panel on Climate
Change pp 159–254
Jeong J H, Resop J P, Mueller N D, Fleisher D H, Yun K,
Butler E E and Kim S-H 2016 Random forests for global and
regional crop yield predictions PLoS One 11 e0156571
Kucharik C J and Ramankutty N 2005 Trends and variability in
US corn yields over the twentieth century Earth Interact. 9
1–29
Kumar S, Merwade V, Kinter J L and Niyogi D 2013 Evaluation of
temperature and precipitation trends and long-term
persistence in CMIP5 twentieth-century climate simulations
J. Clim. 26 4168–85
Leibensperger E M, Mickley L J, Jacob D J, Chen W T, Seinfeld J H,
Nenes A and Rind D 2012 Climatic effects of 1950–2050
changes in US anthropogenic aerosols: II. Climate response
Atmos. Chem. Phys. 12 3349–62
Livneh B, Bohn T J, Pierce D W, Munoz-Arriola F, Nijssen B,
Vose R and Brekke L 2015 A spatially comprehensive,
hydrometeorological data set for Mexico, the US, and
Southern Canada 1950–2013 Sci. Data 2150042
Livneh B, Rosenberg E A, Lin C, Nijssen B, Mishra V,
Andreadis K M and Lettenmaier D P 2013 A long-term
hydrologically based dataset of land surface fluxes and states
for the conterminous United States: update and extensions
J. Clim. 26 9384–92
Lobell D B and Asseng S 2017 Comparing estimates of climate
change impacts from process-based and statistical crop
models Environ. Res. Lett. 12 015001
Lobell D B and Bonfils C 2008 The effect of irrigation on regional
temperatures: a spatial and temporal analysis of trends in
California, 1934–2002 J. Clim. 21 2063–71
Lobell D B and Burke M B 2010 On the use of statistical models to
predict crop yield responses to climate change Agric. For.
Meteorol. 150 1443–52
Lobell D B and Field C B 2007 Global scale climate–crop yield
relationships and the impacts of recent warming Environ. Res.
Lett. 2014002
Maraun D and Widmann M 2018 Statistical Downscaling and Bias
Correction for Climate Research (Cambridge: Cambridge
University Press)
Mascioli N R, Previdi M, Fiore A M and Ting M 2017 Timing and
seasonality of the United States ‘warming hole’Environ. Res.
Lett. 12 034008
Maurer E P, Wood A W, Adam J C, Lettenmaier D P and Nijssen B
2002 A long-term hydrologically based dataset of land surface
fluxes and states for the conterminous United States J. Clim.
15 3237–51
9
Environ. Res. Lett. 14 (2019)114008
Meehl G A, Arblaster J M and Branstator G 2012 Mechanisms
contributing to the warming hole and the consequent US
east-west differential of heat extremes J. Clim. 25
6394–408
Meehl G A, Arblaster J M and Chung C T Y 2015 Disappearance of
the southeast US ‘warming hole’with the late 1990s transition
of the Interdecadal Pacific Oscillation Geophys. Res. Lett. 42
5074–81
Misra V, Michael J P, Boyles R, Chassignet E P, Griffin M and
O’brien J J 2012 Reconciling the spatial distribution of the
surface temperature trends in the Southeastern United States
J. Clim. 25 3610–8
Mueller N D, Butler E E, McKinnon K A, Rhines A, Tingley M,
Holbrook N M and Huybers P 2015 Cooling of US Midwest
summer temperature extremes from cropland intensification
Nat. Clim. Change 6317–22
Nikiel C A and Eltahir E A B 2019 Summer climate change in the
midwest and great plains due to agricultural development
during the twentieth century J. Clim. 32 5583–99
Ortiz-Bobea A, Wang H, Carrillo C M and Ault T R 2019 Unpacking
the climatic drivers of US agricultural yields Environ. Res. Lett.
14 064003
Partridge T F, Winter J M, Osterberg E C, Hyndman D W,
Kendall A D and Magilligan F J 2018 Spatially distinct
seasonal patterns and forcings of the US Warming Hole
Geophys. Res. Lett. 45 2055–63
Pei L, Moore N, Zhong S, Kendall A D, Gao Z and Hyndman D W
2016 Effects of Irrigation on summer precipitation over the
United States J. Clim. 29 3541–58
Ritchie J T 1972 Model for predicting evaporation from a row crop
with incomplete cover Water Resour. Res. 81204–13
Rogers J C 2013 The 20th century cooling trend over the
southeastern United States Clim. Dyn. 40 341–52
Siebert S, Kummu M, Porkka M, Döll P, Ramankutty N and
Scanlon B R 2014 A global dataset of the extent of irrigated
land from 1900 to 2005 Hydrol. Earth Syst. Sci. Discuss. 11
13207–58
Teasdale J R and Cavigelli M A 2017 Meteorological fluctuations
define long-term crop yield patterns in conventional and
organic production systems Sci. Rep. 7688
Thompson L M 1986 Climatic change, weather variability, and corn
production1 Agron. J. 78 649
USDA NASS 2018 USDA—National Agricultural Statistics Service
—Quick Stats [Data file](https://nass.usda.gov/Quick_
Stats/)(Retrieved: 26 March 2019)
USDA NRCS 2017 Soil Survey Geographic Database (SSURGO).
NRCS [Data file].(Retrieved: 26 March 2018)
USGCRP 2017 Climate Science Special Report: Fourth National
Climate Assessment ed D J Wuebbles et al vol 1 (Washington,
DC: US Global Change Reseacrh Program)(https://doi.org/
10.7930/J0J964J6)
Vose R S, Easterling D R, Kunkel K E, LeGrande A N and
Wehner M F 2017 Temperature changes in the United States
ed D J Wuebbles et al Climate Science Special Report: Fourth
National Climate Assessment vol 1 (Washington, DC: US
Global Change Research Program)pp 185–206
Winter J M, Bowen F L, Partridge T F and Chipman J W 2019 Future
extreme event risk in the rural Northeastern United States
Ann. Am. Assoc. Geogr. 109 1110–30
World Bank 2019 Agricultural Land (sq. km).[Data file](https://
data.worldbank.org/indicator/AG.LND.AGRI.K2?
locations=US)(
Retrieved: 22 March 2019)
10
Environ. Res. Lett. 14 (2019)114008
Content uploaded by Trevor F Partridge
Author content
All content in this area was uploaded by Trevor F Partridge on Sep 16, 2019
Content may be subject to copyright.
Available via license: CC BY 3.0
Content may be subject to copyright.