This content has been downloaded from IOPscience. Please scroll down to see the full text.

Download details:

IP Address: 71.192.160.26

This content was downloaded on 26/04/2016 at 18:58

Please note that terms and conditions apply.

Uncertainty in future agro-climate projections in the United States and benefits of greenhouse

gas mitigation

View the table of contents for this issue, or go to the journal homepage for more

2016 Environ. Res. Lett. 11 055001

(http://iopscience.iop.org/1748-9326/11/5/055001)

Home Search Collections Journals About Contact us My IOPscience

Environ. Res. Lett. 11 (2016) 055001 doi:10.1088/1748-9326/11/5/055001

LETTER

Uncertainty in future agro-climate projections in the United Statesand benefits of greenhouse gasmitigation

ErwanMonier1, Liyi Xu1 andRichard Snyder2

1 Joint Programof the Science and Policy ofGlobal Change,Massachusetts Institute of Technology, Cambridge,Massachusetts, USA2 Department of Land, Air andWater Resources, University of California, Davis, California, USA

E-mail: emonier@mit.edu

Keywords: climate change, agriculture, uncertainty, natural variability, cropmodeling, benefits ofmitigation, time of emergence

Supplementarymaterial for this article is available online

AbstractScientific challenges exist on how to extract information from thewide range of projected impactssimulated by cropmodels driven by climate ensembles. A stronger focus is required to understand andidentify themechanisms and drivers of projected changes in crop yield. In this study, we investigatethe robustness of future projections offivemetrics relevant to agriculture stakeholders (accumulatedfrost days, dry days, growing season length, plant heat stress and start offield operations).We use alarge ensemble of climate simulations by theMIT IGSM-CAM integrated assessmentmodel thataccounts for the uncertainty associatedwith different emissions scenarios, climate sensitivities, andrepresentations of natural variability. By the end of the century, theUS is projected to experience fewerfrosts, a longer growing season,more heat stress and an earlier start offield operations—although themagnitude and even the sign of these changes vary greatly by regions. Projected changes in dry days areshownnot to be robust.We highlight the important role of natural variability, in particular for changesin dry days (a precipitation-related index) and heat stress (a threshold index). Thewide range of ourprojections compares well the CMIP5multi-model ensemble, especially for temperature-relatedindices. This suggests that using a single climatemodel that accounts for key sources of uncertaintycan provide an efficient and complementary framework to themore common approach ofmulti-model ensembles.We also show that greenhouse gasmitigation has the potential to significantlyreduce adverse effects (heat stress, risks of pest and disease) of climate change on agriculture, whilealso curtailing potentially beneficial impacts (earlier planting, possibility formultiple cropping). Amajor benefit of climatemitigation is potentially preventing changes in several indices to emerge fromthe noise of natural variability, even by 2100. This hasmajor implications considering that anysignificant climate change impacts on crop yieldwould result in nation-wide changes in the agriculturesector. Finally, we argue that the analysis of agro-climate indices shouldmore often complement cropmodel projections, as they can provide valuable information to better understand the drivers ofchanges in crop yield and production and thus better inform adaptation decisions.

1. Introduction

Climate change is expected to have a substantialimpact on agricultural production because agricultureis highly dependent on climate variables such asprecipitation, temperature, and radiation. The vulner-ability of agriculture and food security to climatechange has been extensively investigated, generallyrelying on biophysical models [1–4] or empiricalmodels [5–8]. Such studies, even though they use very

different sets of crop models, indicate strong negativeimpacts of climate change without adaptive measures,but with a large uncertainty in the range of impacts.The wide range of projected impacts on agriculture isdriven by both the uncertainty in climate projectionsand the structural (or parameterization) differencesbetween crop models. While using climate ensembleswith crop models is increasingly common to projectthe potential impacts of climate change on agriculture,scientific challenges exist on how the uncertainty is

OPEN ACCESS

RECEIVED

16October 2015

REVISED

12March 2016

ACCEPTED FOR PUBLICATION

6April 2016

PUBLISHED

26April 2016

Original content from thisworkmay be used underthe terms of the CreativeCommonsAttribution 3.0licence.

Any further distribution ofthis workmustmaintainattribution to theauthor(s) and the title ofthework, journal citationandDOI.

© 2016 IOPPublishing Ltd

analyzed and the information is extracted frommodelsthat are sometimes used as black boxes [9]. For thisreason, a stronger focus is required to understand andidentify the mechanisms and drivers of projectedchanges in crop yield, especially in light of naturalvariability and potential future changes in extremeevents.

Various metrics and indices have also been devel-oped to address the vulnerability of agricultural pro-duction under climate change. Many of these indicesfocus on moisture availability or drought as the mostsignificant climate impact on crop yields. The Palmerdrought severity index, standardized precipitationindex, NOAA drought index, and Palmer Z-index arewidely used [10–12]. New indices, such as the standar-dized precipitation evapotranspiration index [13], soilmoisture deficit index and evapotranspiration deficitindex [14], continue to be developed and provide valu-able information to farmers [15]. While each of theseclimate indices provides important insight on agri-cultural responses under climate change, [10] showthat there are significant variations in model perfor-mance depending on the choice of drought indices.More recently, studies have focused on agro-climateor agro-meteorological indices that are designed toassess the potential changes in crop exposure to temp-erature (heat and cold) and water stresses [16–18]. Inaddition, metrics relevant to management practices,such as start of field operations and growing seasonlength, provide additional value to farmers and deci-sion makers [19]. Therefore, a comprehensive evalua-tion of both agro-climate and management metricscan help address agriculture vulnerability under cli-mate change.

In this study, we investigate future agro-climateprojections using indices shown to be relevant to landmanagement stakeholders, identify the associatedimpacts on agriculture, and estimate the benefits ofgreenhouse gas mitigation. Soil water balance metricsare highly relevant to stakeholders and are commonlyincluded in climate change studies, but other metricsare also important and often ignored. We examine therobustness of future projections of these agro-climateindices using a large ensemble of integrated economicand climate projections prepared for the US Environ-mental Protection Agency’s Climate Change Impacts

and Risks Analysis (CIRA) project [20]. Using thislarge ensemble, we investigate the role of multiplesources of uncertainty in global and regional climateprojections—including the emissions scenario, theglobal climate system response (climate sensitivity)and natural variability—on the impact of climatechange and greenhouse gas mitigation on agriculture.We further compare our analysis to a multi-modelensemble based on 31 different climate models fromthe coupled model intercomparison project phase 5(CMIP5, see [32]). Section 2 describes the agro-cli-mate indices, the climate model and observationaldata and themethodology used in this study. Section 3presents the results of the analysis, focusing on the roleof uncertainty and the benefits of mitigation. Section 4provides a summary and discussion, with concludingremarks in section 5.

2.Methodology

2.1. Agro-climate indicesWe follow [18] and use five agro-climate indicesdeveloped by [19] that were deemed ‘very’ or ‘quite’useful by land management stakeholder focus groups.Not only do these metrics have little conceptualoverlap, they also account for various stresses of landproductivity and management practices. Table 1 pro-vides a description and methodology for the calcul-ation of these five indices. Accumulated frost days(AFD) serves as a frost index that relates to both coldstress and incidence of pest and disease. Dry days(DD), the number of days with precipitation below athreshold, is used here as a drought index, instead ofusing more sophisticated indices related to soil moist-ure that are not necessarily well simulated by climatemodels [21, 22]. Growing season length (GSL), calcu-lated as the number of days between the last frost(minimumdaily temperature below0 °C) in the springand the first hard frost in the fall [23, 24], relates to thetiming and length of the growing season, which canhave far reaching consequences for plant and animalecosystems [25]. Plant heat stress (PHS), the numberof days with daily maximum temperature above athreshold, identifies the risk of heat stress and theassociated yield decline. Various thresholds can beused for specific plants or crops; in this study we use a

Table 1. Selected agro-climate indices, including type, unit, and description of calculation based on [18] and [19], customized for theUnitedStates.

Index Type Units Description

Accumulated frost days (AFD) Count Days DayswhereTmin< 0 °CDry days (DD) Count Days Dayswhen P<1 mm

Growing season length (GSL) Count Days Days between the last frost in the spring and thefirst frost in the fall

Plant heat stress (PHS) Count Days DayswhenTmax> 29 °CStart of field operations (SFO) Date Day of year Daywhen the sumofTavg from1st January is greater than 200 °C

Tmin, Tavg and Tmax refer to, respectively, the daily minimum, daily mean and daily maximum temperature at2 mheight (in °C);P refers to dailymean precipitation (inmm).

2

Environ. Res. Lett. 11 (2016) 055001

threshold of 29 °C, which corresponds to maize, amajor US crop [6], although different thresholds (i.e.30 °C for soybeans or 32 °C for cotton) do not changethe conclusions of the our study. Finally, the start offield operations (SFO), calculated as the date ofthermal accumulation of 200 °C, is an index derivedfrom workshops with agricultural stakeholders thatrefers to the earliest date in a year when a fieldmight beusefully cultivated [19].

2.2. Climatemodel dataWe use a 45-member ensemble of simulations usingtheMIT IntegratedGlobal SystemModel-CommunityAtmosphere Model (IGSM-CAM) modeling frame-work [26] developed for the US EnvironmentalProtection Agency’s climate CIRA project [20]. Thisensemble is made up of three consistent socioeco-nomic and emissions scenarios: a reference scenario(REF) with unconstrained emissions and two green-house gas mitigation scenarios that rely on a uniformglobal carbon tax to stabilize radiative forcing at4.5Wm−2 (POL4.5) and 3.7Wm−2 (POL3.7) by2100. More details on the emissions scenarios andeconomic implications, along with how they relate tothe representative concentration pathway (RCP) sce-narios [27], are given in [28]. For each emissionsscenario, the IGSM-CAMwas run with three values ofclimate sensitivity (CS=2.0°, 3.0° and 4.5 °C),corresponding to the likely range and best guess,obtained via radiative cloud adjustment (see [29]). Foreach climate sensitivity, a five-member ensemble wasrun to produce different representations of naturalvariability, thus resulting in a 15-member ensemblefor each emissions scenario. A detailed procedure forthe design of the large IGSM-CAM ensemble used inthis study, along with an analysis of the climateprojections for the US, can be found in [30]. Becausewe use integrated economic and climate projectionsobtained using an Integrated Assessment Model(IAM), we can directly attribute the differences infuture agro-climate projections between the differentemissions scenarios to explicit policy choices, thusidentifying the benefits of greenhouse gas mitigation.Furthermore, while this ensemble is derived using asingle climatemodel, it accounts for the uncertainty inemissions scenarios, global climate response andnatural variability, which accounts for a substantialshare of the full uncertainty in future climate projec-tions [30]. We analyze 20-year time periods over fivedifferent representations of natural variability, result-ing in 100 years to define changes in 2050 (defined asthe period 2041–2060) and 2100 (defined as the period2081–2100) relative to present day (defined as theperiod 1991–2010), in order to obtain robust estimatesof the anthropogenic signal and thus identify thebenefits of greenhouse gasmitigation.We also identifythe various uncertainties represented in the modelingframework.

To provide some context to the agro-climate pro-jections using the IGSM-CAM ensemble, we compareour results to the CMIP5 multi-model ensembleunder the RCP8.5 and RCP4.5 scenarios. We computethe agro-climate indices using historical and futuresimulations from 31 climate models (see the supple-mentalmaterials for a complete list) after interpolatingthe required input data to the same 2°×2.5° grid asthe IGSM-CAM. The 31 models considered werethose for which the required daily inputs for bothemissions scenarios were available at the time of thestudy. Only one run is used for eachmodel, even whenan ensemble (i.e. with different initial conditions)was run.

2.3.Observational dataWe evaluate the capability of the IGSM-CAM tosimulate the five agro-climate indices for present-dayconditions using two independent observational data-sets: the National Aeronautics and Space Administra-tion Modern Era Retrospective-Analysis for Researchand Applications (MERRA) reanalysis [33], at0.5°×0.66° resolution, and the National Centers forEnvironmental Prediction/National Center for Atmo-spheric Research (NCEP/NCAR) Reanalysis [34], atT62 gaussian grid (approximately 2°×2° resolution).We rely on two different reanalysis datasets toillustrate—not quantify, as this would require a morecomplete analysis—the uncertainty in observationaldatasets that include the role of lower resolution thatclimatemodels suffer from.

2.4. Time of emergenceWe investigate the robustness of the agro-climateprojections by estimating when the signal (S) ofanthropogenic changes is emerging against the noise(N) of natural variability. There is no single metric ofemergence and various studies have used differentdefinitions of S and N to estimate the signal-to-noiseratio (S/N). For example, [35] define the warmingsignal by regressing temperature at each grid cell of aclimate model simulation against a smoothed versionof its global mean temperature, and they define thenoise as the interannual standard deviation from apre-industrial control simulation. In [36], the authorsdefine the signal as a 10 year mean anomaly and thenoise as the interannual standard deviation from atransient historical simulation.Meanwhile, [37] definethe signal of precipitation change as the 20 year meananomaly of a multi-model ensemble mean and thenoise as intermodel spread and internal multi-decadalvariability. Finally, [38] analyzes the time of emergenceof seasonal temperature changes using a large ensem-ble of climate simulations with different initial condi-tions, where the signal is the ensemble mean and thenoise the standard deviation of internal variabilitycomputed for each year across the ensemble (after a10-year running mean). In this study, we estimate the

3

Environ. Res. Lett. 11 (2016) 055001

signal (S) for each agro-climate index as the 20 yearrunning mean anomaly from present day for eachclimate simulation and the noise (N) as the interann-ual standard deviation from a pre-industrial controlsimulation. We then define the time of emergence asthe first year in which the S/N exceeds a thresholdof 2.

3. Results

Figure 1 shows the evaluation of the IGSM-CAMsimulation of the five agro-climate indices for present-day conditions compared to the two reanalysis data-sets. Overall, the MIT IGSM-CAM exhibits a verygood capability to reproduce the magnitude andspatial features of the agro-climate indices. The tworeanalysis products show a good agreement with eachother in both themagnitude and spatial distribution ofthe indices. Nonetheless, noticeable inconsistenciesexist that can be attributed to differences in observa-tional datasets assimilated in the climate models,differences between the climate models used (includ-ing resolution), and the methodology for data assim-ilation. Generally, the IGSM-CAMbias falls within therange of the two observation datasets and is largelydriven by the lower resolution of the climate model.The AFD and SFO indices show a strong north–southgradient, with largest values over the Rocky Moun-tains, the Great Lakes and New England. The GSL andPHS show the opposite pattern, with the largest valuesin the South. The strongest DD values are oversouthwestern states and the Great Plains. As expected

from any climate model, systematic biases are presentand consistent with previous evaluations of the model[26, 31]: the IGSM-CAM exhibits a warm bias in theMidwest and a dry bias in the Southeast. This warmbias can be identified in the AFD and PHS indices, andto a lesser extent in the GSL and SFO indices while thedry bias is associated with a positive bias in the DDindex. Altogether, the IGSM-CAM shows reasonableskills at reproducing the major characteristics of theagro-climate indices over theUS.

Figure 2 shows maps of forced changes in eachagro-climate index in 2100 relative to present day, foreach climate sensitivity considered (CS=2.0°, 3.0°and 4.5 °C) and for the REF and POL4.5 scenarios. Foreach emissions scenario and climate sensitivity, theforced changes are estimated as themean over the five-member ensemble with different representations ofnatural variability, thus producing a 100-year mean(20-year period window, five simulations). The IGSM-CAM simulates a very wide range of changes in agro-climate projections, from little change under thePOL4.5 scenario for the low climate sensitivity to verylarge changes under the REF scenarios for the high cli-mate sensitivity. The greenhouse gas mitigation sig-nificantly reduces the projected changes, even whenconsidering the uncertainty in the global climate sys-tem response represented by the climate sensitivity:the projected changes under the POL4.5 scenario withthe high climate sensitivity (CS=4.5 °C) is alwayslower than the projected changes under the REF sce-nario with the low climate sensitivity (CS=2.0 °C).In addition, the patterns of change are similar for eachclimate scenario (climate sensitivity and emissions

Figure 1.USmaps of present-day (1991–2010) (a) accumulated frost days (AFD), (b) dry days (DD), (c) growing season length (GSL),(d) plant heat stress (PHS) and (e) start offield operations (SFO) for theMERRA reanalysis, NCEPReanalysis and simulated by theMIT IGSM-CAM. Themean over thefivemembers with different representations of natural variability for CS=3.0 °C is shown fortheMIT IGSM-CAM.

4

Environ. Res. Lett. 11 (2016) 055001

scenario), largely because of the use of a single modeland the large averaging period used (100 years). TheAFD index is projected to decrease the most in theRocky Mountains, the Great Lakes and New England(around 60–100 days under REF, half under POL4.5).A similar pattern can be seen for the SFO, which pro-jects to start earlier (negative values) especially in theseregions, by as much as 75–100 days under REF andhalf under POL4.5. Increases in the GSL are largestover the Northwest and the NorthEast (from 70 to 100days under REF, from 15 to 50 days under POL4.5).PHS is projected to increase the most over the westernUS, the Gulf Coast and the East Coast (from 40 to80 days under REF, 10–30 days under POL4.5).Finally, DD are projected to increase in the western US(as much as 30 days under REF, half under POL4.5)and decrease elsewhere (30–40 days under REF, halfunder POL4.5), especially over the Great Plains. Thisdipole pattern is consistent with the tendency of the

IGSM-CAM, as well as the Community Climate Sys-tem Model version 3, which shares the same atmo-spheric component (CAM).

Figure 3 shows maps of changes in each agro-cli-mate index in 2100 relative to present day, for eachensemble member with different representations ofnatural variability for the simulations with a climatesensitivity of 3.0 °C under the POL4.5 scenario. Thisanalysis identifies the uncertainty in natural variability—in particular multi-decadal variability—since themaps show 20-year mean changes. The impact of nat-ural variability varies greatly by index. For changes inAFD and in SFO, different representations of naturalvariability mainly affect the magnitudes of the pro-jected changes, but not the patterns of change. On theother hand, changes in the other three indices presentdifferent magnitudes and patterns, and even differentsigns. This is particularly striking given the long periodof averaging. While all members show a general

Figure 2.USmaps of projected forced changes in (a) accumulated frost days (AFD), (b)dry days (DD), (c) growing season length(GSL), (d) plant heat stress (PHS) and (e) start offield operations (SFO) in 2100 (2081–2100) relative to present day (1991–2010),simulated by theMIT IGSM-CAM for each climate sensitivity considered (CS=2.0°, 3.0° and 4.5 °C) and for the REF scenario (toppanel) and the POL4.5 scenario (bottompanel). For each emissions scenario and climate sensitivity, the forced changes are estimatedas themean over thefive-member ensemble with different representations of natural variability.

5

Environ. Res. Lett. 11 (2016) 055001

tendency for increases in DD in the West (and decrea-ses elsewhere), different members exhibit differentmagnitudes and extents of change. For example,member 2 projects a weak decrease in DD in the cen-tral US but a clear increase in theWest, and member 3displays a clear decrease over most of the US with onlya little increase on the West Coast. This implies lessrobustness in the projections of changes in DD com-pared to changes in AFD and SFO. This is consistentwith previous findings on the large uncertainty asso-ciated with natural variability for precipitation chan-ges [31, 39, 40]. The GSL and PHS are also clearly

impacted by the role of natural variability. While allmembers show patterns generally consistent with theensemble mean shown in figure 2—for GSL, the lar-gest increases in the Northwest and the smallestincreases in the South; for PHS, the largest increases inthe western US, the Gulf Coast and the East Coast—individual members disagree on the magnitude andspatial extent of the largest changes, and can even dis-agree on the sign in specific regions. For example,member 3 projects widespread decreases in PHS overmajor parts of the Great Plains and some decreases inGSL over a small part of Texas. This implies little

Figure 3.USmaps of projected changes in (a) accumulated frost days (AFD), (b) dry days (DD), (c) growing season length (GSL), (d)plant heat stress (PHS) and (e) start of field operations (SFO) in 2100 (2081–2100) relative to present day (1991–2010), simulated bytheMIT IGSM-CAM for eachmember with different representations of natural variability for CS=3.0 °CandPOL4.5 scenario.

6

Environ. Res. Lett. 11 (2016) 055001

robustness in the projections of changes in GSL andPHS in these regions, which is surprising given thegeneral assumption of robustness in temperature-related simulations. The same analysis for the REF sce-nario in 2050 and in 2100 is shown in the supple-mental materials. It reveals that the relative role ofnatural variability is lessened under a scenario with alarger forcing. At the same time, the absolute rangeassociated with natural variability remains constantamong scenarios and time periods, corresponding tothe irreducible error in the projections, as shownin [30].

Figure 4 shows a summarized analysis of the rangeof projected changes in each agro-climate index, area-averaged over the US, in 2050 and 2100 relative to pre-sent day. The range is estimated over the 15-memberIGSM-CAM ensemble with different climate sensitiv-ities and representations of natural variability for eachemissions scenarios and over the CMIP5 multi-modelensemble (31 models) under the RCP8.5 and RCP4.5.We also show the 1 and 2 standard deviation of thenatural variability derived from a pre-industrial con-trol simulation of the IGSM-CAM to provide a briefsignal-to-noise analysis. TheUS as a whole is projectedto experience fewer frosts, a longer GSL, an earlier

SFO, an increase in heat stress. The range of changes isparticularly wide under the REF scenario, especially by2100, with the upper bounds close to twice as large asthe lower bound. For example, increases in GSL underREF by 2100 range from 38 to 82 days, and decreasesin AFD range from 32 to 60 days. The implementationof either greenhouse gas mitigation scenario con-sidered in this study cuts by half most of the changesprojected under the unconstrained scenario. In addi-tion, the range of changes for the unconstrained andmitigation scenarios do not overlap, thus indicatingrobust benefits ofmitigation. At the same time, there islittle difference between the two mitigation scenarios,given the overall uncertainty. Finally, the lack ofrobustness in the projections of changes in DD, eludedearlier, are further substantiated by this analysis, asthey are within the noise from natural variability evenby 2100. The comparison between the IGSM-CAMensemble and the CMIP5 ensemble reveals similarchanges in temperature-related indices. While theexact magnitude of the changes is not in precise agree-ment, the range is in agreement. Since the RCP scenar-ios and the CIRA scenarios were designedindependently, a simple metric like the radiative for-cing in 2100 is not sufficient to expect perfect

Figure 4.Projected changes in (a) accumulated frost days (AFD), (b) dry days (DD), (c) growing season length (GSL), (d) plant heatstress (PHS) and (e) start offield operations (SFO), area-averaged over theUS, in 2050 (2041–2060) and 2100 (2081–2100) relative topresent day (1991–2010) simulated by theMIT IGSM-CAM for all three emissions scenarios (REF, POL4.5 and POL3.7) and by 31CMIP5models for the RCP4.5 andRCP8.5 scenarios. Box plots show the range over the 15-member IGSM-CAMensemble withdifferent climate sensitivities and representations of natural variability for each emissions scenario and over the CMIP5multi-modelensemble (31models) under the RCP8.5 andRCP4.5. The black horizontal lines show themean of the IGSM-CAMsimulations forthemedium climate sensitivity (CS=3.0 °C) for each emissions scenario and themean over the 31CMIP5models for the RCP8.5andRCP4.5. Dark (light) gray shading represent the 1 (2) standard deviation of the natural variability estimated from a pre-industrialcontrol simulationwith the IGSM-CAM.

7

Environ. Res. Lett. 11 (2016) 055001

agreement. For the DD index, both ensembles projecta range of changes that span both increases anddecreases, but the CMIP5 range is wider. In addition,the CMIP5 ensemble projects a mean increase in DDwhile the IGSM-CAM mean is negative. At the sametime, the statistical significance of these changes islikely to be weak since they fall within the noise of nat-ural variability.

We explore in more detail the S/N by estimatingthe time of emergence, in each simulation, of the USmean changes in the five agro-climate indices, andcomputing the range for each emissions scenario (seefigure 5). This analysis reveals a large uncertainty in theestimates of the time of emergence, indicating very dif-ferent behaviors between indices and scenarios. Chan-ges in DD do not emerge from the noise before 2100for all three scenarios, confirming the lack of robust-ness in projections of precipitation changes. At thesame time, the time of emergence of changes in SFOoccurs between 2025 and 2050 in all the simulations,implying little impact of the mitigation scenarios. Forall other indices, the benefits of mitigation are clear.Under REF, changes emerge from the noise by 2070 atthe latest (and as early as 2020). The implementationof either mitigation scenarios allows for the possibilitythat the projected changes remain within the noise ofnatural variability by 2100. That is generally the casefor simulations with the lower climate sensitivity(CS=2.0 °C), which illustrates the need to account

for the uncertainty in the global climate systemresponse in analysis of the benefits of climate mitiga-tion. Finally, the difference between the two policyscenarios is generally small. It is only noticeable forprojections of changes in PHS—by reducing the prob-ability of emergence before 2100—and for changes inAFD—by increasing the mean estimate of the time ofemergence, but not its range.

4. Summary anddiscussion

Under climate change, this study generally projects theUS as a whole to experience fewer frosts, a longer GSL,an earlier SFO, an increase in heat stress and no robustchanges in DD. However, these changes have specificregional patterns. The northern US, especially over theRocky Mountains, the Great Lakes region and NewEngland project to benefit from less cold damage andearlier planting, ensuring maturation and the possibi-lity for multiple cropping—although fewer frostscould also lead to higher risks of pest and disease. Thesouthern US is expected to suffer from a stronger heatstress without the associated benefit of a significantincrease in the GSL. This north–south disparity inagro-climate projections is consistent with the changesin yield projected by [8]. The West is shown toexperience more heat stress and more DD, which

Figure 5.Time of emergence of the projectedUSmean changes relative to present day (1991–2010) in accumulated frost days (AFD),dry days (DD), growing season length (GSL), plant heat stress (PHS) and start offield operations (SFO) simulated by theMIT IGSM-CAM for all three emissions scenarios (REF, POL4.5 and POL3.7). Box plots show the range over the 15-member IGSM-CAMensemble with different climate sensitivities and representations of natural variability for each emissions scenario. The vertical blacklines show themean of the IGSM-CAMsimulations for themedium climate sensitivity (CS=3.0 °C) for each emissions scenario.The time of emergence is shown for a signal-to-noise ratio (S/N) greater than 2. Dotted lines past 2100 indicate that the signal has notemerged from the noise by 2100.

8

Environ. Res. Lett. 11 (2016) 055001

could result in declining yields and negative implica-tions forwater resources and irrigated agriculture.

These projections are associated with large uncer-tainties in magnitude, spatial pattern and even sign.This study finds that the magnitude of the changes islargely controlled by the climate sensitivity and emis-sions scenario. Meanwhile, natural variability cancause large differences in the regional patterns of theprojected changes—especially for changes in DD, GSLand PHS—including reversals of the sign of the pro-jected changes locally. That is true even using a 20 yearaveraging period, indicating that future changes inagro-climate indices are not always predictable locally.This study further highlights the lack of robustness ofprojected changes in precipitation, as demonstrated bythe lack of emergence of USmean changes in DD fromthe noise of natural variability, even by 2100. Our find-ings suggest that changes in temperature-related agro-climate indices are generally more predictable thanmetrics based on precipitation. The substantial role ofnatural variability on future climate projections hasgained a great deal of interest over the past few years[30, 31, 39, 41–44], and we hope this study sheds somelight on the implications for projections of future cli-mate change impacts on US agriculture. In particular,we caution policymakers and landmanagement stake-holders to properly examine the robustness of agro-climate projections when developing and evaluatingstrategies to adapt agriculture to climate change.

A comparison of the IGSM-CAM ensemble to theCMIP5 multi-model ensemble provides furtherinsight into the uncertainty in agro-climate projec-tions over the US.We find that a single climate model,with different emissions scenarios, climate sensitivitiesand representations of natural variability, simulates arange of changes similar to 31 different climate mod-els. Our analysis suggest that multi-model uncertaintycan be largely be explained by differences in the cli-mate sensitivity of the models and differences in initi-alization that leads to different representations ofnatural variability in the different simulations. Thisfinding is particularly true for temperature-relatedindices, but less so for projections of precipitationchanges, which are less robust to start with. While theIGSM-CAM projects both increases and decreases inDD, it does not reproduce the wide range of changessimulated by the CMIP5 ensemble. Nonetheless, weargue that sampling key sources of uncertainty in asingle climate model provides a complementary fra-mework to the commonly used multi-model ensem-ble. In addition, using an IAM to derive integratedeconomic and climate projections provides a majoradvantage, i.e. differences between scenarios can beattributed to an explicit choice of climate policy andthe benefits of greenhouse gas mitigation can bedirectly estimated (see [28, 45]). In contrast, becausethe underlying socio-economic trajectories and prio-rities are not consistent in the RCP scenarios, differ-ences between the scenarios cannot be attributed to

policy choices. Meanwhile, our framework evenallows us to compare two greenhouse mitigation sce-narios that only differ by the stringency of the carbontax applied.

The analysis shows that the projected changes inthe five agro-climate indices are significantly reducedunder the two policy scenarios compared to the refer-ence scenario, especially by 2100. On average, theimplementation of either greenhouse gas mitigationscenario cuts by half the changes projected under theunconstrained scenario. As a result, greenhouse gasmitigation has the potential to significantly reduceadverse effects of climate change (i.e. higher heatstress, higher risks of pest and disease from fewer frostdays), while also curtailing potentially beneficialimpacts (i.e. earlier planting, a longer growing seasonwith possibility for multiple cropping, and less frostdamage). We also find that climate mitigation canpotentially prevent changes in several indices toemerge from the noise of natural variability, even by2100. This is likely to be a major benefit from mitiga-tion given that any significant climate change impactson crop yield, whether beneficial or damaging, willresult in nation-wide changes in the agriculture sector.The cost of adaptation at that scale, such as the north-ward displacement of crop production, is difficult toquantify. Finally, we find that differences between thetwo mitigation scenarios are difficult to distinguishand that the benefits of mitigation are present in 2050,but small. The benefits of climate mitigation on pro-jections of future changes in agro-climate indices reso-nates with prior studies using crop models [3, 8]. Atthe same time, we realize that increases in CO2 con-centrations and adaptive management can providesignificant mitigation of the negative effects of climatechange [46–48].

5. Conclusion

This study shows that projections of agro-climateindices that are relevant to stakeholders can providegreat insight into the fate of future climate changeimpacts on agriculture. While these projections aresubject to substantial uncertainty, we show that usinga single climate model that accounts for key sources ofuncertainty (i.e. emissions scenario, global climatesystem response, natural variability) provides anefficient and complementary framework to the morecommon approach of multi-model ensemble (i.e.CMIP5 ensemble). We highlight the important role ofnatural variability, especially for projections ofchanges in DD and heat stress, leading to uncertaintyin themagnitude and location of the largest changes oreven the sign of the projected changes. For this reason,studies of climate change impacts on agriculture mustconsider these uncertainties by relying on largeensembles of climate projections that sample themajor sources of uncertainty—especially natural

9

Environ. Res. Lett. 11 (2016) 055001

variability. In addition, using integrated economic andclimate projections, we can directly estimate thebenefit of climate mitigation. We find that climatemitigation has substantial benefits: it cuts in half thechanges projected under an unconstrained scenario,and it potentially prevents changes from emergingfrom the noise of natural variability. Finally, we arguethat agro-climate indices, in combination with cropmodel projections, can provide valuable informationto better understand the drivers of changes in cropyield and production, thus better informing adapta-tion decisions.

Acknowledgments

This work was partially funded by the US Environ-mental Protection Agency’s Climate Change Division,under Cooperative Agreement#XA-83600001, by theUS Department of Energy, Office of Biological andEnvironmental Research, under grant DE-FG02-94ER61937, and by the National Science FoundationMacrosystems Biology Program Grant #EF1137306.The Joint Program on the Science and Policy of GlobalChange is funded by a number of federal agencies anda consortium of 40 industrial and foundation spon-sors. (For the complete list seehttp://globalchange.mit.edu/sponsors/current.html). This research usedthe Evergreen computing cluster at the Pacific North-west National Laboratory. Evergreen is supported bythe Office of Science of the US Department of EnergyunderContractNo.DE-AC05-76RL01830.

References

[1] Asseng S et al 2013Uncertainty in simulatingwheat yieldsunder climate changeNat. Clim. Change 3 827–32

[2] Rosenzweig C et al 2014Assessing agricultural risks of climatechange in the 21st century in a global gridded cropmodelintercomparison Proc. Natl Acad. Sci. USA 111 3268–73

[3] BeachRH et al 2015Climate change impacts onUS agricultureand forestry: benefits of global climate stabilization Environ.Res. Lett. 10 095004

[4] WiebeK et al 2015Climate change impacts on agriculture in2050 under a range of plausible socioeconomic and emissionsscenarios Environ. Res. Lett. 10 085010

[5] Lobell DB, Field CB, Cahill KN andBonfils C 2006 Impacts offuture climate change onCalifornia perennial crop yields:model projectionswith climate and crop uncertaintiesAgric.ForestMeteorol. 141 208–18

[6] SchlenkerWandRobertsM J 2009Nonlinear temperatureeffects indicate severe damages to us crop yields under climatechangeProc. Natl Acad. Sci. USA 106 15594–8

[7] Blanc É 2012The impact of climate change on crop yields inSub-SaharanAfricaAm. J. Clim. Change 1 1–3

[8] SueWing I,Monier E, SternA andMundraA 2015USmajorcrops’uncertain climate change risks and greenhouse gasmitigation benefitsEnviron. Res. Lett. 10 115002

[9] Challinor A J, SmithMS andThornton P 2013Use of agro-climate ensembles for quantifying uncertainty and informingadaptationAgric. ForestMeteorol. 170 2–7

[10] Quiring SMandPapakryiakouTN2003An evaluation ofagricultural drought indices for theCanadian PrairiesAgric.ForestMeteorol. 118 49–62

[11] Rhee J, Im J andCarboneG J 2010Monitoring agriculturaldrought for arid and humid regions usingmulti-sensor remotesensing dataRemote Sens. Environ. 114 2875–87

[12] Piao S et al 2010The impacts of climate change onwaterresources and agriculture inChinaNature 467 43–51

[13] Vicente-Serrano SM,Beguería S and López-Moreno J I 2010Amultiscalar drought index sensitive to global warming: thestandardized precipitation evapotranspiration index J. Clim.23 1696–718

[14] Narasimhan B and Srinivasan R 2005Development andevaluation of SoilMoistureDeficit Index (SMDI) andEvapotranspirationDeficit Index (ETDI) for agriculturaldroughtmonitoringAgric. ForestMeteorol. 133 69–88

[15] Snyder R, Pedras C,Montazar A,Henry J andAckleyD 2015Advances in ET-based landscape irrigationmanagementAgric.WaterManage. 147 187–97

[16] Feng S andHuQ2004Changes in agro-meteorologicalindicators in the contiguousUnited States: 1951–2000Theor.Appl. Climatol. 78 247–64

[17] TerandoA, Keller K and EasterlingWE2012 Probabilisticprojections of agro-climate indices inNorth AmericaJ. Geophys. Res. 117D08115

[18] HardingA, RivingtonM,MineterM andTett S 2015Agro-meteorological indices and climatemodel uncertainty over theUKClim. Change 128 113–26

[19] MatthewsK, RivingtonM, BuchanK,MillerD andBellocchiG2008Characterising the agro-meteorological implications ofclimate change scenarios for landmanagement stakeholdersClim. Res. 37 59

[20] Waldhoff S T,Martinich J, SarofimM,DeAngelo B,McFarland J, Jantarasami L, Shouse K, Crimmins A,Ohrel S and Li J 2015Overview of the special issue: amulti-model framework to achieve consistent evaluation ofclimate change impacts in the United StatesClim. Change131 1–20

[21] GaoX andDirmeyer PA 2006Amultimodel analysis,validation, and transferability study of global soil wetnessproducts J. Hydrometeorol. 7 1218–36

[22] Dirmeyer PA,GaoX, ZhaoM,GuoZ,Oki T andHanasakiN2006GSWP-2:multimodel analysis and implications for ourperception of the land surfaceBull. Am.Meteorol. Soc. 871381–97

[23] Kunkel K E, EasterlingDR,HubbardK andRedmondK 2004Temporal variations in frost-free season in theUnited States:1895–2000Geophys. Res. Lett. 31 L03201

[24] Walsh J et al2014Our changing climateClimateChange Impactsin theUnited States: TheThirdNationalClimateAssessment edJMelillo et al (USGlobalChangeResearchProgram) ch2, pp19–67 (http://s3.amazonaws.com/nca2014/low/NCA3_Full_Report_02_Our_Changing_Climate_LowRes.pdf?download=1)

[25] LinderholmHW2006Growing season changes in the lastcenturyAgric. ForestMeteorol. 137 1–4

[26] Monier E, Scott J R, Sokolov A P, Forest C E andSchlosser CA 2013 An integrated assessmentmodelingframework for uncertainty studies in global and regionalclimate change: theMIT IGSM-CAM (version 1.0)Geosci.Model Dev. 6 2063–85

[27] VanVuurenDP et al 2011The representative concentrationpathways: an overviewClim. Change 109 5–31

[28] Paltsev S,Monier E, Scott J, Sokolov A andReilly J 2015Integrated economic and climate projections for impactassessmentClim. Change 131 21–33

[29] SokolovA andMonier E 2012 changing the climate sensitivityof an atmospheric general circulationmodel through cloudradiative adjustment J. Clim. 25 6567–84

[30] Monier E, GaoX, Scott J R, Sokolov AP and Schlosser CA 2015A framework formodeling uncertainty in regional climatechangeClim. Change 131 51–66

[31] Monier E andGaoX 2015Climate change impacts on extremeevents in theUnited States: an uncertainty analysisClim.Change 131 67–81

10

Environ. Res. Lett. 11 (2016) 055001

[32] Taylor KE, Stouffer R J andMeehl GA 2012An overview ofCMIP5 and the experiment designBull. Am.Meteorol. Soc. 93485–98

[33] RieneckerMM et al 2011MERRA:NASA’smodern-eraretrospective analysis for research and applications J. Clim. 243624–48

[34] Kalnay E et al 1996TheNCEP/NCAR40 year reanalysisprojectBull. Am.Meteorol. Soc. 77 437–71

[35] Hawkins E and SuttonR 2012Time of emergence of climatesignalsGeophys. Res. Lett. 39 L01702

[36] Mahlstein I, Knutti R, Solomon S and PortmannR 2011 Earlyonset of significant local warming in low latitude countriesEnviron. Res. Lett. 6 034009

[37] Giorgi F andBi X 2009Time of emergence (toe) of ghg-forcedprecipitation change hot-spotsGeophys. Res. Lett. 36 L06709

[38] DeserC, Terray L and Phillips A S 2016 Forced and internalcomponents of winter air temperature trends overNorthAmerica during the past 50 years:mechanisms andimplications J. Clim. in press (doi:10.1175/JCLI-D-15-0304.1)

[39] Hawkins E and SuttonR 2011The potential to narrowuncertainty in projections of regional precipitation changeClim.Dyn. 37 407–18

[40] Monier E, SokolovA, Schlosser A, Scott J andGaoX 2013Probabilistic projections of 21st century climate change overNorthern Eurasia Environ. Res. Lett. 8 045008

[41] Hawkins E and SuttonR 2009The potential to narrowuncertainty in regional climate predictionsBull. Am.Meteorol.Soc. 90 1095–107

[42] Hawkins E 2011Our evolving climate: communicating theeffects of climate variabilityWeather 66 175–9

[43] DeserC, Knutti R, Solomon S and Phillips A S 2012Communication of the role of natural variability in futureNorthAmerican climateNat. Clim. Change 2 775–9

[44] DeserC, Phillips A, Bourdette V andTengH2012Uncertaintyin climate change projections: the role of internal variabilityClim.Dyn. 38 527–46

[45] Reilly J et al 2013 Valuing climate impacts in integratedassessmentmodels: theMIT IGSMClim. Change 117561–73

[46] Padgham J 2009Agricultural DevelopmentUnder aChangingClimate: Opportunities andChallenges for AdaptationWorldBank (http://siteresources.worldbank.org/INTARD/Resources/climate_change_combined.pdf)

[47] Stöckle CO,NelsonR L,Higgins S, Brunner J, GroveG,BoydstonR,WhitingMandKruger C 2010Assessment ofclimate change impact on EasternWashington agricultureClim. Change 102 77–102

[48] Lobell D B andGourdji SM 2012 The influence of climatechange on global crop productivity Plant Physiol. 1601686–97

11

Environ. Res. Lett. 11 (2016) 055001