SUPPLEMENTARY INFORMATION

Size: px
Start display at page:

Download "SUPPLEMENTARY INFORMATION"

Transcription

1 SUPPLEMENTARY INFORMATION DOI: 1.138/NCLIMATE1585 Adaptation of US maize to temperature variations Ethan Butler and Peter Huybers Department of Earth and Planetary Sciences Harvard University 2 Oxford St., Cambridge MA, 2138 Contents S1 Supplemental Figures 2 S2 Case Study: Butler County 12 S3 Comparison with Schlenker and Roberts, NATURE CLIMATE CHANGE 1

2 S1 Supplemental Figures 5 N 4 N 3 N 11 W 1 W 9 W 8 W R 2 Figure S1: Squared-cross-correlation between yield data and individual county models. The results for all displayed counties are significant (Fig. S2). Irrigated counties are outlined in black

3 5 N 4 N 3 N 11 W 1 W 9 W 8 W p value Figure S2: Model significance. P-values from an F-test comparing the full model to a reduced model including only an intercept and time trend for all 1666 counties, with the colorbar truncated at p=.5. Omitted counties with p-values greater than.5 are in deep red and counties omitted for insufficient data are in gray. The final pool of counties is 113. States not reporting planting and harvest times are shown in white. Note that a scatter plot of p-values and KDD sensitivity had no apparent structure.

4 .1 Killing Degree Day Sensitivity (β 3 ) Mean Killing Degree Days Figure S3: KDD sensitivity versus mean KDDs for irrigated counties. The fit to irrigated counties is substantially weaker than for rainfed counties (linear (gray) R 2 =.12, logarithm (red) R 2 =.21, and inverse (gray) R 2 =.25) possibly because of the differing levels of irrigation between counties. For consistency, we use the logarithm of mean KDD to model the adaptation of irrigated counties.

5 5 N 4 N 3 N 11 W 1 W 9 W 8 W Δ GDD 2 C Warming Figure S4: Increase in Growing Degree Days from a 2 C warming. Changes in each counties Growing Degree Days correspond well with the growing season length of each county. The majority of days during the growing season are well above the minimum threshold for GDD accumulation so that changes in GDDs are essentially linear with changes in mean temperature. Irrigated counties are outlined in black.

6 5 N 4 N 3 N 11 W 1 W 9 W 8 W Δ KDD 2 C Warming Figure S5: Increase in Killing Degree Days from a 2 C warming. Changes in each county s Killing Degree Days correspond with the local mean temperature in each region. Cold counties are generally below the 29 C threshold and accumulate few KDD with modest warming, while hot counties are already near or above the threshold and accumulate more KDDs. This helps explains why hot counties incur greater losses from the the warming, despite having a lower sensitivity. Irrigated counties are outlined in black.

7 5 N N N W 1 W 9 W 8 W Yield/Growing Degree Day Figure S6: Climatological GDDs (contours) and GDD sensitivity (shading). Climatological GDDs are weakly negatively correlated with unirrigated GDD sensitivity (r=-.3). The regions with the shortest growing season tend to have the highest GDD sensitivity. In the North, where the growing season is potentially too short for the crops to reach maturity, yields respond well to seasons with above average GDD; whereas in the South, where GDDs are abundant, yield is largely insensitive to fluctuations in GDD.

8 .25.2 GDD Sensitivity (β 2 ) Mean Growing Degree Days Figure S7: GDD sensitivity versus GDD climatology. There is a weak relationship between GDD climatology and GDD sensitivity (Fig. S6). As with KDD, this fit was performed separately for irrigated counties. Counties with extremely low sensitivity to GDD (GDD sensitivity <.1) were omitted from this fit (shown in grey) and the regression was weighted by the inverse of the bootstrap variance estimate of the GDD sensitivity parameter. The net effect of reducing GDD sensitivity as GDD climatology warms is a mean yield loss of 8% from a 2 C warming, as opposed to a 6% loss when changes in GDD sensitivity are ignored.

9 5 N 4 N 3 N 11 W 1 W 9 W 8 W Correlation Coefficient Figure S8: Correlation between mean KDD and mean precipitation. The correlation is almost everywhere negative between mean KDD and mean precipitation. The strong negative correlation in the south further supports that local adaptation accounts for the low sensitivity to KDD, as opposed to compensating meteorological factors. In general, years with high KDD are accompanied by low precipitation.

10 54 53 Mean Sum of Squared Resdiduals Optimum Temperature Figure S9: Goodness-of-fit for the temperature threshold above which killing degree days are computed. The goodness-of-fit is measured as the sum of the squared residuals when fitting Eq. 1 in the main manuscript to the county-level yield data. The best fit is achieved with a temperature threshold of 3 C, but we use a threshold of 29 C throughout this study for consistency with previous work. Use of a 3 C threshold reduces the estimated losses from a 2 C warming to 11% without adaptation and 4% with adaptation. Note that although the magnitudes differ according to the threshold, the relative differences are comparable using either threshold, supporting our main conclusion that adaptation has major implications for forecasting the effects of climate change upon crop yield.

11 .1.2 log((bushels/acre))/kdd Mean Killing Degree Days Figure S1: Alternative analysis using the logarithm of yield. Other studies (i.e. refs 1 and 2) often use log(yield) instead of magnitude. The basic relationship between climatological KDD and log(yield) sensitivity is quite similar to that between climatological KDD and yield sensitivity (Fig. 2). The y-axis has been truncated for clarity at -.1, omitting the four most sensitive values, all of which had very low mean KDDs, though the points were included in the fit.

12 S2 Case Study: Butler County It is useful to consider a specific case study as a means of illustrating the overall analysis process, and here we focus on Butler County, Ohio, at 39.4 N and 84.5 W. Butler County s mean yields are 118 bushels/acre relative to the average of 18 bushels/acre over all counties. Butler is one of the better modeled counties, with the model explaining 86% of the yield variability (in the 75 th percentile of variance explained, fig. S11a), but the results presented here are nonetheless generalizable to results from other counties. An obvious relationship exists between peaks in KDD anomaly and troughs in annual yield from Butler (Fig. S11b). Furthermore, Butler serves to illustrate how GDD and KDD vary relatively independently. For example, the worst yields for Butler in this study occurred in 1983 (52.3 bushels/acre) with a GDD anomaly of +3.6 degree-days and a KDD anomaly of +2 degree-days. In 1985, a year with one more GDD but a KDD anomaly of degree-days, the yields were over double those in 1983 (122.8 bushels/acre). This demonstrates the great sensitivity of yields in Butler to variations in KDD, and a central question for determining Butler s response to future warming is whether such large losses could be mitigated if higher levels of KDDs become the norm. In the case of no adaptation, our model predicts that a 2 C warming would lead to a 19% reduction in yield from 118 to 96 bushels/acre (Fig. S11c) in Butler, whereas with adaptation, mean yields decline by 11%. Losses with and with-out adaptation are calculated by running the model with historical temperatures plus 2 C, giving a different time-history of GDDs and KDDs (Fig. S11d). For example, in year 1 our model gives a 3 bushels/acre greater loss without adaptation, relative to with adaptation, whereas in years 3 and 27 there is a 21 bushels/acre greater loss. (Years are reported generically under the warming scenario with year 1 corresponding to 1981 and year 28 corresponding to 28 in the historic record.) All statistics that we report in the main manuscript are calculated by averaging across years for which historical data are available, which for the future warming scenario implicitly assumes that the distribution of temperatures around a given mean is unchanged.

13 Yield [Bushels/Acre] Historic Yield Predicted Yield Yield [Bushels/Acre] Projected Yield with Adaptation Projected Yield without Adaptation (a) Year (c) Year [with 2 C warming] GDD anomaly (b) Year KDD anomaly Anomaly relative to Historic Mean GDD KDD (d) Year [with 2 C warming] Figure S11: Case study for Butler, OH. (a) Butler, OH yields (black) and model fit (dashed). (b) Yearly variation in GDD and KDD anomaly. (c) Estimated yield without adaptation (dashed) for a 2 C warming and with adaptation (solid). Note that year 1 corresponds to 1981 and year 28 corresponds to 28 from the historical record. (d)the comparable increase in GDD anomaly is not enough to offset the effects of increased KDDs, even with a reduced KDD sensitivity. Note that while anomalies increase by similar amounts, the fractional increase in KDD is substantially larger. There are missing data for Butler County in years 1993, 1995, 1996, 1997, 1999, and 2 because of gaps greater than one week in the maximum temperature data obtained from station 1213 at 39.4 N and 85. W.

14 S3 Comparison with Schlenker and Roberts, 29 In this section we seek to reconcile our results with those of Schlenker and Roberts (29, see ref. 1, hereafter SR9) who, in supplemental material, offer an argument that the adaptation of different regions to heat stress does not have implication for yield loss resulting from warming. In particular, SR9 show a remarkable consistency between the expected yield loss when using temperature transfer functions derived from Southern counties as when using a temperature transfer function derived from the entire dataset of Eastern counties (SR9, Fig. A9). This result calls into question whether regional adaptation implies any degree of adaptability to long term climate change, and appears contrary to the findings in the main text. There are, however, several possible explanations for the insensitivity of fractional yield losses to selection of regional adaptation functions. SR9 indicate that this regional insensitivity is a tradeoff between the temperature at which yield damages are incurred and the magnitude of those damages. In particular, they note that although yield is more sensitive to high temperatures in the North than in the South, damages are incurred at lower temperatures in the South. (No biophysical interpretation is offered, but one possibility relates to moisture availability and overall lower water availability and/or greater demand in Southern counties, as opposed to those in the North.) A second possibility is that comparison of the results of the Southern adaptation function to the adaptation function derived from the whole of the Eastern counties fails to adequately capture regional distinctions. SR9 divide Eastern U.S. counties into Northern, Middle, and Southern groups, but only compare the Southern to the all-county estimate. The differences between their Southern transfer function relative to the all-county estimate is much smaller than between the Southern county estimates and either the Middle or Northern estimates. The similarity between the Southern and all-county adaptation functions may occur because the Southern region contains more counties than are in either the Middle or Northern group. Furthermore, the magnitude of losses estimated for high temperatures in the Middle and Northern transfer functions appears roughly double that of the Southern function, whereas the Southern and all-county transfer functions are roughly equal. To explore these possibilities in greater depth and to facilitate direct comparison of SR9 s results with ours, we have applied SR9 s methodology to the dataset that we use in the main text. Three other differences from the treatment of the data in the main text are also adopted in order to facilitate direct comparison. First, irrigated and unirrigated crops are combined together. While SR9 accounted for irrigated counties by omitting counties west of 1 W, there are a number of irrigated counties east of this meridian according to the USDA/NASS database (ref. 2). Irrigated counties tend to have lower sensitivity to high temperatures. Thus, this combination likely leads to lower sensitivity and a worse overall fit of a single temperature response function to the yield dataset than when only unirrigated crops are included. Second, we use the logarithm of yield, as opposed to the magnitude of yield. As reported in the main text, the changes in yield for a 2 C warming with and without adaptation are similar when using the logarithm and magnitude, and we find a similar result when using SR9 s method, though only focus on the logarithmic case here. Third, following SR9, a fixed growing season is used across the entirety of the domain. This leads to inclusion of temperature data that had no direct influence on yield because, for example, planting occurs as much as two months earlier in some Southern counties relative to Northern ones. The argument has been made that including time variable planting may be undesirable because

15 it introduces endogeneity between temperature and yield residuals (D. Lobell, private communication). Our view is that yield data does, in fact, reflect variable planting and harvesting time and that inferring sensitivity to temperature variations is better undertaken using the actual temperature forcing. Future studies should explore the interdependence between temperature and the times of planting and harvesting both for purposes of model fitting and estimation of adaptability. That said, we have computed results both within this section and for those reported in the main text using fixed and variable planting and harvesting times. Although the temperature distributions change, the overall pattern of response is consistent, and in this section we only report results using fixed planting and harvesting time The transfer functions that we estimate from temperature to yield (Fig. S12) are similar to that of SR9 both in magnitude and shape. Like with SR9, the Southern county transfer function shows lower magnitudes of sensitivity than the Middle and Northern regions. However, there is no discernible trade-off between this sensitivity and the temperature at which damages begin to be incurred. Note also that the all-eastern county transfer is most similar to the Northern county result, as opposed to the Southern county estimate. This is, in fact, consistent with the simple observation noted earlier that the region having the most counties will tend to control the overall result. In the present dataset, which is the same as in the main manuscript, many Southern counties are excluded on account of missing data or lacking planting and harvesting times. Using the technique of SR9, we find that a 2 C warming leads to an average yield loss of 1%, similar to the 11% loss we find using our technique with a logarithmic transformation of yield. Importantly, however, SR9 s technique gives 8% losses when using the Southern transfer function and 13% losses when using the Northern transfer function, indicating that Northern crops are more sensitive to temperature increases and consistent with the findings reported in the main manuscript. Given substantial regional differences in climate and differences in the estimated crop response, it is also useful to examine the county-level patterns of changes in yield (Figs. S13-S15). A map of fractional yield losses calculated for a 2 C warming shows that using the all-eastern-county transfer function gives large losses in the South and trending toward small gains in the North (Fig. S14). This regional pattern is consistent with the results reported in the main text (c.f. Fig. 3). Note that inclusion of more Southern counties would almost certainly increase the average yield loss, and that this likely explains why SR9 found higher average yield losses. When the Southern transfer function is applied for a two C warming (Fig. S15), the fractional losses in the South are decreased by as much as 1%. Conversely, if the Northern transfer function is applied, losses in the South are exacerbated by as much as 1%, whereas yields in the North increase even further. Perhaps unsurprisingly, the optimal transfer function for purposes of mitigating losses from warming differs according to region. These results provide context for interpreting SR9 s finding of little overall change in sensitivity when applying their all-county transfer function and that derived from the South. This insensitivity appears to result from two facts: that their all-county transfer function is similar to the Southern transfer function, and that application of the Southern regional transfer function is better for Southern counties but worse for Northern counties such that the overall effect on yield is similar to the all-county results. Furthermore, when looking at regional responses, we find that regional transfer functions tend to be more optimal for the regions from which they are derived, thus indicating spatial adaptation in agreement with the results in the main manuscript.

16 .2.2 (a) (b) log(yield) anomaly.2.4 log(yield) anomaly distribution temperature ( C).2 distribution temperature ( C).2 (c) (d) log(yield) anomaly.2.4 log(yield) anomaly distribution temperature ( C) distribution temperature ( C) Figure S12: Damage functions computed using the methodology of Schlenker and Roberts (29) but using the data described in this manuscript. Results are for (a) theentireeasternu.s.andregional subsets covering (b) the South, c) the middle, and (d) the North. Regional subsets follow those shown in SR29, Figure A7, and the results can be directly compared with their Figures 1a and A8. Although we use a more tightly screened and shorter datasets, there is generally a close correspondence between these results and those of SR29.

17 5 N 4 N 3 N 11 W 1 W 9 W 8 W US fractional yield loss Figure S13: Fractional changes in yield from a 2 C warming using the temperature sensitivity function shown in Fig. S12a. Importantly, almost all of the damages occur in the South and the North actually showsagaininyield.

18 5 N 4 N 3 N 11 W 1 W 9 W 8 W yield anomaly using Southern damage function Figure S14: Similar to S13 but for the difference between the all-eastern-county results shown there and when using the transfer function associated with Southern counties (i.e. Fig. S12b).

19 5 N 4 N 3 N 11 W 1 W 9 W 8 W yield anomaly using Northern damage function Figure S15: Similar to S13 but for the difference between the all-eastern-county results shown there and when using the transfer function associated with Northern counties (i.e. Fig. S12d).