Translating Climate Forecasts into Agricultural Terms: Advances and Challenges James Hansen, Andrew Challinor, Amor Ines, Tim Wheeler, Vincent Moron presented at the International Workshop on Climate Prediction and Agriculture Advances and Challenges WMO, Geneva, 11 May 2005 Motivation Information relevant to decisions Ex-ante assessment for credibility and targeting Fostering and guiding management Overview Six years ago Dominance of historic analogs Doubts about crop predictability Recent advances

The challenge, and potential approaches Synthetic weather conditioned on climate forecasts Use of daily climate model output Statistical prediction of crop simulations Downscaling and upscaling Opportunities and challenges Embedding crop models within climate models Enhanced use of remote sensing, spatial data bases Robustness of alternative coupling approaches Forecast assessment and uncertainty Climate research questions Six Years Ago: Dominance of Historic Analogs Advantages Concerns Small sample size, confidence, artificial skill Are differences in distribution real? How to use with dynamic prediction systems without discarding information?

Variance explained Intuitive probabilistic interpretation Accounts for any differences in signal strength May incorporate useful higher-order statistics 50% 40% 30% 20% 10% 0% 3 6 12 16 Number of "phases" cross-validated no cross-validation Six Years Ago: Doubts About Crop Predictability Spatial variability of rainfall limits predictability at farm scale Accumulation of error from SSTs, to local climatic means, to crop response

Impact of wrong forecast on farmers risk Barrett, 1998. Am. J. Agric. Econ. 80:1109-1112 The Challenge Nonlinearities. Crop response to environment can be nonlinear, non-monotonic. G ra in y ie ld (M g /h a ) 5 4 3 2 1 0 0 200 400 600 Dynamics. Crops respond not to mean conditions but to dynamic interactions: Soil water balance Phenology

The scale mismatch problem. 800 OND precipitation (mm) The Scale Mismatch Problem Crop models: Homogeneous plot spatial scale Daily time step (w.r.t. weather) GCMs: Spatial scale 10,000-100,000 km2 Sub-daily time step, BUT... Output meaningful only at (sub)seasonal scale Tend to over-predict rainfall frequency, under-predict mean intensity Temporal scale problem more difficult than spatial scale. Effect of Spatial Averaging Inverse-distance interpolation of daily weather data, north Florida, at a scale comparable to a GCM grid cell. Hansen & Jones, 2000. Agric. Syst. 65:43-72.

Effect of Spatial Averaging Spatial averaging distorts variability, increases frequency, decreases mean intensity. In ten sity (m m /d ) R elative freq uency R ain fall to tal (m m ) Similar spatial averaging occurs within GCM. 0.8 150 0.6 100 0.4 50 0.2 observed 0 12 interpolated

Jan Mar May Jul Sep Nov Month 0.0 Jan Mar May Jul Sep Nov Month 8 4 0 Jan MarMay Jul SepNov Month Effect of Spatial Averaging Simulated maize yields, CERES-Maize Information Pathways observed climate predictors ?

predicted crop yields Information Pathways observed climate predictors categorize analog years crop model predicted crop yields Information Pathways observed climate predictors

categorize statistical climate model analog years downscaled dynamic model stochastic generator crop model hindcast ( weather) predicted crop yields Information Pathways observed climate predictors

categorize statistical climate model analog years downscaled dynamic model crop model (observed weather) stochastic generator crop model hindcast ( weather) predicted crop statistical yield model yields Approaches Classification and selection of historic analogs (e.g., ENSO phases)

Synthetic daily weather conditioned on forecast: stochastic disaggregation Statistical function of simulated response Nonlinear regression Linear regression with transformation or GLM Probability-weighted historic analogs (Corrected) daily climate model output Advances: Synthetic Weather Inputs Two Approaches: Adjusting generator input parameters: Flexibility to produce statistics of interest Assumed role of intensity vs. frequency Constraining generator outputs: No assumptions re. frequency vs. intensity Option 2. Constraining generated output First step: - Iterative procedure Using climatological parameters, accept the first realization with Rm near target: |1-Rm/Rm,S|j <= 5% Second step: - Apply multiplicative rescaling to exactly match target monthly target. Hansen & Ines, Submitted. Agric. For. Meteorol. Constraining generator outputs reproduces correlations better than adjusting inputs.

Scenario Tifton, Georgia Gainesville, Florida RM vs. RM vs. I I vs. RM vs. RM vs. I I vs. Observed daily rainfall 0.649 0.577 -0.165 0.668 0.706 0.046 Disaggregated monthly rainfall constrain RM 0.681 0.676 -0.004

0.649 0.697 0.014 condition condition I 0.822 0.473 0.013 0.831 0.121 0.052 0.491 0.856 0.071 0.458 0.837 0.052

Constraining generator output requires fewer replicates for given accuracy. Tifton a b 1.0 1.5 1.5 0.8 0.6 1.0 1.0 c 1.2 2.0 2.0 Katumani

1.4 0.4 0.5 0.5 0.0 1.0 0.0 1.0 d 0.2 0.0 1.0 e 0.8 0.8 0.8 0.6 0.6

0.6 0.4 0.4 0.4 0.2 0.2 0.2 f R RMSE, Mg ha -1 2.5 Gainesville 2.5 0.0 1 10

100 0.0 1000 1 constrain RM condition I condition 0.0 10 100 Number of realizations 1000 1 10 100 1000 Maize simulated from disaggregated monthly GCM hindcasts, Katumani, Kenya Simulated

2000 7000 R=0.44 6000 1600 Rm pi 1400 Yield, kg ha -1 1800 RMSE, kg ha-1 Rm Hindcasts 5000 4000 3000 2000 1000 1200

0 1960 1000 1 10 100 1965 1970 1975 1000 1980 1985 1990 1995 2000 2005 Year

N o . o f r e a liz a t io n s 0 .6 Simulated Hindcasts 7000 6000 R=0.41 0 .5 0 .4 pi 0 .3 0 .2 Yield, kg ha -1 R Rm

5000 4000 3000 2000 1000 0 .1 1 10 100 N o . o f r e a liz a t io n s 1000 0 1960 1965 1970 1975 1980 1985 Year

1990 1995 2000 2005 Advances: Use of Daily Climate Model Output Options Calibrate simulated yields Challinor et al., 2005. Tellus 57A:198-512 Correct GCM mean bias Additive shift for temperatures Multiplicative shift for rainfall Rainfall frequency-intensity correction Ines & Hansen, In preparation Correcting Bias in Daily GCM Output: Rainfall Frequency 1 F(xhist=0.0) F(xGCM=0.0) GCM Historical 0

0 calibrated threshold Correcting Bias in Daily GCM Output: Rainfall Intensity 1 obs ,m xi F (FGCM ,m ( xi )) 1 F(x) F(xi) 0 GCM Historical 0 xi 0 x'i

Daily rainfall (x), mm ECHAM4 & observed OND 0.8 daily rainfall (1970-95) 0.9 0.7 Intensity corrections: 0.5 0.4 0.3 0.2 EG: empirical (GCM) to gamma (observed) GG: gamma (GCM and 0.0 1970 1975 1980 1985 observed) Mean rainfall frequency, -1wd d -1 0.6 0.1

1990 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 14 1970 1975 1980 1985 1990 1975 1980 1985 1990

1995 1980 1985 1990 1995 12 1995 10 8 6 4 2 Mean rainfall amount, mm d Katumani, Kenya -1 Mean rainfall intensity, mm wd Corrects rainfall total, frequency, intensity.

0.9 0 81970 7 Obs 6 EG GG 5 4 Rm Uncorr 3 2 1995 1 0 1970 1975

Year Mean monthly rainfall (Rm), mm d-1 Predicts yields from GCM, perhaps better than stochastic disaggregation 8 7 6 Obs rMOS=0.59 MOS rGG =0.74 GG 5 4 3 2 1995 1 0 1970 1975

1980 1985 Year CERES-Maize simulated with: Disaggregated MOS-corrected monthly hindcasts Gamma-gamma transformation of daily rainfall 1990 1995 Advances: Statistical Prediction of Crop Simulations Seasonal predictors of local climate potential predictors of crop response Predictand: Yields simulated with observed weather Eliminates need for daily weather conditioned on climate forecast Poor statistical behavior Nonlinear Regression Katumani maize prediction example: Yields as f(PC1) Mitscherlitch functional form: y a b 1 e x p c x

y=3.33+1.34(1-exp(-0.133x)) R2 = 0.400 Cross-validation K Nearest Neighbor Unequally-weighted analogs Weights w: Based on rank distance (predictor state space) wj Interpreted as probabilities Forecast a weighted mean: yt Optimize k A non-parametric regression 1/ j k 1/ i i 1 n wy i i 1,i t i

Linear Regression & Transformation: Regional-Scale Wheat, Qld, Australia Wheat simulations: water satisfaction index ECHAM4.5, persisted SSTs, optimized (MOS) Yield prediction by c-v linear regression Box-Cox normalizing transformation Forecast distribution: Regression residuals in transformed space n antecedent X n within-season weather years Hansen et al., 2004. Agric. For. Meteorol. 127:77-92 Linear Regression & Transformation: Regional-Scale Wheat, Qld, Australia N 1 May 1 June 1 July 1 August 200 0 200 400 km Correlation

<0.34 (n.s.) 0.34-0.45 0.45-0.55 0.55-0.65 0.65-0.75 0.75-0.85 > 0.85 Climatology GCM-based ENSO phase 2.0 Grain yield (Mg ha-1) 1.5 1.0 2.0 1.5 1.0 0.5 2.0 1.5 Observed 90th percentile 75th percentile

50th percentile 25th percentile 10th percentile 1.0 0.5 1May1Jun 1Jul 1Aug Harvest1May1Jun 1Jul 1Aug Harvest1May1Jun 1Jul 1Aug Date of forecast Harvest 1989 (neutral) 1988 (La Nia) 0.5 1982 (El Nio) Linear Regression & Transformation: Regional-Scale Wheat, Qld, Australia Advances: Downscaling & Upscaling

C o rre la tio n 0.8 0.6 0.4 Spatial climate downscaling: 0.2 Jan-Mar Apr-Jun 0.0 point ~11~21 state Scale Methods advancing Obs. vs. pred. rainfall, Cear, NE Brazil, as function of aggregation. Gong et al., 2003. J. Climate 16:3059-71. Uncertain impact on skill Crop model upscaling: Understanding and methods for aggregating point models Increasing set of reduced form large-area models Predictability (r) of groundnut yields with large area model, W India. Challinor et

al., 2005. Tellus 57A:198-512 Opportunities & Challenges: Crop Models Within Climate Models Run crop models within GCM or RCMs Allow crop to influence atmosphere Alternative land surface scheme Intended benefit is atmosphere response to crop Likely to require calibration of crop results for foreseeable future Match scale of climate model grid Opportunities & Challenges: Remote Sensing, Spatial Data Bases Enhanced georeferenced soil, land use, cultivar data bases Assimilation of real-time, contiguous antecedent weather into forecasts Estimation of cropped areas, dates Correction of simulated state variables Eventual farm-specific crop forecasts? Opportunities & Challenges Robustness of Alternative Approaches? 5 4 3 2 1 0

5 r = 0.57 Regression r = 0.58 Stochastic disaggregation 4 3 2 1 0 5 k nearest neighbors, 1 PC r = 0.53 r = 0.55 Stochastic disaggregation + 1960 4 1970 1980 1990

3 2 r = 0.53 1 k nearest neighbors, 2 PCs observed predicted 0 1960 1970 1980 1990 Hansen & Indeje, 2004. Agric. For. Meterol. 125:143 Opportunities & Challenges: Forecast Assessment and Uncertainty Does predictability (climate and impacts) change from year to year? Artifact of skewness? Real impacts of climate state? Captured by GCM ensembles? Interpretation of forecasts based on categorical vs. continuous predictors?

Consistency of hindcast error vs. GCM ensemble distributions? Are differences in dispersion real? Raw Transformed skewness 1.243 p ENSO influence on: means 0.0001 *** dispersion 0.0001 *** -0.032 0.0004 *** 0.91 n.s. December rainfall Junin, Argentina, 1934-2001 La Nina neutral ENSO phase El Nino

La Nina neutral ENSO phase El Nino Opportunities & Challenges: Forecast Assessment and Uncertainty Does predictability (climate and impacts) change from year to year? Artifact of skewness? Real impacts of climate state? Captured by GCM ensembles? Interpretation of forecasts based on categorical vs. continuous predictors? Consistency of hindcast error vs. GCM ensemble distributions? Opportunities & Challenges: Climate Research Questions Past prediction efforts driven by skill Relative shifts Large areas 3-month climatic means Stimulating interest in weather within climate

Skill at sub-seasonal time scales Higher-order rainfall statistics Shifts in timing, onset, cessation Methods to translate into weather realizations