Translating climate forecasts into agricultural terms:

advances and challenges

Jw Hansen, A. Challinor, A Ines, T Wheeler, Vincent Moron

To cite this version:

Jw Hansen, A. Challinor, A Ines, T Wheeler, Vincent Moron. Translating climate forecasts
into agricultural terms: advances and challenges. Climate Research, 2006, 33 (7-8), pp.27-41.
�10.3354/cr033027�. �hal-02894588�

HAL Id: hal-02894588

https://hal.science/hal-02894588
Submitted on 6 Sep 2021

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.

Distributed under a Creative Commons Attribution| 4.0 International License

 CLIMATE RESEARCH
Vol. 33: 27–41, 2006 Published December 21
Clim Res

OPEN
ACCESS

Translating climate forecasts into agricultural

terms: advances and challenges
James W. Hansen1,*, Andrew Challinor2, Amor Ines3, Tim Wheeler4, Vincent Moron1
1
International Research Institute for Climate and Society, 61 Route 9W, Palisades, New York 10964-8000, USA
2
Department of Meteorology, University of Reading, Reading RG6 6BB, UK
3
Department of Biological and Agricultural Engineering, Texas A&M University, College Station, Texas 77843, USA
4
Department of Agriculture, University of Reading, PO Box 236, Reading RG6 6AR, UK

ABSTRACT: Seasonal climate prediction offers the potential to anticipate variations in crop production
early enough to adjust critical decisions. Until recently, interest in exploiting seasonal forecasts from
dynamic climate models (e.g. general circulation models, GCMs) for applications that involve crop
simulation models has been hampered by the difference in spatial and temporal scale of GCMs and
crop models, and by the dynamic, nonlinear relationship between meteorological variables and crop
response. Although GCMs simulate the atmosphere on a sub-daily time step, their coarse spatial
resolution and resulting distortion of day-to-day variability limits the use of their daily output. Crop
models have used daily GCM output with some success by either calibrating simulated yields or
correcting the daily rainfall output of the GCM to approximate the statistical properties of historic
observations. Stochastic weather generators are used to disaggregate seasonal forecasts either by
adjusting input parameters in a manner that captures the predictable components of climate, or by
constraining synthetic weather sequences to match predicted values. Predicting crop yields, simulated
with historic weather data, as a statistical function of seasonal climatic predictors, eliminates the need
for daily weather data conditioned on the forecast, but must often address poor statistical properties of
the crop–climate relationship. Most of the work on using crop simulation with seasonal climate fore-
casts has employed historic analogs based on categorical ENSO indices. Other methods based on clas-
sification of predictors or weather types can provide daily weather inputs to crop models conditioned
on forecasts. Advances in climate-based crop forecasting in the coming decade are likely to include
more robust evaluation of the methods reviewed here, dynamically embedding crop models within cli-
mate models to account for crop influence on regional climate, enhanced use of remote sensing, and
research in the emerging area of ‘weather within climate.’

KEY WORDS: Yield forecasting · General circulation model · GCM · Crop simulation model ·
Stochastic weather generator · Calibration · Probabilistic forecasting
Resale or republication not permitted without written consent of the publisher

1. INTRODUCTION circulation models (GCMs) that model the physical

processes and dynamic interactions of the global cli-
Forecasts of climate fluctuations with a seasonal mate system in response to sea and land surface
(i.e. several months) lead-time are possible because boundary forcing. Probabilistic forecasts are obtained
the atmosphere responds to the more slowly varying from ensembles of GCM integrations initialized with
ocean and land surfaces, an example being climate different atmospheric conditions. Periodic regional
fluctuations associated with the El Niño-Southern climate outlook forums in Africa and Latin America
Oscillation (ENSO) in the tropical Pacific. Several have issued seasonal climate forecasts targeting
climate prediction centers routinely issue prob- agriculture and other climate-sensitive sectors since
abilistic seasonal forecasts based on dynamic general 1997.

*Email: [email protected] © Inter-Research 2006 · www.int-res.com

28 Clim Res 33: 27–41, 2006

By providing advance information early enough to of crop yield response at the farm scale, early enough
adjust critical agricultural decisions, seasonal climate to allow farmers to modify critical pre-planting deci-
prediction appears to offer significant potential to con- sions, might be a fundamental constraint to the use of
tribute to the efficiency of agricultural management, forecasts by risk-averse farmers (Barrett 1998, Blench
and to food and livelihood security. However, there is a 2003). The argument was based on 2 assumptions.
gap between the information that comes routinely (1) Variability of rainfall over small spatial scales
from climate prediction centers and regional climate implies that seasonal rainfall predictability is limited to
outlook forums, and the needs of farmers and other regional spatial scales. (2) Because crop yield is not a
agricultural decision makers. Applications of climate simple function of seasonal total rainfall, the accumula-
forecasts within agriculture are concerned with im- tion of errors going from seasonal climatic predictors
pacts on production and environmental and economic (e.g. SSTs), to local seasonal means, to crop response,
outcomes, and not with climate fluctuations per se. If implies that predictions of effects such as crop re-
farmers are to benefit from seasonal climate forecasts, sponse will be less accurate than predictions of climatic
the information must be presented in terms of produc- means. Recent research challenges these assumptions
tion outcomes at a scale relevant to their decisions, (Hansen 2005). Limited evidence suggests that much
with uncertainties expressed in transparent, proba- of the skill of regional seasonal forecasts holds up at a
bilistic terms. Market and food security early warning local scale (Gong et al. 2003), and that predictability of
applications also need to translate climate information crop yield response can be as great or greater than
into production outcomes, but generally at a different predictability of seasonal climatic means (Cane et al.
spatial scale and lead time. 1994, Hansen et al. 2004).
Stimulated in part by the socioeconomic conse- Our objective here is to survey progress in translat-
quences and widespread public awareness associated ing seasonal climate prediction into forecasts of agri-
with the very strong 1997–1998 El Niño event, interest cultural production that are relevant to agricultural
in agricultural application of seasonal climate predic- decision-making, through the integration of climate
tion gained momentum in the late 1990s. Research models with process-oriented agricultural simulation
efforts have used dynamic, process-oriented crop models. While most applications address crop or forage
simulation models as a means of translating climate yields, relevant applications also include environmen-
forecasts into crop yield prediction and as a basis for tal quality impacts (Mavromatis et al. 2002, Zhang
evaluating potential management responses. This work 2003). We highlight advances over the last decade, as
has depended heavily on analog methods based on well as key challenges and emerging opportunities fac-
categorical indicators of the ENSO (e.g. Jones et al. ing us in the coming decade. Advances in the use of
2000, Meinke & Hochman 2000, Podestá et al. 2002, seasonal climate forecasts with agricultural simulation
Everingham et al. 2003, Meinke & Stone 2005). Tropical models contribute to (1) translating climate forecasts
Pacific sea surface temperatures (SSTs) or the Southern into more relevant information about impacts within
Oscillation Index (SOI) are classified into a small num- the system being managed; (2) ex-ante assessment of
ber of categories or ‘phases.’ Weather data from past benefit to motivate support and insights to target inter-
years, with the same predictor category as the forecast ventions; and (3) guiding management responses
period, are used as input to crop models. The set of through the use of model-based systems that support
simulated outcomes provides a probabilistic forecast. discussion and decision-making (Hansen 2005).
Through the 1990s, this work seldom attempted to in-
corporate operational dynamic climate forecasts, and
borrowed little from the concurrent development of 2. THE CLIMATE–CROP MODEL CONNECTION
methodology for translating climate change scenarios PROBLEM
(often based on the same GCMs used for seasonal fore-
casting) into estimates of agricultural impacts. Despite Operational seasonal climate forecasts are generally
strong interest in using GCM-based seasonal forecasts issued as averages in time (≥ 3 mo) and space. Because
for agricultural applications, progress has — until re- of the effect of spatial and temporal averaging on the
cently — been slow, due in part to limited accessibility of random noise resulting from the chaotic nature of the
GCM results, methodological challenges related to atmosphere, the proportion of variability that is pre-
the spatio-temporal scale mismatch between GCMs dictable at a seasonal lead-time due to boundary forc-
and crop model requirements, and concerns about ing tends to increase with increasing spatial and tem-
characterizing and interpreting forecast probabilities. poral scale up to a point. Furthermore, computational
The 1997–1998 El Niño also stimulated debate about capacity limits the spatial resolution of GCMs used for
the value of climate prediction to a range of societal seasonal prediction to a fairly coarse grid scale, cur-
problems. One concern was that limited predictability rently on the order of 10 000 km2.
 Hansen et al: Translating climate forecasts into agricultural terms 29

Crop production is a function of dynamic, nonlinear gating daily rainfall from 17 stations to a scale typical
interactions between weather, soil water and nutrient of GCMs resulted in the over-prediction of mean simu-
dynamics, management, and the physiology of the lated yields by 28%, due to overestimation of rainfall
crop. Relating predicted climatic variations, averaged efficiency associated with an increased proportion of
in space and time, to crop response is not straightfor- intermediate rainfall events in the aggregated series.
ward. Crop response tends to be nonlinear and some-
times non-monotonic over a realistic range of environ-
mental variability. Furthermore, crops do not respond 3. ADVANCES IN METHODS FOR LINKING
to conditions averaged through the growing season, CLIMATE AND CROP MODELS
but to dynamic interactions between weather, soil
water and nutrient dynamics, and the stage of crop A range of methods for linking crop simulation mod-
development. In rainfed production systems, the els to seasonal climate forecast models have been
interaction between rainfall and the soil water balance advanced. We survey recent advances in methodology
is particularly important. Crop characteristics, soil under 4 categories (Hansen & Indeje 2004): (1) crop
hydraulic and fertility properties, stresses, and man- simulation with daily climate model output, (2) use of
agement mediate sensitivity to weather conditions synthetic daily weather conditioned on climate fore-
within the growing season. Finally, a range of interact- casts, (3) statistical prediction of crop response simu-
ing weather variables mediates many aspects of crop lated with historic weather, and (4) classification and
growth and development. To capture the dynamic, analog methods. The discussion includes some meth-
nonlinear interactions between weather, soil water and ods that have been developed for simulating agricul-
nutrient dynamics, and physiology and phenology of tural impacts of GCM-based climate change scenarios,
the crop, process-oriented crop simulation models but that appear to have potential for yield prediction
typically operate on a daily time step and a spatial based on seasonal forecasts.
scale of a homogeneous plot (although sampling the The applications include both field scales with a
heterogeneity of soil, weather and management inputs focus on farmer decisions, and regional scales that are
allows simulated results to be interpreted at a range relevant to food security early warning and market
of scales). applications. Simulating crop response to weather at
Global and regional dynamic climate models operate aggregate scales has progressed in 2 parallel direc-
on sub-daily time steps, but the spatial averaging that tions. Process-oriented crop models that have been
occurs within grid cells distorts the temporal variability developed for field-scale applications can be scaled up
of daily weather sequences (Osborn & Hulme 1997). by (1) representing heterogeneity of environment and
Any distortion of daily weather variability can seri- management with spatial data sets, (2) probabilistic
ously bias crop model simulations (Semenov & Porter sampling of environmental variables, (3) calibration of
1995, Mearns et al. 1996, Riha et al. 1996, Mavromatis model input parameters or (4) model outputs against
& Jones 1998a, Hansen & Jones 2000, Baron et al. reported crop data at the scale of interest (Hansen &
2005). One of the most serious effects is a tendency to Jones 2000). The alternative is to simulate aggregate
over-predict frequency of wet days and under-predict crop response with models that are simplified to oper-
their mean intensity (Mearns et al. 1990, 1995, Mavro- ate on a large spatial scale while maintaining enough
matis & Jones 1998a, Goddard et al. 2001). The direc- complexity to capture the major components of yield
tion of resulting crop model error cannot be easily responses to climate variability. Examples range
anticipated. On the one hand, when canopy cover is from water-satisfaction indices based on simplified
incomplete and evaporative demand is high, frequent soil water balance (Frere & Popov 1979), to process-
low-intensity showers do not recharge soil water oriented models such as the General Large-Area Model
reserves in deeper layers, but favor increased evapora- (GLAM) designed to simulate annual crop yields at a
tion from the soil surface, thereby increasing water GCM grid scale (Challinor et al. 2004, 2005).
stress (de Wit & van Keulen 1987). On the other hand,
increasing the frequency of rainfall events tends to
reduce the duration of dry periods between rain 3.1. Crop simulation with daily climate model output
events, thereby reducing water stress (Carbone 1993,
Mearns et al. 1996, Riha et al. 1996, Hansen & Jones Despite the tendency of GCMs to seriously distort daily
2000). Baron et al. (2005) suggested that millet in variability, daily GCM output has been used as input to
Sahelian West Africa can use only intermediate (10 to crop models with some success through either the
30 mm d–1) rainfall events efficiently, as smaller rainfall calibration of yields simulated with raw GCM output,
events are largely lost to soil evaporation while more simple rescaling to correct GCM mean bias, or the
intense rainfall is lost to runoff and drainage. Aggre- application of a more sophisticated simultaneous cor-
30 Clim Res 33: 27–41, 2006

rection of GCM rainfall frequency and intensity. of daily GCM rainfall after calibrating frequency. To
Mavromatis & Jones (1998b) used uncorrected daily derive daily rainfall data for a maize simulation model
output from runs of the HadCM2 GCM as input to the at a semiarid location in Kenya, Ines & Hansen (2006)
CERES-Wheat model for studying potential impacts of mapped the cumulative distribution of GCM rainfall
climate change on regional winter wheat production in FGCM,m (x), truncated below the calibrated threshold for
France. Yields simulated with GCM weather data month m, onto the distribution of observed daily rain-
approximated mean yields simulated with observed fall Fobs,m (x), using the transformation,
weather during the past century, and captured a yield –1
x ’i = F obs,m [FGCM,m (x i )] (2)
trend associated with the recent trend in observed
temperature, but under-represented year-to-year vari- for each i th day of GCM rainfall (Fig. 1b). The calibra-
ability. Challinor et al. (2005) used daily meteorologi- tion, using a fitted gamma distribution for observed
cal variables from 9 seasonal hindcast runs from each rainfall intensity, and either a gamma or empirical
of 7 GCMs as input to the GLAM crop model to predict distribution of GCM rainfall intensity, substantially
groundnut yields over western India. Historic district reduced biases of both mean and variance of monthly
groundnut yields aggregated to the GCM grid scale totals, frequency and mean intensity of GCM rainfall.
showed lowest overall prediction error when simulated Baron et al. (2005) demonstrated that disaggregating
yields were calibrated to observed district yields, the spatial averages of daily rainfall from 17 stations
regardless of whether mean bias in the GCM output (approximating the scale of a GCM grid cell) in Sene-
was first corrected. gal to a network of 81 ‘virtual stations’ corrected much
Studies (Mavromatis & Jones 1998a, Hansen & Jones of the bias in rainfall frequency and simulated millet
2000, Baron et al. 2005) have demonstrated the impact yield that resulted from spatial aggregation. They used
on day-to-day variability and crop simulation results, a spatial disaggregation algorithm based on a transfor-
of aggregating daily weather data to a spatial scale mation of a multivariate Gaussian process to a shifted
typical of GCM grid cells and operational seasonal gamma rainfall distribution, designed to generate syn-
forecasts. Several approaches have been proposed to thetic sets of rainfall that match a specified aerial aver-
disaggregate GCM output and other area-averaged age for a given day, but that have statistical properties
daily data sources (e.g. satellite rainfall estimates) to (i.e. frequency, intensity distribution, spatial structure)
the scale of individual stations in a manner that cor- that are consistent with observations at a set of stations
rects the biases. The simplest option for calibrating within the area (Onibon et al. 2004).
daily GCM output to match observed mean local Ines & Hansen (2006) found that using daily GCM
climate is to apply a simple shift (e.g. Ines & Hansen rainfall calibrated to station data at a semiarid location
2006). An additive shift is appropriate for temperature
and solar radiation. A multiplicative adjustment, e.g. 1
a
x ’i = xi,GCM xobs 兾xGCM (1)
is more appropriate for precipitation, as it preserves the F(xobs = 0.0)
sequence of zero values associated with dry days,
where x i,GCM and x’i refer to raw and calibrated GCM
rainfall on day i, respectively, and xobs and xGCM are F(xGCM = 0.0)
long-term mean observed and simulated rainfall, re- 0
0 x0
spectively, for a given time of year. However, because (x0 ⬅ calibrated threshold)
the multiplicative shift corrects total rainfall by adjust-
ing intensity and not frequency, it cannot correct the b
observed tendency of GCMs to over-predict frequency 1
F(xi )
and under-predict mean intensity (see Section 2).
Schmidli et al. (2006) and Ines & Hansen (2006) pro-
F(x)

posed calibrating both the frequency and intensity dis-

tribution of GCM rainfall. The tendency for GCM rain- GCM
fall to be more frequent than observations can be Historical
corrected simply by calibrating a daily GCM rainfall 0
0 xi 0 x ’i
threshold, such that the relative frequency of simu- Daily rainfall (x), mm
lated rainfall above the threshold matches the long-
Fig. 1. Schematic representation of daily GCM bias cor-
term observed frequency for e.g. a given calendar rection: (a) frequency, (b) intensity (Ines & Hansen 2006).
month (Fig. 1a). Schmidli et al. (2006) used a simple F: cumulative distribution function; x: daily rainfall amount;
multiplicative shift to correct the intensity distribution obs: historical observations; GCM: model simulations
 Hansen et al: Translating climate forecasts into agricultural terms 31

in Kenya resulted in systematic under-prediction of Marletto et al. 2005), and by sampling past years in
simulated maize yields, even though the calibration proportion to forecast shifts from climatological tercile
largely corrected mean and variance of GCM monthly probabilities to estimate parameters (Wilks 2002). Con-
rainfall totals, frequency and intensity. They attributed ditioning precipitation parameters on seasonal fore-
the simulated yield bias to a tendency for the GCM casts requires either some assumption about the rela-
rainfall to be more strongly autocorrelated than ob- tive contribution of occurrence and intensity to target
served rainfall, resulting in excessive clustering of rainfall, or empirical estimation relating hindcasts to
rainfall events and unrealistically long dry spells dur- the historic rainfall frequency and intensity record.
ing the growing season. The potential utility of daily Because weather generators are stochastic, many rep-
climate forecast model output for predicting crop licates may be required to approximate target means
response may therefore be limited more by the ability or other statistics of interest with acceptable accuracy.
of GCMs to simulate rainfall with a realistic time struc- An alternative approach is to constrain the gener-
ture than by biased simulation of rainfall frequency ated daily sequences to match target monthly values.
and intensity. However, there is some evidence from A simple additive shift may be sufficient to constrain a
Northeast Brazil that a high-resolution climate model generated series of temperatures to match a target
nested within GCM output fields can simulate daily monthly mean. For rainfall, this is accomplished by
rainfall with more realistic spell lengths than the sampling and testing generated sequences until the
underlying GCM (Sun et al. in press). total is sufficiently close to a target value, then correct-
ing the generated sequence to exactly match the target
(Hansen & Indeje 2004, Kittel et al. 2004, Hansen &
3.2. Synthetic weather conditioned on climate forecasts Ines 2005). This approach requires no a priori assump-
tion about the relative contribution of occurrence and
Seasonal or sub-seasonal (e.g. monthly) climate fore- intensity to target rainfall, but samples synthetic rain-
casts can be disaggregated using a stochastic weather fall sequences that are consistent with the occurrence
model to produce synthetic daily time series that cap- and intensity components of the weather generator,
ture the predictable, low-frequency components of parameterized with historic data. It can, however,
seasonal or sub-seasonal variability, while reproducing accommodate adjustments to parameters of the fre-
important statistical properties of the high-frequency quency and intensity processes. Hansen & Ines
variability in the historic daily record. Two approaches (2005) applied this approach to disaggregate observed
have been advanced. The first, more common ap- monthly precipitation at sites in the Southeast USA,
proach is to adjust the parameters of a stochastic and both observed and hindcast precipitation at a site
generator in a manner that is consistent with the fore- in Kenya, as input to the CERES-Maize model. Con-
cast. The second is to constrain the generated daily straining generated monthly rainfall to match observa-
sequences to exactly match target monthly or seasonal tions largely reproduced the cross-correlation between
means. observed amount, frequency and mean intensity of
The input parameters of simple stochastic weather rainfall more accurately than conditioning weather
generators can be manipulated to reproduce predicted generator parameters on monthly rainfall, and re-
statistical properties of interest, such as means, vari- quired roughly an order of magnitude fewer realiza-
ances, and the relative influence of the number of tions to approach the asymptotic maximum correlation
storms (i.e. frequency) and the type of storm (i.e. the with yields simulated with observed daily rainfall.
intensity distribution) on total rainfall. Several studies
of the behavior of stochastic weather models, moti-
vated largely by climate change impact assessment, 3.3. Statistical prediction of simulated crop response
provide a solid foundation for using weather genera-
tors to produce synthetic daily sequences that are con- The approaches discussed in Sections 3.1 & 3.2
ditioned on seasonal forecasts (Wilks 1992, Katz 1996, involve conditioning crop model weather input data on
Mearns et al. 1997). Methods for conditioning weather the climate forecast. An alternative approach is to treat
generator parameters on seasonal predictions or pre- yields simulated with historic daily weather data as a
dictors include: estimating parameters from years with statistical predictand, and condition the crop model
a particular categorical predictor value (Katz & Par- output on the forecast. By bypassing the need to derive
lange 1993, Grondona et al. 2000, Katz et al. 2003), weather data inputs conditioned on the seasonal fore-
regressing parameters against a seasonal predictor cast, the use of a statistical model trained on crop
(Woolhiser et al. 1993), predicting from GCM output model outputs eliminates one source of error. On the
fields using multivariate statistical downscaling (Can- other hand, this approach is constrained to treating the
telaube & Terres 2005, Feddersen & Andersen 2005, seasonal forecast and its relationship to crop response
32 Clim Res 33: 27–41, 2006

as essentially static within a growing season. While that to our knowledge has not yet been applied to
our review focuses on forecasting using dynamic crop predicting crop yields in response to forecast seasonal
models, statistical prediction from GCM output fields climate variations.
has also been applied to remotely-sensed forage vege-
tation indices (Indeje et al. 2006) and de-trended crop
production statistics (G. Baigorria, pers. comm.). 3.4. Classification and analog methods
Crops tend to show non-linear, non-monotonic rela-
tionships with their environment over some range of Several practical benefits account for the continued
variability, complicating direct statistical prediction. dominance of the historical analog approach for crop
Other potential problems that violate assumptions of yield prediction described in Section 1 (Meinke &
ordinary least-squares regression include residuals that Stone 2005). The approach is easily adapted to any
are non-normally distributed, and residual variance spatial or temporal scale for which historic data are
that varies systematically with predictor. Approaches available. If the predictors used provide any predictive
to dealing with these challenges include nonlinear information about higher-order variations beyond sea-
regression, linear regression following normalizing sonal climatic means that influence crop response,
transformation, generalized linear models, and non- analog years will incorporate that predictability into
parametric models. crop simulations. Distributions derived from analogs
As an example of nonlinear regression, Hansen & will account for any differences in dispersion, in addi-
Indeje (2004) predicted simulated maize yields at a tion to mean shifts associated with different states
site in southern Kenya as a cross-validated function of of ENSO. Finally and perhaps most important, dis-
the first principal component of GCM rainfall over tributions of outcomes simulated for the analog years
the region. They chose a Mitscherlich function, associated with a given category provide an intuitive
means of estimating and communicating forecast
ŷ = a + b (1 – e–cx) (3)
uncertainty in probabilistic terms.
based on its widespread use for modeling plant re- The analog method also has important limitations.
sponse to water and other growth factors. Diagnostics Confidence, artificial forecast skill and biased estima-
showed some evidence that residual variance varied tion of uncertainty are concerns in those cases when
systematically with the predictor— a mild violation of the number of categories and limited record length
the assumptions of least-squares regression. lead to small sample sizes within each category (Sec-
Where the relationship between predicted climate tion 4.3). More important, analogs based on ENSO or
variations and simulated crop response is only weakly other empirical indices do not necessarily capture the
nonlinear, transforming the predictand and potentially best that climate science or operational forecast sys-
the predictor may correct nonlinearity, non-normality tems have to offer. While statistical climate prediction
of regression residuals and heterogeneity of residual models have generally approached their predictive
variance sufficiently to permit ordinary linear regres- limits, dynamic climate forecast models, which inte-
sion. Hansen et al. (2004) used an optimal power series grate global sea and land surface forcing, sometimes
(Box & Cox 1964) transformation to normalize mildly- outperform the best statistical models, and are expected
skewed simulated yield distributions before predicting to improve with improvements in models, data assimi-
district and state wheat yields, simulated with ob- lation, computer capacity and post-processing meth-
served antecedent rainfall and historic within-season ods. Stone et al. (2000) proposed using the analog
rainfall, as a linear function of a regional GCM rainfall approach with GCM output fields classified into dis-
predictor in northeastern Australia. Such data transfor- crete categories by cluster analysis.
mations may not handle non-monotonic crop–climate The analog method described above treats each past
relationships or extreme departures from linearity year falling within the given predictor category as
and normality sufficiently to permit ordinary least- equally probable. It is a special case of a more general
squares linear regression. Because aggregating in set of methods based on classification of predictors or
space smoothes year-to-year variability of crop yields weather types. If there is a basis for predicting that the
and, by the Central Limit Theorem, reduces depar- coming season is more likely to resemble some past
tures from normality, we hypothesize that linear years than others, we can use the predicted probabili-
regression, possibly with a normalizing transforma- ties to derive a probability-weighted forecast distribu-
tion, may be more suited for yield forecasts at an tion, or calculate weighted mean or other distribution
aggregate scale than at a field scale. Generalized statistics. The common method of issuing operational
linear models (McCullagh & Nelder 1989), designed to seasonal climate forecasts as shifted probabilities of
extend the benefits of linear regression where data are each of the climatological terciles, can be used directly
not normally distributed, are a promising alternative to assign weights to analog years or to resample past
 Hansen et al: Translating climate forecasts into agricultural terms 33

years in proportion to the forecast probabilities. For ability of rainfall characteristics (i.e. frequency, distrib-
example, Everingham et al. (2002) and Bezuidenhout ution of dry and wet spells, seasonal total) with encour-
& Singles (2006) sampled analog years in proportion aging skill and realism (V. Moron, pers. comm.). It
to tercile forecasts from the South African Weather appears to be a promising approach to conditioning
Service to forecast sugarcane production. daily weather data inputs on aspects of sub-seasonal
The k-nearest neighbor (KNN) method selects and variability that are predictable at a seasonal lead time,
assigns probability weights to a subset of k past years but it has not yet been tested for crop simulation.
based on their similarity, in predictor state space, to a Non-homogeneous hidden Markov models (NHMM)
given predictor state (Lall & Sharma 1996). Weights wj integrate weather classification with stochastic weather
of the k nearest neighbors, ordered on the basis of their models (Section 3.2) (Hughes & Guttorp 1994, Charles et
similarity to the value of the current predictor vector, al. 1999, Hughes et al. 1999). Observed rainfall patterns
are calculated as: are classified into discrete types. Transition between
states in a NHMM is a Markov process, with transition
wj = ( j Σ ki= 1i –1)–1 (4)
probabilities conditioned on a given set of predictors.
where j and i are indices of the given historic year and The NHMM is parameterized using daily sequences of
the other k nearest neighbor years, sorted by distance spatial weather patterns, and is capable of representing
(i.e. closest = 1) from the current predictor vector. For the historic spatial structure in the weather patterns that
all j > k, wi is set to 0. Using the KNN method to sam- it simulates. The NHMM has been applied to disaggre-
ple past seasons showed comparable results to other gating seasonal rainfall predictions (Robertson et al.
methods that Hansen & Indeje (2004) tested for GCM- 2004, 2006), and to disaggregate rainfall data in space
based maize prediction in Kenya. It has also been used and time as input to a maize simulation model over the
successfully for predicting reservoir inflow from sea- Southeast USA (Robertson et al. in press).
sonal rainfall predictors in Northeast Brazil (DeSouza
& Lall 2003). The KNN analog approach can also be
applied on shorter time steps to probabilistically sam- 4. UNCERTAINTY IN CLIMATE-BASED CROP
ple subsets of past weather observations based on the FORECASTING
degree of similarity of current and historical values of a
given feature vector that may include atmospheric Transparent presentation of uncertainty in proba-
indicators from SST-forced GCM outputs (Clark et al. bilistic terms is crucial to appropriate application of
2004, Gangopadhyay et al. 2005). Appropriate selec- advance information, particularly when risk aversion
tion criteria can preserve moments of the historical dis- influences decisions. Underestimating the uncertainty
tribution, as well as observed spatial and temporal cor- of a forecast can lead to excessive responses that are
relations and correlations among variables. inconsistent with a decision makers’ risk tolerance,
Weather classification works in the same way, except and can damage the credibility of the forecast pro-
that historic data are clustered into discrete circulation vider, while overestimating uncertainty leads to under-
patterns or ‘weather types’ identified e.g. by cluster confidence and lost opportunity to prepare for adverse
analysis, that explain a substantial portion of the vari- conditions and take advantage of favorable conditions.
ability and spatial patterns of rainfall. The ability of a Climate variability and crop model (including input)
GCM, driven by SSTs, to produce daily regional circu- error are the major sources of uncertainty in yield fore-
lation patterns with realistic frequency and seasonality casting. One way to characterize the uncertainty asso-
provides a basis for re-sampling historic local rainfall ciated with climate variability is to simulate yields with
observations based on similarity of circulation patterns antecedent weather observations up to a given forecast
simulated by a GCM and from reanalysis data used date within the season for a current or hindcast year,
as a proxy for observed wind fields (V. Moron, pers. and sample weather data for remainder of season from
comm.). Moron proposed a 2-stage sampling proce- all other years (Fig. 2). The resulting distribution ap-
dure. To capture interannual variability, past seasons proximates the climatic component of uncertainty. In-
are sampled in proportion to their similarity to the cur- formation about antecedent weather and its effect on
rent year based on the distance between principal stored soil moisture provides a degree of predictability
components of GCM and reanalysis wind fields. Daily of yields that increases as the forecast date advances
rainfall is then sampled randomly from the pool of days through the growing season, and an increasing propor-
within the sampled past seasons with the same tion of weather data is observed, rather than sampled.
weather type that the GCM simulates. The process is Several proposed and operational crop-forecasting
repeated for the sequence of daily weather types that systems integrate weather observations through the
the GCM simulates through the season. The approach current date with sampling from climatology for the re-
predicted a substantial portion of the year-to-year vari- mainder of the growing season (Thornton et al. 1997,
 34 Clim Res 33: 27–41, 2006

4.1.1. Forecast distributions from hindcast residuals

initialization

Forecast

Harvest
Model

The first approach is to estimate a forecast distribu-

date
tion as the distribution of hindcast residuals, centered
1 yk,1 on the expected value of the current forecast. To illus-
2 yk,2 trate, Fig. 4a shows a hypothetical 1960–2000 yield
3 yk,3 time series, derived from sampling a multivariate nor-
Weather data year

. . mal distribution, of observations (y) and hindcasts (ŷ )

. .
. . calibrated to observations by linear regression. Sub-
tracting predictions from observations yields a time
series of hindcast residuals
k
. εi = yi – ŷi (5)
.
.
(Fig. 4b), which are then sorted to derive a residual dis-
n yk,n tribution (Fig. 4c). The forecast distribution for the next
year (2001) is obtained by adding its expected value,
ŷ = 1.9 Mg ha–1, to each ε (Fig. 4d). The method
Time of year
accounts for the overall prediction error of the forecast
Fig. 2. Schematic of simulating yields y with antecedent system, and is generally applicable to statistical or
weather data in year k and all other years of within-season
weather data to characterize the climatic component of uncer-
dynamic forecast models.
tainty in yield simulations at a given point within a growing The analog method discussed earlier (Sections 1 & 3.4)
season provides a simple, intuitively appealing way to derive

Samba 1998, Bannayan et al. 2003, Lawless &

Model uncertainty
a
Semenov 2005). Model error represents the
remaining discrepancy between observed Climate uncertainty
Uncertainty

yields and yields simulated with observed

weather.
A skillful seasonal climate forecast reduces
the climatic component of uncertainty. Since
the proportion of total uncertainty that is due
to climate decreases through the growing
season (Fig. 3a), the relative contribution of Planting Anthesis Harvest
seasonal forecasts to overall predictability Growing season
tends to be greatest early in the season, and b
to decrease as the season progresses and an
Uncertainty

increasing proportion of weather is observed,

rather than predicted or sampled (Fig. 3b).
On the other hand, reducing model error
through e.g. improved measurement or cali-
bration of model inputs, updating crop state
variables based on remote sensing (see Sec- Planting Anthesis Harvest Planting Anthesis Harvest
tion 5.2), or modeling additional yield-limit-
ing factors, is likely to have a greater relative
c
Uncertainty

impact on overall uncertainty later in the

growing season (Fig. 3c).

4.1. Deriving crop forecast distributions

Planting Anthesis Harvest Planting Anthesis Harvest
Methods that have been developed for
Growing season Growing season
deriving and evaluating probabilistic cli- Fig. 3. Changing contribution of climatic and model uncertainty (a)
mate forecasts are generally relevant to fore- through the growing season; and with impact of reducing (b) model and
casts of agricultural impacts. (c) climatic uncertainty
 Hansen et al: Translating climate forecasts into agricultural terms 35

Observed Predicted a
4

Yield (Mg ha–1)

3

ŷ2001 = 1.9 Mg ha–1

1

1.5
b
1.0
Residual (Mg ha–1)

0.5

0.0

0.5

–1.0
1960 1970 1980 1990 2000
Year
1.0

c d
0.8
Cumulative probability

0.6

0.4

0.2
2001 forecast
Climatological
0.0
–1.5 –1.0 –0.5 0 0.5 1.0 1.5 0 1 2 3 4
Residual (Mg ha–1) Yield (Mg ha–1)
Fig. 4. Steps in deriving a probabilistic forecast from hindcast residuals: (a) synthetic time series of observations (yi) and corre-
lated, calibrated predictions (ŷi), (b) time series of residuals (εi = ŷi – yi), (c) cumulative density function (CDF) of residuals and
(d) CDF of 2001 forecast and climatology

probabilistic forecasts of climate variations and their agri- 4.1.2. Forecast distributions from dynamic climate
cultural impacts. Yields simulated with weather sampled model ensembles
from the set of past years falling within the category that
corresponds to current conditions are taken as a forecast Initializing GCMs with different sampled atmos-
distribution. It is easy to show that distributions derived pheric conditions improves skill and gives an indica-
from historical analogs are a special case of residual- tion of the uncertainty associated with initial condi-
based distributions, which use the subset of residuals tions (Barnston et al. 2003, Palmer et al. 2004). The use
about the mean from those years that fall within the given of several different GCMs captures uncertainty associ-
predictor class. Hansen et al. (2004) compare cross- ated with model structure and assumptions (Palmer et
validated probabilistic wheat yield forecasts based on al. 2005). The spread of resulting predictions can be
SOI phase analogs and regression from GCM predictors. interpreted as a measure of forecast uncertainty, but
36 Clim Res 33: 27–41, 2006

must be calibrated before forecasts can be expressed cern for the analog method in those cases when the num-
as probability distributions at a local scale (Doblas- ber of categories and limited record length lead to small
Reyes et al. 2005, Palmer et al. 2005). However, there is sample sizes within each category. Robinson & Butler
not yet a consensus about the most appropriate cali- (2002) suggest that many studies that used analogs as a
bration method. basis for prediction may have overestimated prediction
Probabilistic forecasting based on GCM ensembles skill (of climate or impacts) or the potential economic
can be extended to crop yield prediction. For example, value of forecasts for particular decisions.
Challinor et al. (2005) used daily output from each mem- A simple numeric example illustrates the problem. To
ber of both single- and multiple-GCM ensembles to as- mimic an analog-based prediction system, we generated
sess probabilistic forecasts of observed district-level a series of 48 pairs of correlated (r = 0.64) random normal
groundnut yields and crop failure in western India. Can- variates, arbitrarily selected one series as predictors and
telaube & Terres (2005) used monthly climatic means the other as predictands, and grouped sorted predictors
predicted from each ensemble member, statistically into 3, 6, 12, and 16 equally-sized classes — a classifica-
downscaled and disaggregated to daily values, using tion scheme that is analogous to classifying years into
a stochastic weather generator (Feddersen & Andersen El Niño, Neutral and La Niña phases based on SST ob-
2005) to produce probability density estimates of gridded servations in the eastern tropical Pacific (Trenberth
and national wheat yields across Europe. 1997). The proportion of variance in the predictand time
Although crop yield is a continuous quantity, the series that the analogs explained is
probability of yields falling below some threshold — a ––
r 2 = 1 – Σi (yi – ŷi)2兾Σi (yi – Y )2 (6)
discrete event — may be more relevant than the mag-
nitude of yields for applications related, for example, to where yi and ŷi are the ith predictand and prediction,
––
food crisis response or crop insurance. In their study of and Y is the mean of the predictand series. For each
groundnut yield forecasting in western India, Chal- period i, ŷi is taken as the mean of observations falling
linor et al. (2005) used a probabilistic, categorical skill within the period’s predictor class, either including
metric (the relative operating characteristics [ROC] period i, or with cross-validated estimates that omit yi
curve) to demonstrate skillful GCM-based predictions from the calculated mean. Unbiased estimation of fore-
of crop failure. We extend their analysis to compare the cast uncertainty requires that observations from the
predictability of extreme crop failure to more moderate period being predicted do not influence the prediction.
failure. The inverted ROC (IROC) replaces the false Cross-validation reduces this bias (Efron & Gong 1983,
alarm rate (i.e. crop failures that were forecast but not Michaelsen 1987). As expected, the proportion of vari-
observed) used on the x-axis of the ROC curve with a ance that the analogs explain increases as the number
false alarm ratio (i.e. fraction of incorrect failure fore- of categories increases (Fig. 6). However, when the ith
casts), allowing events occurring with different fre- observation is excluded from the ith forecast through
quencies to be compared (Lalaurette unpubl.). For any
crop failure threshold, the nearer the intersection of 1.0
the IROC curve and the no-bias line is to the point Yield < 200 kg ha–1
[0,1], the more skillful the forecast. Based on IROC Yield < 500 kg ha–1
curves (Fig. 5), simulation of crop yield failure is more 0.8
skillful based on a 500 kg ha–1 than a 200 kg ha–1
threshold for both the multi-GCM ensemble and a sin-
0.6
gle-GCM ensemble. Prediction skill at a 3 to 6 mo lead-
Hit rate

time is significant for the higher yield threshold, which

matches the yield below which the cost of groundnut
0.4
cultivation exceeds its value (Rao et al. 2000).

0.2
4.2. Artificial skill and biased probabilities

The error of statistical forecast models tends to be 0

0 0.2 0.4 0.6 0.8 1.0
smaller for the period used to calibrate the model than for
False alarm ratio
predictions outside the calibration data. Artificial fore-
cast skill and systematic underestimation of the disper- Fig. 5. Inverted relative operating characteristics (ROC) curve
for the simulation of crop failure. Solid and dot-dashed
sion of probabilistic forecasts are therefore inherent risks lines — thick: multi-GCM ensemble, thin: single-GCM en-
in statistical forecasting, including empirical calibration semble. Dashed line across graph: simulated crop failure
of dynamic forecast models. They are of particular con- occurrence equals observed climatology
 Hansen et al: Translating climate forecasts into agricultural terms 37

cross-validation, increasing the number of categories influence both the mean (by Kruskal-Wallis 1952 test)
decreases the proportion of variance predicted and and variance (Levene’s 1960 test) of December rainfall
hence the uncertainty that remains. The difference be- in Junin, Argentina (Fig. 7a). After applying a normal-
tween cross-validated and non-cross-validated results izing power series transformation (Box & Cox 1964) to
is an indication of artificial skill and of bias in the dis- reduce the positive skewness of the rainfall series, the
persion of probabilistic forecasts (Michaelsen 1987, mean separation remains, but variances are constant
Meilke et al. 1997, Drosdowsky & Allen 2000), although among ENSO phases (Fig. 7b). In this instance, the
more efficient methods are available for correcting substantial differences in dispersion were an artifact of
prediction error bias (Kohavi 1995, Efron & Tibshirani predicting a skewed distribution, and not an indication
1997). Limiting bias when deriving and evaluating of fundamental shifts in predictability in different
probabilistic climate-based crop forecasts in practice ENSO states. Although rainfall amounts tend to be
requires a combination of selection of credible predic- positively skewed, the existence and direction of skew-
tors with a mechanistic basis, and conservative statisti- ness of rainfed crop yields is difficult to anticipate, due
cal methods such as independent validation and statis- to the generally concave nonlinearity of crop yield
tical hypothesis testing. response to rainfall variability.
Hindcast residuals can account for the effects of
skewness on forecast dispersion, for example by apply-
4.3. Year to year consistency of forecast uncertainty ing a normalizing transformation to the predictand and
potentially the predictor time series, deriving a fore-
Methods for deriving probabilistic forecasts — hind- cast distribution in transformed space, then applying
cast residuals, historical analogs and GCM ensemble an inverse transformation to put the forecast distribu-
distributions — differ in their ability to handle any tions into the original yield units (Hansen et al. 2004).
changes in predictability from year to year. There is GCM ensemble distributions seem to be the best way
some evidence that predictability and hence the dis- to account for any decadal changes in forecast un-
persion of forecast distributions changes over decadal certainty. Resolving the extent to which predictability
time scales, and that GCM ensemble distributions can of climatic variations and crop response change from
capture part of those variations (Grimm et al. 2005, year to year beyond the effect of skewness, and the
Moron 2005). In other cases, analysis of GCM ensem- degree to which GCMs can predict these variations, is
ble simulations show that variability of SST patterns beyond the scope of this review.
influence climatic means but have little influence
on the spread of seasonal mean atmospheric states
(Kumar et al. 2000). 5. EMERGING ISSUES, OPPORTUNITIES, AND
One source of apparent variation in predictability CHALLENGES
between years is the skewness of the underlying distri-
butions. For strongly skewed variables, the magnitude Methods for linking crop and climate models, and
of forecast residuals, and therefore the spread of a field and aggregate scales (see Section 3), and for eval-
forecast distribution, tends to increase in the direction uating probabilistic forecasts (see Section 4) are likely
of skewness. To illustrate, ENSO phases significantly to see significant advances in the coming years. One of

50 Cross-validated
Raw a Transformed b
Skewness: 1.243 Skewness: 0.032
December rainfall

No cross-validation
Variance explained (%)

40

30

20

10
La Niña Neutral El Niño La Niña Neutral El Niño
0 ENSO phase
3 6 12 18
Fig. 7. Box plots and skewness of separation of means and
Number of ‘phases’ variance of December rainfall in Junin, Argentina, by ENSO
Fig. 6. Proportion of variance explained by a hypothetical phase, (a) before and (b) after a normalizing Box-Cox trans-
analog prediction system with and without cross-validation. formation. ENSO influences means in both cases, and dis-
See Section 4.2 for description of ‘phases’ persion in the raw dataset (p < 0.001 in all cases)
38 Clim Res 33: 27–41, 2006

the greatest immediate challenges is the near-absence 5.2. Enhanced use of remote sensing and spatial data
of empirical comparison of the various methods. A
comparison of stochastic disaggregation, nonlinear We anticipate that enhanced use of a range of spatial
regression and weighted analog methods for simulat- data sets from ground observations (e.g. soil surveys,
ing maize yields from GCM output at a single site in crop management) and remote sensing (e.g. rainfall,
Kenya (Hansen & Indeje 2004) was inconclusive. vegetation indices) will contribute substantially both to
Because of peculiarities of the various global and the skill and to the spatial specificity of climate-based
regional climate models; differences in the nature, spa- crop predictions in the coming years. Spatial databases
tial structure and predictability of climate variability are available or under development in many parts of
among locations and seasons; and differences in crop the world for soil properties, land cover and stochastic
sensitivity to within-season weather variability due to weather generator parameters. Satellite remote sens-
soil properties, crop characteristics and management, ing provides a great quantity of spatially explicit infor-
a great deal of empirical testing will be necessary mation about the land surfaces and atmosphere, at
before we can make any robust conclusions about the spatial resolutions that continue to improve with new
suitability of the various approaches. Areas that are sensors.
likely to result in substantial future improvements in Remote sensing has the potential to make several
our ability to predict agricultural impacts of climate contributions to climate-based crop forecasting. (1)
variations at a seasonal lead-time include: (1) expanded Satellite rainfall estimates provide near-real-time in-
evaluation of alternative combined climate–crop fore- formation in locations where rainfall is not directly
casting methods, (2) embedding crop models within measured or rain gauge data are not accessible. Com-
climate models, (3) enhanced use of remote sensing bined with soil and management information, spa-
and spatial data, and (4) new avenues of climate pre- tially contiguous rainfall data offers the potential to
diction research. simulate crop yields anywhere across the landscape
(Thornton et al. 1997, Reed & Maidment 1999).
(2) Remote sensing offers some potential to monitor
5.1. Embedding crop models within climate models cropped areas, planting dates and phenological stages
(Ines & Honda 2005). (3) Remote sensing vegetation
It is increasingly apparent that vegetation can affect indices provide information about the state of the crop
climate (Pitman et al. 1993, Lawrence & Slingo 2004, canopy that can be used to update the state variables
Osborne et al. 2004). At an aggregate scale, agri- of a crop simulation model during the growing
cultural production influences the atmosphere by season, calibrate model input parameters, or statisti-
altering surface roughness, albedo, temperature, and cally correct final yield simulations (Bouman 1992,
moisture flux. Annual crops in particular have quite Delecolle et al. 1992, Moulin et al. 1998). The quality
different seasonal cycles from natural vegetation. The of remote-sensing data is expected to improve in
methods described in Section 3 for translating climate the near future, given plans to launch improved
forecasts into agricultural response account for the sensors for precipitation (Tropical Rainfall Measuring
influence of climate on crops but do not allow any Mission, TRMM; Advanced Microwave Scanning
feedback from the crops to the climate. Dynamically Radiometer for Eos, AMSR-E), soil moisture (AMSR-E;
integrating dynamic climate models with detailed Soil Moisture and Ocean Salinity, SMOS) and vege-
crop and soil simulation models that replace existing tation (Moderate Resolution Imaging Spectroradio-
land surface schemes would allow 2-way feedback meter, MODIS).
between crops and climate within a growing season Seasonal climate forecasts and remote sensing of
(Osborne 2004, Betts 2005). The challenge of match- the state of the crop complement each other. Remote
ing the scale of the climate model grid would have to sensing has the potential to reduce the crop model
be addressed, either by scaling up a field-scale crop component of uncertainty by providing refined esti-
model to account for the heterogeneity within a grid mates of crop state variables up to the time of the
cell (Hansen & Jones 2000), or by using a crop model forecast, while skillful seasonal forecasts reduce
that is optimized for the relatively coarse scale climatic uncertainty from the time of the forecast
(Challinor et al. 2004). The primary benefit of such through the remainder of the season (see Section 4).
2-way coupling would be improved prediction of local Because this application of remote sensing applies
climate in the latter part of the season. Because of only to crop forecasts made during the growing sea-
the biases that dynamic climate models show, we son, integrating remote sensing into climate-based
anticipate that yield predictions from fully coupled crop forecasting may have greater value for food
climate–crop models would require substantial cali- security and market applications than for farm-level
bration. applications.
 Hansen et al: Translating climate forecasts into agricultural terms 39

5.3. New avenues of climate research Barnston AG, Mason SJ, Goddard L, DeWitt DG, Zebiak SE
(2003) Multimodel ensembling in seasonal climate forecast-
ing at IRI. Bull Am Meteorol Soc 84:1783–1796
Climate prediction research has tended to focus on
Baron C, Sultan B, Balme M, Sarr B, Traore S, Lebel T, Janicot
climatic means averaged in time over ≥ 3 mo periods, S, Dingkuhn M (2005) From GCM grid cell to agricultural
and over substantial spatial areas. Although this con- plot: scale issues affecting modelling of climate impact.
vention maximizes prediction skill by reducing non- Philos Trans R Soc B 360:2095–2108
covariant random variability, for agricultural appli- Barrett CB (1998) The value of imperfect ENSO forecast
information: discussion. Am J Agric Econ 80:1109–1112
cations it does so at the expense of relevance. With Betts RA (2005) Integrated approaches to climate-crop mod-
increased attention to forecast applications, particu- eling: needs and challenges. Philos Trans R Soc B 360:
larly in agriculture, and growing awareness of the 2049–2065
tradeoffs between skill and value, climate prediction Bezuidenhout CN, Singels A (2006) Operational forecasting of
South African sugarcane production. I. System description.
research is paying increasing attention to downscaling
Agric Syst (in press), doi:10.1016/j.agsy.2006.02.001
in space and time. Blench R (2003) Forecasts and farmers: exploring the limita-
Seasonal forecasts can, in principle, be calibrated tions. In: O’Brien K, Vogel C (eds) Coping with climate vari-
and evaluated at a local scale, although attempts ability: the use of seasonal climate forecasts in southern
to quantify the effect on prediction skill have so far Africa. Ashgate, Aldershot, p 59–71
Bouman BAM (1992) Linking physical remote sensing models
been few (e.g. Gong et al. 2003). Incorporating under- with crop growth simulation models, applied to sugarbeet.
standing of fine-scale climatic influences — such as Int J Remote Sens 13:2565–2581
orography, land –water interfaces, or land cover — Box GEP, Cox DR (1964) An analysis of transformations. J R
into either statistical downscaling models or high- Stat Soc B 26:211–243
Cane MA, Eshel G, Buckland RW (1994) Forecasting Zim-
resolution, regional dynamic climate modeling is likely
babwean maize yield using eastern equatorial Pacific sea
to further enhance prediction skill at the local scale surface temperature. Nature 16:3059–3071
that is relevant to farm impacts and decisions. Cantelaube P, Terres JM (2005) Seasonal weather forecasts for
Although it is impossible to predict the timing of crop yield modelling in Europe. Tellus 57A:476–487
daily weather events through a season, it is reasonable Carbone GJ (1993) Considerations of meteorological time
series in estimating regional-scale crop yield. J Clim 6:
to assume that the large-scale ocean–atmosphere 1607–1615
interactions that give rise to predicable shifts in sea- Challinor AJ, Wheeler TR, Slingo JM, Craufurd PQ, Grimes
sonal means may also influence higher-order statistics DIF (2004) Design and optimisation of a large-area process-
of synoptic weather events that are important to agri- based model for annual crops. Agric For Meteorol 124:
99–120
culture, such as the frequency and persistence of rain-
Challinor AJ, Slingo JM, Wheeler TR, Doblas-Reyes FJ (2005)
fall events, the distribution of dry spell durations, the Probabilistic hindcasts of crop yield over western India
timing of season onset and the probabilities of intense using the DEMETER seasonal hindcast ensembles. Tellus
rainfall events or temperature extremes. For now, the 57A:498–512
predictability of these higher-order statistics at a sea- Charles SP, Bates BC, Hughes JP (1999) A spatio-temporal
model for downscaling precipitation occurrence and
sonal lead-time remains largely unquantified. We amounts. J Geophys Res 104(D24):31657–31669
anticipate that the emerging focus on what has been Clark MP, Gangopadhyay S, Brandon D, Werner K, Hay L,
coined ‘weather within climate’ will gain momentum; Rajagopalan B, Yates D (2004) A resampling procedure for
lead to improvements in prediction of the higher-order generating conditioned daily weather sequences. Water
Resour Res W04304
weather statistics that determine agricultural impacts,
Delecolle R, Maas SJ, Guerif M, Baret F (1992) Remote sensing
and better characterization of predictability at finer and crop production models: present trends. ISPRS J Photo-
spatial and temporal scales; and perhaps challenge the grammetry Remote Sens 47:145–161
convention of presenting operational forecasts only as DeSouza FA, Lall U (2003) Seasonal to interannual ensemble
seasonal climatic means at an aggregate spatial scale. streamflow forecasts for Ceara, Brazil: applications of a
multivariate, semi-parametric algorithm. Water Resour Res
39:1307–1321
Acknowledgements. This work was supported by grant/coop- de Wit CT, van Keulen H (1987) Modelling production of field
erative agreement number NA67GP0299 from the National crops and its requirements. Geoderma 40:253–265
Oceanic and Atmospheric Administration (NOAA). The views Doblas-Reyes FJ, Hagedorn R, Palmer TN (2005) The rationale
expressed herein are those of the authors, and do not neces- behind the success of multi-model ensembles in seasonal
sarily reflect the views of NOAA or any of its sub-agencies. forecasting. II. Calibration and combination. Tellus 57A:
234–252
Drosdowsky W, Allen R (2000) The potential for improved
LITERATURE CITED statistical seasonal climate forecasts. In: Hammer GL,
Nicholls N, Mitchell C (eds) Applications of seasonal cli-
Bannayan M, Crout NMJ, Hoogenboom G (2003) Application mate forecasting in agricultural and natural ecosystems.
of the CERES-Wheat model for within-season prediction of Kluwer, Dordrecht, p 77–87
winter wheat yield in the United Kingdom. Agron J Efron B, Gong G (1983) A leisurely look at the bootstrap, the
95:114–125 jackknife, and cross-validation. Am Stat 37:36–48
40 Clim Res 33: 27–41, 2006

Efron B, Tibshirani R (1997) Improvements on cross-validation: outputs for crop simulation studies. Agric For Meteorol
the .632+ bootstrap method. J Am Stat Assoc 92(438): 138:44–53
548–560 Jones JW, Hansen JW, Royce FS, Messina CD (2000) Potential
Everingham YL, Muchow RC, Stone RC, Inman-Bamber NG, benefits of climate forecasting to agriculture. Agric Ecosys
Singels A, Bezuidenhout CN (2002) Enhanced risk man- Environ 82:169–184
agement and decision-making capability across the sugar- Katz RW (1996) Use of conditional stochastic models to gen-
cane industry value chain based on seasonal climate fore- erate climate change scenarios. Clim Change 32:237–255
casts. Agric Syst 4459–477 Katz RW, Parlange MB (1993) Effects of an index of atmo-
Everingham YL, Muchow RC, Stone RC, Coomans DH (2003) spheric circulation on stochastic properties of precipitation.
Using Southern Oscillation Index phases to forecast sugar- Water Resour Res 29:2335–2344
cane yields: a case study for northeastern Australia. Int J Katz RW, Parlange MB, Tebaldi C (2003) Stochastic modeling
Climatol 23:1211–1218 of the effects of large-scale circulation on daily weather
Feddersen H, Andersen U (2005) A method for statistical down- in the southeastern US. Clim Change 60:189–216
scaling of seasonal ensemble predictions. Tellus 57A: Kittel TGF, Rosenbloom NA, Royle JA Daly C and 9 others
398–408 (2004) VEMAP Phase 2 Bioclimatic Database. I. Gridded
Frere M, Popov GF (1979) Agrometeorological crop monitoring historical (20th century) climate for modeling ecosystem dy-
and forecasting. FAO Plant Production and Protection namics across the conterminous USA. Clim Res 27:151–170
Paper No. 17, FAO, Rome Kohavi R (1995). A study of cross-validation and bootstrap for
Gangopadhyay S, Clark M, Rajagopalan B (2005) Statistical accuracy estimation and model selection. In: Mellish
downscaling using k-nearest neighbors. Water Resour Res CS (ed) Proc 14th Int Joint Conf Artificial Intelligence.
W02024, doi:10.1029/2004WR003444 Morgan Kaufmann, San Francisco, p 1137–1143. Available
Goddard L, Mason SJ, Zebiak SE, Ropelewski CF, Basher R, at: http://robotics.stanford.edu/~ronnyk
Cane MA (2001) Current approaches to seasonal-to-inter- Kruskal WH, Wallis WA (1952) Use of ranks in one-criterion
annual climate predictions. Int J Climatol 21:1111–1152 variance analysis. J Am Stat Assoc 47:583–621
Gong X, Barnston A, Ward M (2003) The effect of spatial Kumar A, Barnston AG, Peng P, Hoerling MP, Goddard L
aggregation on the skill of seasonal precipitation forecasts. (2000) Changes in the spread of the variability of the
J Clim 16:3059–3071 seasonal mean atmospheric states associated with ENSO.
Grimm AAM, Sahai BAK, Ropelewski CCF (2005) Interdecadal J Clim 13:3139–3151
variability of atmospheric teleconnections and simulation Lall U, Sharma A (1996) A nearest neighbor bootstrap for time
skills of climate models. Geophys Res Abstr 7, 05733; series resampling. Water Resour Res 32:679–693
available at: www.cosis.net/abstracts/EGU05/05733/EGU05- Lawless C, Semenov MA (2005) Assessing lead-time for pre-
J-05733.pdf dicting wheat growth using a crop simulation model. Agric
Grondona MO, Podestá GP, Bidegain M, Marino M, Hordu H For Meteorol 135:302–313
(2000) A stochastic precipitation generator conditioned on Lawrence DM, Slingo JM (2004) An annual cycle of vegetation
ENSO phase: a case study in southeastern South America. in a GCM. Part I. Implementation and impact on evapora-
J Clim 13:2973–2986 tion. Clim Dyn 22:87–105
Hansen JW (2005) Integrating seasonal climate prediction and Levene H (1960) Robust tests for equality of variance. In: Olkin
agricultural models for insights into agricultural practice. I (ed) Contributions to probability and statistics: essays in
Philos Trans R Soc B 360:2037–2047 honor of Harold Hotelling. Stanford University Press, Palo
Hansen JW, Jones JW (2000) Scaling-up crop models for Alto, CA, p 278–292
climate variability applications. Agric Syst 65:43–72 Marletto V, Zinoni F, Criscuolo L, Fontana G and 5 others
Hansen JW, Indeje M (2004) Linking dynamic seasonal climate (2005) Evaluation of downscaled DEMETER multi-model
forecasts with crop simulation for maize yield prediction in ensemble seasonal hindcasts in a northern Italy location by
semi-arid Kenya. Agric For Meteorol 125:143–157 means of a model of wheat growth and soil water balance.
Hansen JW, Ines AMV (2005) Stochastic disaggregation of Tellus 57A:488–497
monthly rainfall data for crop simulation studies. Agric For Mavromatis T, Jones PD (1998a) Comparison of climate change
Meteorol 131:233–246 scenario construction methodologies for impact assessment
Hansen JW, Potgieter A, Tippett M (2004) Using a general studies. Agric For Meteorol 91:51–67
circulation model to forecast regional wheat yields in Mavromatis T, Jones PD (1998b) Evaluation of HadCM2 and
Northeast Australia. Agric For Meteorol 127:77–92 direct use of daily GCM data in impact assessment studies.
Hughes JP, Guttorp P (1994) A class of stochastic models for Clim Change 41:583–614
relating synoptic atmospheric patterns to regional hydro- Mavromatis T, Jagtap S, Jones J (2002) El Niño Southern
logic phenomena. Water Resour Res 30:1535–1546 Oscillation effects on peanut yield and nitrogen leaching.
Hughes JP, Guttorp P, Charles SP (1999) A non-homogeneous Clim Res 22:129–140
hidden Markov model for precipitation occurrence. J R Stat McCullagh P, Nelder JA (1989) Generalized linear models
Soc Ser C Appl Stats 48:15–30 (2nd edn). Chapman & Hall, London
Indeje M, Ward MN, Ogallo LJ, Davies G, Dilley M, Anyamba Mearns LO, Schneider SH, Thompson SL, McDaniel LR (1990)
A (2006) Predictability of the normalized difference vegeta- Analysis of climate variability in general circulation models:
tion index in Kenya and potential applications as an indica- comparison with observations and changes in variability in
tor of rift valley fever outbreaks in the Greater Horn of doubled CO2 experiments. J Geophys Res D95:20469–20490
Africa. J Clim 19:1673–1687 Mearns LO, Giorgi F, McDaniel L, Shields C (1995) Analysis of
Ines AVM, Honda K (2005) On quantifying agricultural and daily variability of precipitation in a nested regional climate
water management practices from low spatial resolu- model: comparison with observations and doubled CO2
tion RS data using genetic algorithms: a numerical study results. Global Planet Change 10:55–78
for mixed-pixel environment. Adv Water Resour 28: Mearns LO, Rosenzweig C, Goldberg R (1996) The effects of
856–870 changes in daily and interannual climatic variability on
Ines AVM, Hansen JW (2006) Bias correction of daily GCM CERES-Wheat: a sensitivity study. Clim Change 32:257–292
 Hansen et al: Translating climate forecasts into agricultural terms 41

Mearns LO, Rosenzweig C, Goldberg R (1997) Mean and vari- Reed SM, Maidment DR (1999) Coordinate transformations for
ance change in climate scenarios: methods, agricultural using NEXRAD data in GIS-based hydrologic modeling.
applications, and measures of uncertainty. Clim Change J Hydrolog Eng 4:174–182
37:367–396 Riha SJ, Wilks DS, Simeons P (1996) Impacts of temperature
Meilke PW Jr, Berry KJ, Landsea CW, Gray WM (1997) A and precipitation variability on crop model predictions.
single-sample estimate of shrinkage in meteorological Clim Change 32:293–311
forecasting. Weather Forecast 12:847–858 Robertson AW, Kirshner S, Smyth P (2004) Downscaling of
Meinke H, Hochman Z (2000) Using seasonal climate forecasts daily rainfall occurrence over Northeast Brazil using a
to manage dryland crops in northern Australia. In: Hammer hidden Markov model. J Clim 17:4407–4424
GL, Nicholls N, Mitchell C (eds) Applications of seasonal Robertson AW, Kirshner S, Smyth P, Charles SP, Bates BC
climate forecasting in agricultural and natural ecosystems. (2006) Subseasonal to-interdecadal variability of the Aus-
Kluwer, Dordrecht, p 149–165 tralian monsoon over north Queensland. Q J R Meteor Soc
Meinke H, Stone RC (2005) Seasonal and inter-annual climate 132(615):519–542
forecasting: the new tool for increasing preparedness to Robertson AW, Ines AVM, Hansen JW (in press) Downscaling
climate variability and change in agricultural planning and of seasonal precipitation for crop simulation. J Appl Meteor
operations. Clim Change 70:221–253 Climatol
Michaelsen J (1987) Cross-validation in statistical climate fore- Robinson JB, Butler DG (2002) An alternative method for
cast models. J Appl Meteorol 26:1589–1600 assessing the value of the Southern Oscillation Index (SOI),
Moron V (2005) Skill of Sahelian rainfall index in two atmo- including case studies of its value for crop management in
spheric General Circulation model ensembles forced by the northern grainbelt of Australia. Aust J Agric Res 53:
prescribed sea surface temperatures. CLIVAR Exchanges 423–428
10(14):19–22 Samba A (1998) Les logiciels Dhc de diagnostic hydrique
Moulin S, Bondeau A, Dellecolle R (1998) Combining agricul- des cultures. Prévision des rendements du mil en zones
tural crop models and satellite observations: from field to soudano-sahéliennes de l’Afrique de l’Ouest. Sécheresse 9:
regional scales. Int J Remote Sens 19:1021–1036 281–288
Onibun H, Lebel T, Afouda A, Guillot G (2004) Gibbs sampling Schmidli J, Frei C, Vidale PL (2006) Downscaling from GCM
for conditional spatial disaggregation of rain fields. Water precipitation: a benchmark for dynamical and statistical
Resour Res W08401 downscaling methods. Int J Clim 26:679–689
Osborn TJ, Hulme M (1997) Development of a relationship Semenov MA, Porter JR (1995) Climatic variability and the
between station and grid-box rainday frequencies for modelling of crop yields. Agric For Meteorol 73:265–283
climate model evaluation. J Clim 10:1885–1908 Stone R, Smith I, McIntosh P (2000) Statistical methods for
Osborne TM (2004). Towards an integrated approach to simu- deriving seasonal climate forecasts from GCMs. In:
lating crop-climate interactions. PhD dissertation, Univer- Hammer GL, Nicholls N, Mitchell C (eds) Applications of
sity of Reading seasonal climate forecasting in agricultural and natural
Osborne TM, Lawrence DM, Slingo JM, Chalinor AJ, Wheeler ecosystems. Kluwer, Dordrecht, p 135–147
TR (2004) Influence of vegetation on the local climate and Sun L, Li H, Ward MN, Moncunill D (in press) Climate varia-
hydrology in the tropics: sensitivity to soil parameters. Clim bility and corn yields in semi-arid Ceará, Brazil. J Appl
Dyn 23:45–61 Meteorol Climatol
Palmer T, Alessandri A, Andersen U, Cantelaube P and 21 Thornton PK, Bowen WT, Ravelo AC, Wilkens PW, Farmer G,
others (2004) Development of a European multi-model Brock J, Brink JE (1997) Estimating millet production for
ensemble system for seasonal to inter-annual prediction. famine early warning: an application of crop simulation
Bull Am Meteorol Soc 85:853–872 modelling using satellite and ground-based data in Burkina
Palmer TN, Doblas-Reyes FJ, Hagedorn R, Weisheimer A Faso. Agric For Meteorol 83:95–112
(2005) Probabilistic prediction of climate using multi-model Trenberth KE (1997) Short-term climate variations: recent
ensembles: from basics to applications. Philos Trans R Soc B accomplishments and issues for future progress. Bull Am
360:1991–1998 Meteorol Soc 78:1081–1096
Pitman AJ, Durbridge TB, Henderson-Sellers A, McGuffe K Wilks DS (1992) Adapting stochastic weather generation algo-
(1993) Assessing climate model sensitivity to prescribed rithms for climate change studies. Clim Change 22:67–84
deforested landscapes. Int J Climatol 13:879–898 Wilks DS (2002) Realizations of daily weather in forecast sea-
Podestá G, Letson D, Messina C, Royce F and 6 others (2002) sonal climate. J Hydrometeorol 3:195–207
Use of ENSO-related climate information in agricultural Woolhiser D, Keefer T, Redmond K (1993) Southern Oscillation
decision making in Argentina: a pilot experience. Agric effects on daily precipitation in the southwestern United
Syst 74:371–392 States. Water Resour Res 29:1287–1296
Rao KN, Gadgil S, Rao PRS, Savithri K (2000) Tailoring strate- Zhang XC (2003) Assessing seasonal climate impact on water
gies to rainfall variability: the choice of the sowing window. resources and crop production using CLIGEN and WEPP
Curr Sci 78:1216–1230 models. Trans Am Soc Agric Eng 46:685–693

Submitted: September 19, 2005; Accepted: September 15, 2006 Proofs received from author(s): November 13, 2006