APPLICATION OF THE WEATHER GENERATOR
FOR CROP GROWTH SIMULATIONS
IN CLIMATE CHANGE IMPACT STUDIES
Martin Dubrovsky
Zdenek Zalud
Institute of Atmospheric Physics
Prague, Czech Republic
Mendel University of
Agriculture and Forestry
Brno, Czech Republic,
zalud@mendelu.cz
http://www.ufa.anet.cz/dub.htm
related contributions in this conference:
 Dubrovsky M., 1999: “Met&Roll: The weather generator for crop growth modelling.” [software demonstration section]
 Zalud, Dubrovsky and Stastna: “Modelling climate change impacts on maize and wheat growth and development” [poster]
web: http://www.ufa.anet.cz/dub.htm#wg (this poster, Met&Roll, publications, ...)
Acknowledgement: The study is sponsored by the Grant Agency of the Czech Republic, contract GACR 205/97/P159.
Introduction
To assess impacts of potential climate change on crop yields, crop growth models are used to simulate
yields in present vs. changed climate conditions [Fig 1 in Zalud et al. (1999) /the adjacent poster; referred
to as ZDS in next/]. Daily weather series required by the crop models may be synthesised by stochastic
weather generator. In a present study, stochastic weather generator Met&Roll [Dubrovsky, 1997;
Dubrovsky, 1999] and crop growth models CERES-Maize and CERES-Wheat were used. This contribution
addresses some methodological aspects, results of the crop simulations are discussed elsewhere [ZDS].
Approaches to simulating impact of climate change on crop yields
In order the findings obtained by comparing model yields in present vs. changed climates have a statistical
significance, it is desirable to perform multi-year crop model simulation for each scenario [Fig.1 in ZDS].
The summary statistics, such as means and standard deviations [e.g., Fig 5] or quantile characteristics
[Fig 4], are then calculated from a set of model yields under each scenario and used for impact analysis.
Following approaches to multi-year crop growth simulations are used here and in ZDS:
A) The crop model simulations with observed pedological, physiological and cultivation data specific for
each individual year. Observed weather series is used for present climate simulations; the weather
series for changed climate simulations is obtained by direct modification of observed series
according to the climate change scenario. [this approach is used in Fig.5-left and Fig. 6]
B) Pedological, physiological and cultivation data are taken from a single “representative” year and
arbitrarily long synthetic weather series is created by the stochastic weather generator. Parameters of
the generator derived from the observed series are used to generate weather series representing
present climate; parameters of the generator are modified in accordance with climate change scenario
to generate series representing changed climate. [used in sensitivity analysis /Fig. 4/ and in poster by ZDS]
This contribution addresses:
a) testing applicability of the weather generator
 Validation of statistical structure of synthetic weather series generated by Met&Roll.
(comparison of the statistical structure of synthetic and observed weather series) [more details in
Dubrovsky (1997)] [Fig. 1, Fig. 2]
 Validation of variability of grain yields simulated by CERES-Maize and CERES-Wheat crop
models. [Dubrovsky et al., 1999] [Fig. 3]
b) analysis of sensitivity of model grain yields to selected characteristics of weather series
(crop model is run with weather series synthesised by generator with modified parameters [Fig. 4])
c) comparison of approaches A and B to assess impacts of climate change on crop yields [Fig. 5]
See ZDS for
 validation of CERES-Maize and CERES-Wheat crop models
 assessment of maize and wheat yields in 1.5CO2 and 2CO2 climates.
Met&Roll: 4-variate stochastic daily weather generator
model:
(i) precip. occurrence ................................ Markov chain (order=1)
[2 parameters]
(ii) precipitation amount (RAIN)................. Gamma distribution
[2 parameters]
(iii) solar radiation
(SRAD)
+ max. temperature (TMAX)
...... AR model (order=1)
parameters: - two 3x3 matrices
+ min. temperature (TMIN)
- 3  (wet/dry)  (avg’s/std's)
 all parameters of the generator [including matrices of the AR(1) model !new!] may vary during a year
 interannual variability of monthly means may be controlled [!new!]
available procedures:
 analysis (= calculating parameters of the generator’s model and other statistical characteristics from
the observed series)
 modification of parameters of the generator’s model according to given climate change scenario
 stochastic generation of synthetic weather series with use of either original or modified WG’s
parameters
 generation of weather series by direct modification of existing series (according to given climate
change scenario)
freely available from web:
http://www.ufa.anet.cz/dub.htm#met&roll
visit the software demonstration on Wednesday morning
Validation of the stochastic structure of synthetic weather series
motivation: stochastic structure of observed and synthetic weather series should be the same
validation tests were focused on:






parameters of the weather generator
distribution of daily precipitation totals
distribution of length of dry and wet periods
normality of variables of AR model (SRAD, TMAX, TMIN)
annual cycle of lag-0 and lag-1 correlations among variables of the autoregressive model [Fig. 1]
variability of monthly and annual means [Fig. 2]
results (of the comparison of synthetic vs. observed series):
 the weather generator underestimates
- frequency of occurrence of long dry spells
- extreme values of daily precipitation amount
- variability of annual and monthly means (Fig. 2)
 daily sums of global solar radiation (SRAD) and daily extreme temperatures (TMAX and TMIN) do
not follow normal distribution assumed by AR model
 correlations and lag-1 correlations among SRAD, TMAX and TMIN exhibit significant annual
cycle not assumed by older version of the generator’s model (Fig. 1)
 on the whole, the model of the generator performs better in summer months
Fig. 1 Annual cycle of lag-0 correlations among
SRAD, TMAX and TMIN. The sample correlation
coefficients for individual weeks were calculated from
the 30-year observed series. The vertical bars at the
right part of the graphs indicate the 95% confidence
intervals of the all-year correlations.
Fig. 2 Reproduction of the variability of monthly and
annual means by Met&Roll. The figure displays the
ratios of observed to synthetic sample standard
deviations of monthly and annual (Y) means of SRAD,
TMAX, TMIN and RAIN. These ratios were averaged
over 17 stations in the Czech Republic.
Note: the new version of Met&Roll allows to consider
annual cycles of the lag-0 and lag-1 correlations
Note: the new version of Met&Roll allows to increase
the interannual variability of monthly means
Validation of variability of model grain yields
Motivation: How the generator's imperfections (in reproducing stochastic structure of daily weather
series) affect model crop yields simulated by CERES-Maize?
Hypothesis:
the lower variability of synthetic series may imply lower variability of model grain yields
Validation experiment:
 input daily weather series (17 Czech stations):
OBS: 30-year observed series
SYN: 30-year synthetic series generated by Met&Roll
 other (pedological, planting, management, ...) input data:
- the same settings for each model year
 testing characteristic: distribution function of model grain yields from the 30-year run
 comparison of distributions obtained with OBS and SYN weather series:
- visually (quantile characteristics of grain yields → Fig. 3)
- Wilcoxon test
Figure 3. Validation of the variability of model maize yields. The minima, 5th smallest values, medians and
maxima of the grain yields were calculated from the 30-year CERES-Maize simulations with use of
observed (lines + rectangles) and synthetic (circles) weather series related to 17 Czech stations.
[note: the input data for crop model simulations slightly differ from those used in ZDS and in Figs. 5-6]
conclusion:
No statistically significant difference was found (validated by Wilcoxon test) and it is thus assumed
that the synthetic weather series generated by Met&Roll are applicable to crop growth simulations
Sensitivity of grain yields to statistical
structure of weather series [approach B]
motivation:
(i) to gain a notion on possible errors resulting from
a) inaccuracy of climate change scenarios
b) ambiguities in projecting climate change scenario into
parameters of weather generator
(ii) to demonstrate possibilities of the weather generator
experiment (for each climate change scenario):
step 1: modification of weather generator’s parameter(s) [according to
given scenario; see the list below]
step 2: generation of 99-year synthetic weather series (Met&Roll)
step 3: simulation of 99-year growth series (CERES-Maize)
step 4: determining quantile characteristics of model grain yields
results are displayed in Fig. 4
list of scenarios used in the sensitivity analysis:
A: modification of means of daily extreme temperatures (daily
temperature amplitude is preserved)
B: modification of daily temperature amplitude (daily temperature
means are preserved)
C: modification of standard deviations of SRAD, TMAX and TMIN
D: modification of interdiurnal variability in AR model
E: modification of mean daily precipitation amount (by modifying
scale parameter of the Gamma distribution)
F: modification of frequency of wet days
G: modification of the shape of distribution of daily precipitation
amount
H: simultaneous modification of frequency of wet days & mean daily
precipitation amount (monthly precipitation sums are preserved)
I: modification of interdiurnal variability of precipitation occurrence
Figure 4. Quantiles of the sets of grain yields obtained in 99-year crop growth
simulations for various scenarios (see the list). The numbers to the right of each
bar are values of the standardised Wilcoxon statistic for testing the hypothesis
that the distribution of grain yields under a given scenario does not differ from the
present-climate distribution (“no change” scenario).
[note: input data for crop model simulations slightly differ from those used in ZDS and in Figs. 5-6]
Comparison of approaches A and B:
The multi-year crop model simulations performed in both approaches (A: "direct modification of weather
series"; B: "weather generator") were made to demonstrate an effect of increased CO2:
Figure 5 Effect of increased CO2 on CERES-Maize yields. The horizontal bars demarcate (avgstd)
calculated from the model yields obtained in a multi-year crop simulation experiments.
Left: 17-year simulations in approach A; right: 99-year simulations in approach B.
Effect of climate change related to increased CO2 is revealed by comparing "1CO2 weather",
"1.5CO2 weather" and "2CO2 weather" crop yields.
Effect of increased CO2 in the air (= direct effect of CO2) is revealed by comparing "c=1CO2",
"c=1.5CO2", "c=2CO2" crop yields.
conclusion: ... the values of the summary statistics of grain yields differ but the trends indicating the
effect of CO2 are about the same.
Model grain yields in multi-year CERES-Maize simulations
under different CO2 scenarios [in approach A]
Figure. 6. Time series of model grain yields simulated by CERES-Maize with the use of a) observed
weather data (red lines) and b) weather data obtained by direct modification of observed weather data
(black lines for 1.5CO2 climate, green lines for 2CO2 climate). Blue symbols represent observed
grain yields. [means and standard deviations from the 17-year series are displayed in Fig. 5 /left plot/]
[see ZDS for the GCM-based climate change scenario and discussion on the effect of elevated CO2]
Conclusions
1. validation of the weather generator: There were found some
discrepancies in ability of the weather generator to reproduce the
statistical structure of the observed daily weather series.
2. validation of variability of grain yields [Fig. 3]: Despite the
above discrepancies, no statistically significant difference
between distributions of grain yields simulated with use of
observed and synthetic weather series was detected (confirmed
by Wilcoxon test) and it is thus assumed that the synthetic
weather series generated by Met&Roll are applicable to crop
growth simulations
3. sensitivity of model grain yield to statistical structure of
daily weather series [Fig. 4]:
- increased variability of daily weather characteristics around their
mean annual cycle and increased interdiurnal persistence of
the weather series decrease grain yields
4. comparison of approaches to multi-year crop growth
simulation under different climate scenarios [A vs B; Fig. 5]:
The results obtained in multi-year simulations with use of directly
modified weather series and stochastically generated weather
series show similar trends in crop yields related to CO2 changes.
Effect of increased CO2 is discussed in a greater detail in a separate poster
contribution [Zalud et al.].
References
Dubrovsky M., 1997: “Creating daily weather series with use of the weather generator.”
Environmetrics 8, 409-424. [http://www.ufa.anet.cz/envir96.htm]
Dubrovsky M., 1999: “Met&Roll: The weather generator for crop growth modelling.” In: this volume.
Dubrovsky M., Zalud Z. and Stastna M., 1999: “Sensitivity of CERES-Maize yields to statistical
structure of daily weather series.” Climatic Change, in press.
Zalud Z, Dubrovsky M. and Stastna M., 1999: “Modelling climate change impacts on maize and wheat
growth and development.” In: this volume.