Agricultural Systems 77 (2003) 23–38

www.elsevier.com/locate/agsy

Predicting regional rice production in

South Korea using spatial data and
crop-growth modeling
Jin I. Yun*
Department of Ecosystems Engineering/Institute of Life Science and Natural Resources,
Kyung Hee University, Suwon 449-701, South Korea

Received 2 April 2001; received in revised form 31 May 2002; accepted 17 June 2002

Abstract
It is well recognized that aggregation of the inhomogeneous data for soils, management,
and weather within a land unit for crop simulation plays a key role in the scaling-up task of
crop simulation at regional scales. While the other spatial data could be obtained at a desired
level, weather data might not be available in mountainous regions where the production sys-
tem is comprised of many small farms. This study was carried out to develop a scaling-up
scheme of a plot-based crop model to an area-based rice production forecasting system sui-
table for the Korean situation, with an emphasis on the weather data preparation. Daily
meteorological surfaces on a 1
1 km grid spacing were generated by spatial interpolation of
63 point observations from the Korea Meteorological Administration network. Potential
effects of elevation on daily temperature were implemented by applying a seasonal lapse rate
function developed for the terrain. A land cover classification scheme based on a satellite data
analysis was used to identify those pixels corresponding to paddy rice fields, and an empiri-
cally determined land cover effect (cooling) was applied to those pixels to obtain the final
temperature surface over rice paddies. The rice paddy pixels falling within a given cultivation
zone units (CZU) were extracted from the temperature surface to calculate the average
weather representing that CZU. The other weather variables were interpolated by an inverse
distance weighting. By this procedure, the spatially explicit climatic data set for crop simula-
tion (daily maximum/minimum temperature, solar radiation, and precipitation) was prepared
for 1455 CZU’s making up the coterminous South Korea. Soil characteristics and manage-
ment information representing each CZU were available from the Rural Development
Administration of Korea. CERES-rice, a rice crop growth simulation model, was calibrated

* Tel.: +82-31-201-2608; fax; +82-31-204-8116.

‘ E-mail address: [email protected] (J.I. Yun).

0308-521X/03/$ - see front matter # 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0308-521X(02)00084-7
 24 J.I. Yun / Agricultural Systems 77 (2003) 23–38

to have genetic characteristics pertinent to domestic cultivars in Korea. The model was run to
predict rice production of 162 crop-reporting districts for the 3 years (1997–1999) during
which official production statistics are available. The simulation was done with the past
weather data as of September 15 each year, which is about a month earlier than the actual
harvest date. Simulated yields of the 1455 CZUs were aggregated into the corresponding 162
crop reporting districts by acreage-weighted summation, and compared with the reported
yield and production. Forecast yields were less sensitive to the inter-annual climate variation
than the reported yields, and there was a relatively weak correlation between the forecast and
the reported yields. Projected production of rice for each crop- reporting district, which is the
product of yield and acreage, was close to the reported production with the r2 values over 0.97
in all 3 years.
# 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Spatial interpolation; Rice; Crop forecasting; Scaling-up; Crop models; Mountain climate;
Site-specific management

1. Introduction

Rice is a most important food crop in Korea and constitutes a major source of
farmers’ income. Policy-makers as well as farmers are interested in forecasting crop
production across geographic areas before actual harvest. Since 1997, the govern-
ment has been carrying out nationwide surveys on 15 August for crop monitoring
and on 15 September for yield forecasting each year, the latter being approximately
1 month prior to actual harvest date. Yield components data collected from 5000 to
10,000 random plots across the nation are used to calculate the final yield. Plot
yields are aggregated to project the crop production of 162 counties making up the
coterminous South Korean territory. However, this time-consuming and labor-
intensive survey is confronted with obstacles arising mainly from the recently
reduced budget. Under the current budget situation, it would be appropriate to
search for an alternative method.
It is known that more than half of the detrended year-to-year variation in rice
yield can be explained by weather factors (Gommes, 1998). A crop model is a valu-
able tool for evaluating potential effects of environmental, biological and manage-
ment factors on crop growth and development. Unfortunately, most crop-weather
models are applicable to individual plots. If there were some means to broadly apply
crop-weather models to scales that are economically useful and technically feasible,
it would be possible to develop an operational crop forecasting system based on
crop models capable of interpreting weather variability on a regional basis in the
context of crop status and the projected yield. Recent studies in linking crop models
with a Geographical Information System (GIS) have demonstrated a strong feasi-
bility of crop modeling applications at a spatial scale, from the field level for site-
specific management (Han et al., 1995) to the regional level for productivity analysis
(Engel et al., 1997; Thornton et al., 1997).
Besides a large-scale operational crop monitoring system in Europe (Rijks et al.,
‘ 1998), successful applications of crop models to regional scales have been reported
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 25

for millet in Burkina Faso (Thornton et al., 1997), grain sorghum in Australia
(Rosenthal et al., 1998) and wheat in Canada (Chipanshi et al., 1999). Gridded data
at 10–50 km resolution or representative point data are common in such large-scale
crop monitoring systems to scale-up the crop models from plot to regional scales. It
may not be appropriate, however, to apply the same method to Korea, since rice
production systems of this region are comprised of many small farms (mean farm
size less than 2 ha) located in mountainous terrain. Although the highest point does
not exceed 2000 m in this country (e.g. 1950 m for Mt. Halla in Cheju Island, 1915
m for Mt. Chiri in the mainland), the range of elevation found in a 250 m digital
elevation model (DEM) is 253 m in average with the standard deviation of 245 m.
Sloping angles range from 0 to 46.5 degrees with the mean of 7.9 degrees at a 250 m
horizontal resolution (Fig. 1). Besides the complex terrain, spatial variation in soils,
climate, and management practice is far more complicated than those of major
agricultural regions in the world.
It is well recognized that aggregation of the inhomogeneous data for soils, man-
agement, and weather within a single crop reporting district (i.e. a county in the case
of Korea) plays a key role in the scaling-up task. While the other spatial data are
available at a sub-county level in Korea, weather data could not be applied mainly
due to the mountainous terrain feature. Spatial interpolation schemes using GIS
technology have been developed to solve this problem in Korea (Yun et al., 1999, 2000).
However, even after high precision weather data are available, the complicated

Fig. 1. Elevation contour (left) and slope class (right) of South Korea at a 250 m horizontal sampling
‘ resolution.
 26 J.I. Yun / Agricultural Systems 77 (2003) 23–38

distribution pattern of arable land hinders crop model applications at regional scales
in Korea.
The objectives of this study are (1) to develop a scaling-up scheme of a plot-based
crop model to an area-based rice production forecasting system suitable for the
Korean situation, with an emphasis on the weather data preparation, and (2) to test
the applicability of the system by comparison of the simulated yield data with the
actual yield data at the 162 crop reporting districts in Korea for the recent 3 years.

2. Data and methods

2.1. Map data preparation

The smallest land unit where soil and management data are available in Korea is
called ‘Myon’ and coterminous South Korea consists of 1455 such units. Each
county, which is the official district for crop statistics, consists of several Myons.
Each Myon was chosen as a spatial unit for this study, where the crop model will
simulate the same environmental and management condition, and be named ‘culti-
vation zone unit’ (CZU). Areas of CZU range from 7 to 517 km2 with a mean of 65
km2. All CZUs identified in a 1:250,000 analog map were digitized into an ArcView
(ESRI, 1996) shape file of polygon feature.

2.2. Preparation of weather data

Temperature, precipitation and solar radiation are weather variables for most
crop models. Precipitation has little meaning in the lowland irrigated rice cultivation
because rice plants in Korea grow under a flooded condition during most of the
growing season. Solar irradiance within a CZU is not expected to vary significantly,
if measured as a hemispherical total radiation on a horizontal plane. However, local
temperature is controlled by various factors and the spatial variation is expected to
vary substantially even within a CZU. Preparation of daily temperature data is
described in more detail.

2.2.1. Temperature surfaces over mountainous terrain

An observed value of daily maximum/minimum temperature at any point on a
land surface integrates effects of all the control factors including horizontal and
vertical locations, water bodies, inversion, etc. If the land surface were homo-
geneous, the spatial variation of the temperature interpolated by an inverse distance
weighting (IDW) should coincide with the actual temperature. Unfortunately, the
actual land surface consists of diversified covers, and the topography cannot be
represented by the station elevations, frequently making large gaps between the
interpolation and the actual.
Assuming homogeneity of land surface characteristics, temperature at any point
on this land surface may be expressed as the sum of the portion estimated by the
‘ IDW method and the product of a relevant lapse rate multiplied by the elevation
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 27

difference between the real and the ‘virtual’ landscape, the latter being reproduced
by the IDW interpolation of the station elevations.
Since the lapse rate of the atmosphere varies in temporal as well as in spatial
dimension, we should determine the value appropriate for the average time when
daily maximum and minimum temperature occurs over the land surface. This value
is different from the atmospheric lapse rate in that it changes in a horizontal as well
as a vertical direction over the sloping land surface. In this study, the so-called
mountain slope lapse rate (MSLR) was calculated for each day based on a time-
dependent formula. The formula was derived from empirical relationships between
the station elevation and the observed daily maximum/minimum temperature at 63
locations in South Korea for 1999. Although a simple regression between the
observed temperature and the station elevation is common in obtaining regionally
averaged lapse rates, we used a slightly different approach to derive the MSLR.
Since temperature deviation between any two stations at different altitudes consists
of the portions due to geographical variability (latitude, proximity to water bodies,
aspect of the slope, etc.) and weather variability (solar radiation, cloud amount,
etc.), as well as the portion by the elevation difference. Extraction of the ‘signal’
contributed by the elevation effect alone from ‘noise’ caused by other factors was
necessary to delineate meaningful relationships between the temperature and eleva-
tion. In order to remove the noise caused by weather variability, cases with clear sky
(cloud amount less than 0.2) or overcast sky (cloud amount more than 0.8) were
selected from the 22,995 cases (63 stations
365 days). If the number of stations
selected by this classification exceeded 32 on any date, that is, more than half of the
study area is sunny or overcast, then the date was regarded as ‘homogeneous’ with
respect to the sky condition of the study area and the data were subject to the next
step to remove the other noises caused by geographical variability.
Previous studies showed that latitude, longitude, distance from the coastline,
aspect of the slope as well as the elevation are among the major factors determining
the local temperatures in Korea (Shin et al., 1999; Yun and Lee, 2000). We selected a
set of variables best explaining the spatial variation of the minimum temperature on
a given date by a stepwise regression analysis with the 0.05 stay level. Among the
144 days with clear or overcast sky in 1999, we found optimum models with the
coefficient of determination over 0.6 on 77 days. Elevation was included as an inde-
pendent variable in all the models selected. These models explain at least 60% of the
spatial variation in the minimum temperature on a given date by the selected geo-
graphical variables. The partial regression coefficient of each variable in the regres-
sion model represents the amount of temperature change with respect to a step
change in the corresponding variable. Hence, the partial regression coefficients of
the elevation variable in the models are close to the MSLR on the 77 clear or over-
cast dates in 1999.
When these cases were plotted against the corresponding dates, a cyclic pattern
was detected and could be expressed as a cosine function of date. According to the
empirical functions, the absolute value of MSLR (unit:  C/m) is 0.00708+0.003
COS [0.0172 (I120)] for daily maximum temperature, and 0.00695+0.0013 COS
‘ [0.0172 (I30)] for daily minimum temperature, where I is the day of year. Yun et al.
 28 J.I. Yun / Agricultural Systems 77 (2003) 23–38

(2001) applied this method to the 90,000 km2 region in the mountainous Korean
Peninsula to obtain daily temperature at a 1
1 km spacing, and reported the
root mean squared error of less than 1.5  C for both maximum and minimum
temperatures.
Observed daily maximum and minimum temperature data from 63 standard sta-
tions of the Korea Meteorological Administration (KMA) were spatially inter-
polated to produce the temperature surfaces at a 1
1 km grid spacing using the
methods described earlier, starting on 1 January and ending on 31 December each
year, for 7 years from 1993 to 1999.

2.2.2. Zonal aggregation

If we overlay the CZU themes on the meteorological grids and produce an average
of all grid cells within a given CZU, which could be a representative for the CZU.
But the average of the grid cells where rice crops actually grow should be more
appropriate for representing the actual weather, especially in regions with a com-
plicated land use. Land use pattern for production agriculture is strongly affected by
terrain features like elevation, slope, and aspect. For example, rice fields are usually
located in lower-lying areas and apple orchards are common in northward facing
slopes within a CZU. When we calculate the elevation of CZU by zonal averaging,
the mean is 190 m with the standard deviation of 97 m on a 250 m horizontal sam-
pling resolution. Elevation variation within a CZU is even greater, showing the
range of 500 m in average and 1604 m for an extreme case. Average slope of CZU
was 6.4 degrees with the standard deviation of 4.6 degrees. Variation of slopes
within a CZU was 24 degrees in average and 46 degrees for an extreme case. Such a
diversity in terrain feature results in mixed vegetation types rather than a homo-
geneous rice paddy within a CZU. Even within a single grid cell of 1 km2 area, there
could exist a mixture of rice paddy and patches of other vegetations.
Among the grid cells, we identified those with land cover characteristics of rice
paddy based on an analysis of seasonal change in the normalized difference vegeta-
tion index (NDVI). First, we obtained 101 scenes from the NOAA-AVHRR
(advanced very high resolution radiometer) imagery for 1-year period (April 1999–
March 2000) and produced monthly composites of NDVI by using ‘maximum value
composite’ method (Holben, 1986). An iterative self-organizing data analysis scheme
was applied to the monthly composites to classify the land cover characteristics into
five major features at a 1
1 km grid spacing. Pixels classified as ‘forests’ in August
were further analyzed to extract those corresponding to rice paddies. Since rice fields
are flooded with irrigation water before transplanting in Korea, rice paddies are
usually classified as ‘water body’ in May. As crops grow and leaf area indices
increase, rice paddy gets closer to other vegetations in terms of NDVI. We extracted
those pixels classified as ‘water body’ in May from the ‘forests’ pixels in August, and
assigned ‘rice paddy’ feature class to them. When we plot the rice paddy pixels and
compare the scatter with the reported acreage at the CZU level, a good agreement in
spatial distribution pattern is found between them (Fig. 2).
Yoon and Yun (2001) reported a fairly consistent drop in air temperature over
‘ rice paddies compared with the nearby synoptic stations in Korea during rice
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 29

Fig. 2. Comparison of the reported paddy rice acreage by CZU (left) with the pixels classified as rice
paddy by a remote sensing data analysis (right). One pixel corresponds to 100 ha.

growing season. The pixels classified as rice paddy were used to extract the corre-
sponding grid cell values from the meteorological surfaces. The reported cooling
rate (0.6–1.2  C depending on the time of day) was applied to the extracted pixels to
get the final temperature estimates over rice paddies.
By overlaying the polygon CZU coverage on the final temperature grids and cal-
culating the zonal average of the grid cells falling within the corresponding polygon,
we obtained the temperature input data for each CZU.
An inverse distance squared interpolation was used to estimate daily total solar
radiation at each grid cell from 19 solar radiation observation sites of KMA. Pre-
cipitation surfaces were also generated by the same method with the observed data
at 63 KMA stations. Any modification due to land cover change within a CZU was
not considered in the solar radiation and precipitation data preparation. The repre-
sentative value for each CZU was calculated by aggregating all the cell values falling
within the same CZU boundary.

2.3. Calibration of CERES-Rice

The CERES (Crop-Environment Resource Synthesis) model emphasizes the

effects of management and the influence of soil properties on crop performance
(Godwin et al., 1992). Genetic coefficients of the model specifying the phenology of
a given variety are: the thermal units required to complete the juvenile stage, critical
‘ photoperiod, the extent to which phasic development leading to panicle initiation is
 30 J.I. Yun / Agricultural Systems 77 (2003) 23–38

delayed for each hour increase in photoperiod above the critical photoperiod, and
the thermal units for grain fill period. Those specifying growth and yield character-
istics of a given variety are: the number of spikelets per unit dry matter of main
culm, the single grain weight under ideal growing condition, the relative tillering
potential, and the tolerance coefficient for thermal environment.
We determined genetic coefficients of two major Korean cultivars, ‘Odae’ and
‘Palkong’, representing the early- and the mid-to-late maturity class, respectively,
following the method by Hunt et al. (1993). This procedure begins with running the
model with initial values indicated as suitable for a representative Korean cultivar
(Cho and Yun, 1999). Then the model output is compared with actual data, fol-
lowed by successive re-runs with the adjusted parameters. The procedure is iterated
until the simulated and the measured values match. Hence, long-term measurement
data with good quality are the prerequisite for this procedure.
Data derived from long-term field experiments at two national crop experiment
stations (Chorwon and Miryang) were used to carry out this procedure and to verify
the model performance with respect to phenology and yield prediction at a plot scale
(NCES, 1990–1998). Heading date, physiological maturity, grain yield, grain weight,
panicle number, grain number, tops dry weight, and by-product harvest are among
the data collected from the experiments. At the Chorwon site, data were collected
from 1993 to 1997 on the early maturing cultivar Odae. Data were collected at
Miryang from 1989 to 1997 for the mid-to-late maturing cultivar Palkong.

2.4. Crop simulation at CZU level

It was assumed that rice farms in each CZU carry out the same management
practices called ‘recommended management guidelines’ by the Rural Development
Administration (RDA) extension offices. For example, seedlings at 35 days old shall
be transplanted on 20 and 30 May for the early and late maturing cultivars in most
northern parts of Korea, respectively. Planting density of 250,000 plants/ha and 110
kg/ha of nitrogen application were assumed. Irrigation automatically starts at the
0.05 level on a 0 (no stress) to 1 (full stress) water stress scale. The soil is character-
ized by its initial nutrient content, texture, water-holding capacity and field topo-
graphic position in the CERES model. The user determines the number of soil layers
used to run the model. Input files with 5–10 soil layers were prepared for each CZU
based on the soils database of the Rural Development Administration (RDA, 1994).
Daily meteorological surfaces were stored in the Arc/Info Grid format for further
processing (ESRI, 1996). Because the purpose of this study is to compare the per-
formance of the model-based crop forecasting system against the government’s sur-
vey as of 15 September in 3 recent years, observed weather data after this date
should not be used in the simulation for a fair comparison. In operational uses,
forecast weather data should be used to fill in the rest of input file. In this study,
daily weather data from the climatological normals (1961–1990 average; MOST,
1992) were substituted for the 16 September to 31 December portion of the 1997–
1999 data files.
‘
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 31

Soils, weather and management data for each CZU were incorporated into input
files with relevant format to drive the CERES-rice. The model was run to simulate
the growth of two rice cultivars at 1455 CZUs with 7 years of weather data. Among
the output variables, grain yield at physiological maturity was selected for the sub-
sequent analysis.
Because much more diverse cultivars could be grown in a single CZU, the simu-
lated yields of just two cultivars can never be justified as representing the CZU
average. Furthermore, the analysis is complicated because the actual planting
acreages of the early- and the mid- to late-maturing cultivars are unknown. To make
it simple, we assigned either of the higher yield cases from the two simulation results
to be representative of the given CZU.
Since the Korean Ministry of Agriculture and Forestry (MAF) announces the
yield and production of 162 counties at the end of each year, the projected yield and
production should have the same spatial resolution for the purpose of direct com-
parison and validation. Accurate information on the rice planting area is critical to
convert the CZU yield to the county yield. Unfortunately, no acreage information is
updated at CZU level each year. Alternatively, we used the acreage information of
1455 CZUs surveyed at the end of 1995 by the MAF (MAF, 1966). Although the
land use at some CZUs has been rapidly changing in Korea, less than 5% change
(decrease as usual) in the paddy rice acreage is expected in 5 years (MAF, 1996). The
county level yield was obtained through summing the CZU level yields weighted by
the acreage.
Comparison of the absolute value of yield and production may not be adequate
for the system validation purpose because the model not only oversimplifies the real
agroecosystem but also neglects many factors affecting the final crop production, for
example disease and pest outbreaks, lodging by strong winds, and technological
trends. We suggest a relative yield score rather than absolute yield, which can be
expressed as:

Yis ¼ Yi =Ybar
100 ð1Þ

where Yi is the simulated yield of county ‘i’ for any year, Ybar is the spatio-temporal
average of the simulated yields of all the CZUs for the recent 4 years (1993–1996 in
this case). Ysi represents the percent yield of county ‘i’ in any year relative to the past
4-year average yield of the whole country. The same yield score was applied to the
reported yield. In this case, Ybar is the spatio-temporal average of 162 counties for
the 1993–1996 period reported by MAF. Yield scores both reported and simulated
in 1997, 1998, and 1999 were calculated for each county by this formula. Three
years’ data may not be sufficient for a full validation of any methods for crop fore-
casting, but data for validation are limited because the government began the county
level production statistics in 1997.
Total rice crop production (tons harvested in unhulled rice) for each county was
also calculated by multiplication of the simulated yield with the MAF reported
‘ acreage of the county.
 32 J.I. Yun / Agricultural Systems 77 (2003) 23–38

Fig. 3. Measured and predicted heading date (top) and the grain fill period (bottom) of Korean rice
varieties by CERES-rice.

3. Results and discussion

3.1. Model performance at a plot scale

When the duration from transplanting to heading was estimated and compared
with observed data at the two experimental plots, there was a good agreement
between them (Fig. 3a). Summer weather was much cooler than normal in 1993,
while much warmer than normal in 1994. The model predicted the variation in
heading date as affected by unusual temperature anomalies in two extreme years.
Although there was a substantial discrepancy in a few years, overall trend seems to
‘ reflect year-to-year variation in weather pattern during the test period.
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 33

Fig. 4. Measured and predicted grain yield of Korean rice varieties by CERES-rice.

Though the coefficient of determination (r2) was lower than the heading date case,
duration for grain filling (from heading to physiological maturity) was predicted
fairly well by the model (Fig. 3b). But the simulated grain fill period of the early
maturing cultivar, Odae, varied from 38 days in 1994 with warm summer to 52 days
in 1993 with cool summer, while the observed period varied little from 46 to 48 days.
In the mid- to late-maturing cultivar, Palkong, the model predicted approximately
two weeks of variation in the duration of grain fill during the 9-year period, while
the observed variation was only 1 week. Phenological development response of the
model seems to be more sensitive to thermal environment during the growing season
than the actual crop response.
The model predicted higher grain yield than the observed yield in poor-yield years,
but underestimated in good-yield years (Fig. 4). Slope of the fitted line is far less
than 1.0, indicating that the simulated yield response is less sensitive than the actual.

3.2. Yield and production forecasting at county level

The simulated grain yield of 128 crop reporting districts with the rice crop acreage
greater than 1000 ha was plotted against the reported yield for the recent 3 years
(Fig. 5). The widely scattered pattern and the low correlation coefficients imply that
the linkage between simulated and reported yield weakens as the model application
level moves from a plot scale to a regional scale such as CZU or county. Hence, the
spatial variation in a single year also shows less similarity between the projected and
the reported yields (top panels in Fig. 6).
When the productivity of each county is expressed as ‘relative yield score’, the
simulated results show lower productivity than the reported values in general (bot-
tom panels in Fig. 6). The reported rice yields in the recent 3 years were consistently
higher than the average yield of the past 4 years. When we select 37 counties with
‘ more than 10,000 ha acreage, all the counties were reported to have better yield than
 34 J.I. Yun / Agricultural Systems 77 (2003) 23–38

Fig. 5. Comparison of the projected grain yield of September 15 with the MAF reported final harvests for
128 counties in the recent 3 years.

the past 4 years’ average in 1997 and 1999 (MAF, 2000). However, the model pre-
dicted less yields than normal for 11 and four counties out of the 37 counties in 1997
and 1999, respectively. Relative yield scores predicted by the model are consistently
lower than the reported yield scores in all 3 years.
There should be many known and unknown factors contributing to this somewhat
disappointing result. Individual plots are more or less homogeneous in the context
of weather, soils, cultivar, and management practices. When we apply the model to
simulating the crop growth, development and yield, the output is expected to have a
quasi-deterministic nature. But, if the same model is applied to a CZU, which is
comprised of several hundred farms having different soils, weather, and manage-
ment practices, it should be run by ‘representative’ or ‘spatial average’ input data for
these variables, which cannot be free from inherent errors. Though we employed a
scaling-up technique for weather input data, overall performance of the yield fore-
casting system seems far less sufficient. Spatial aggregation schemes for other input
variables should also be improved to get a reliable yield forecast at county scale.
A feasible but partial solution could be found from the forecast weather avail-
ability. Crop simulation ends on 15 September, when there remains about one more
month for actual rice plants to complete grain fill. Since the simulated crops grow
under the weather condition of the climatological normal year after 15 September in
this operation, they do not respond to any effects from weather anomalies during
the grain fill period which could be critical in some years to determination of the
final yield. To test this reasoning, we replaced the climatic normals with the ‘Perfect
forecasts’, i.e. the observed weather data after 15 September in the 3 years, and ran
the model again. There was a substantial improvement in the correlation coefficient
(from 0.47 to 0.57) between the simulated and the reported yield scores in 1997, but
‘ little improvement was found in 1998 and 1999. If long-range forecasts accurate
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 35

Fig. 6. Spatial variation of the grain yield projected by aggregating the CZU yields (top left) and the
reported yields (top right) in 1997. Bottom two panels show the variation of relative yield scores calcu-
lated by the simulated (left) and the reported (right) yield statistics in the same year. Rectangles on the top
left panel indicate the location of 63 weather stations.

enough to substitute the normal weather were available, crop forecasting could be
improved in some abnormal years.
Rainfall data have little influence on growth simulation of paddy rice in South
Korea and the carefully prepared site-specific temperature data seem to have accu-
racy relevant for crop-growth simulation. But solar radiation data were interpolated
from only 19 points while temperature and rainfall data were from 63 points. Hence
inaccuracy in the interpolated solar radiation data might partially cause the poor
model fit.
Another explanation for the regional yield estimation error may be the validation
data set itself. Since the nationwide crop status survey employs a random sampling
and statistical aggregation to obtain the yield estimates, the validation data set
derived from the survey results is only an approximate measure of true regional
‘ yields.
 36 J.I. Yun / Agricultural Systems 77 (2003) 23–38

Fig. 7. Comparison of the projected grain production as of 15 September with the MAF reported final
harvests for 162 counties in the recent 3 years.

Total unhulled grains harvestable from 162 counties were calculated by multiplying
the absolute value of the simulated county mean yield with the MAF reported acreage,
and plotted against the MAF statistics (Fig. 7). There was a good agreement (r2 values
higher than 0.97) between the simulated and the reported production in all 3 years. This
apparent improvement in the model performance, however, comes from the contribu-
tion of the acreage factor. Even though the simulated yield is far different from the
reported, overall range of the yield variation is much smaller than the acreage variation
among the counties. When the products (yield
acreage) are calculated and plotted,
discrepancy in the yields becomes negligible on the 1:1 line of the whole county range.
Nevertheless, it is promising that the trend lines show a close similarity among the
3 years with different weather conditions. It implies that the crop model response to
the inter-annual climate variation is adequate, at least in the context of the county
scale total production, even though the yield response to year-to-year weather var-
iation becomes less sensitive as the spatial scale goes from plot to county level.

4. Conclusions

We suggested a way to link a crop model with spatially explicit weather data to
project crop production at a sub-county scale in mountainous regions like Korea.
Emphasis was put on preparation of the site-specific weather data for the model
input. A seasonal variation of the lapse rate was adopted for daily temperature
‘ interpolation, which has practical implications in mountainous regions. Land cover
 J.I. Yun / Agricultural Systems 77 (2003) 23–38 37

effects were also taken into account by applying the cooling rate to the rice paddies.
Satellite remote sensing data were used to delineate the exact locations of paddy rice
cultivation, and helped to get more relevant spatial average data.
When the rice production of each county was projected by the system about a
month earlier than the actual harvest date, there was reasonable agreement
(r2 > 0.97) with the reported production. Results imply that the simulated produc-
tion explained both the inter-annual and the geographical variation in the reported
production fairly well. We conclude that the system showed a feasibility of opera-
tional crop forecasting.
Relatively poor performance of the system in the context of yield forecasting can
be explained by several factors. Inaccuracy in the interpolated solar radiation data
could be a partial cause for the poor model fit. The validation data set itself could be
another candidate for accounting for the poor model performance. In addition, the
assumption of constant management practices across the country is apparently a
source of error. Operationally, the simulated yield has no contribution from the
weather effects after 15 September. If we find a way to utilize long-range weather
forecasts instead of climatic normals, more reliable yield prediction could be possi-
ble. Any losses by disease and pests as well as severe weather events were not con-
sidered in the simulation. Plant disease and pest modules may be added to the crop
model to solve this problem.
It appears desirable to maintain all the necessary spatial information at the individual
farm or plot level, and aggregate the outputs from each simulation to regional scales
such as a CZU or a county. Some GIS technology combined with a high speed
computation facility should be helpful for this end. Besides the creation of spatial
weather data necessary to run the model from point observations as is shown in this
study, GIS can be used to store and manage a large map database including soils
and management information.

Acknowledgements

This work was supported by Grant No. 1999-2-221-002-5 from the Inter-
disciplinary Research Program of the Korea Science and Engineering Foundation
(KOSEF).

References

Chipanshi, A.C., Ripley, E.A., Lawford, R.G., 1999. Large-scale simulation of wheat yields in a semi-arid
environment using a crop-growth model. Agricultural Systems 59, 57–66.
Cho, K.S., Yun, J.I., 1999. Regional crop evaluation and yield forecast of paddy rice based on daily
weather observation. Korean Journal of Agricultural and Forest Meteorology 1 (1), 12–19. (in Korean
with English abstract).
Engel, T., Hoogenboom, G., Jones, J.W., Wilkens, P.W., 1997. AEGIS/WIN—a computer program for
the application of crop simulation models across geographic areas. Agronomy Journal 89, 919–928.
ESRI, 1996. Using ArcView GIS. Environmental Systems Research Institute, Inc, Redlands, CA.
‘
 38 J.I. Yun / Agricultural Systems 77 (2003) 23–38

Godwin, D., Singh, U., Ritchie, J.T., Alocilja, E.C., 1992. A Users Guide to CERES-rice. International
Fertilizer Development Center, Muscle Shoals, AL, USA.
Gommes, R., 1998. Crop-yield Weather Modelling. Lecture Notes and Exercises for WMO/FAO Roving
Seminar.
Han, S., Evans, R.G., Hodges, T., Rawlins, S.L., 1995. Linking a geographical information system with a
potato simulation model for site-specific crop management. Journal of Environmental Quality 24, 772–777.
Holben, B.N., 1986. Characteristics of maximum-value composite images from temporal AVHRR data.
International Journal of Remote Sensing 7, 1417–1434.
Hunt, L.A., Pararajasingham, S., Jones, J.W., Hoogenboom, G., Imamura, D.T., Ogoshi, R.M., 1993.
GENCALC: software to facilitate the use of crop models for analyzing field experiments. Agronomy
Journal 85, 1090–1094.
MAF, 1996. General Survey on Agricultural Census 1995 (CD-ROM). Bureau of Agricultural Statistical
Information, Ministry of Agriculture and Forestry, Korea.
MAF, 2000. Trend in Crop Production. Bureau of Agricultural Statistical Information, Ministry of
Agriculture and Forestry, Korea.
MOST, 1992. Studies on Digital Maps for Climatological Normals in South Korea. Meteorological
Research Institute. Final Report to the Ministry of Science and Technology, Korea. (in Korean with
English abstract).
NCES, 1998. Annual Report (Paddy Rice). National Crop Experiment Station, Korea.
RDA, 1994. Development of Computer Model and Agricultural Information System for Classifying Farm
Area. Final Report of Rural Development Administration Special Project. (in Korean with English
abstract).
Rijks, D., Terres, J.M., Vossen, P. (Eds.), 1998. Agrometeorological applications for regional crop mon-
itoring and production assessment. Official Publication of the EU, EUR 17735, Luxembourg.
Rosenthal, W.D., Hammer, G.L., Butler, D., 1998. Predicting regional grain sorghum production in
Australia using spatial data and crop simulation modeling. Agricultural and Forest Meteorology 91,
263–274.
Shin, M.Y., Yun, J.I., Suh, A.S., 1999. Estimation of daily maximum and minimum temperature dis-
tribution over the Korean Peninsula by using spatial statistical technique. Journal of Korean Society of
Remote Sensing 15, l9–20. (in Korean with English abstract).
Thornton, P.K., Bowen, W.T., Ravelo, A.C., Wilkens, P.W., Farmer, G., Brock, J., Brink, J.E., 1997.
Estimating millet production for famine early warning: an application of crop simulation modeling
using satellite and ground-based data in Burkina Faso. Agricultural and Forest Meteorology 83, 95–
112.
Yoon, Y.K., Yun, J.I., 2001. Using synoptic data to predict air temperature within rice canopies across
geographic areas. Korean Journal of Agricultural and Forest Meteorology 3, 198–202. (in Korean with
English abstract).
Yun, J.I., Lee, K.H., 2000. Agroclimatology of North Korea for paddy rice cultivation: preliminary
results from a simulation experiment. Korean Journal of Agricultural and Forest Meteorology 2, 47–
61. (in Korean with English abstract).
Yun, J.I., Lee, D.S., Choi, J.Y., Cho, S.I., Park, E.W., Hwang, H., 1999. Elevation-corrected spatial
interpolation for near-real time generation of meteorological surfaces from point observations.
AgroInformatics Journal 1 (1), 28–33. (in Korean with English abstract).
Yun, J.I., Choi, J.Y., Yoon, Y.K., Chung, U., 2000. A spatial interpolation model for daily minimum
temperature over mountainous regions. Korean Journal of Agricultural and Forest Meteorology 2 (4),
175–182. (in Korean with English abstract).
Yun, J.I., Choi, J.Y., Ahn, J.H., 2001. Seasonal trend of elevation effect on daily air temperature in
Korea. Korean Journal of Agricultural and Forest Meteorology 3 (2), 96–104. (in Korean with English
abstract).