Statistics Korea has reported estimates of rice yield based on a sampling method (9.15 estimates of rice yield). They collect the number of hills per 1-m2, the number of effective panicles per hill, the number of filled grains per panicle, and any damages on rice plants from 3,359 sampling fields over the country from Sept. 15 to 21, estimate rice yields in kg/10a. However, this sampling method often failed to accurately estimate rice yields at harvesting stages due to the lack of weather conditions during ripening periods. It is widely known that weather conditions such as precipitation, sunshine hours, and temperature during ripening periods are major factors to influence rice yields especially after heading stages. The objectives of this present report are to develop models estimating rice yields and to suggest their applications for stake holders including policy implications using the APCC MME seasonal forecast datasets.

MODIS (Moderate Resolution Imaging Spectroradiometer) Vegetation Indices such as NDVI (Normalized Difference Vegetation Index) and LAI (Leaf Area Index) were collected from Land Processes Distributed Active Archive Center (LP DAAC) for this study. To calculate NDVI at rice paddy fields in Korea, a mask map of rice paddy fields was developed using a landuse map provided by EGIS (Environmental Geographic Information System). There were two types of meteorological datasets which were used for this study to reflect weather conditions during ripening periods: observed meteorological datasets collected from ASOS (Automated Surface Observing System) and APCC (APEC Climate Center) MME (Multi-Model Ensemble) Hindcasts. The 61 sites of ASOS which have longer than 30 years of climate datasets were selected to collect the observed precipitation, sunshine hours, and temperatures. Those datasets from 1983 to 2010 were collected for this study considering the period of APCC MME haindcasts. Even though the 16 GCMs (Global Circulation Models) were used generally used for the APCC MME technique, 6 GCMs including APCC-CCSM3 (APCC, Korea), MSC_CANCM3 (MSC, Meterological Service of Canada, Canada), MSC_CANCM4 (MSC), NASA (National Aeronautics and Space Administration, USA), PNU (Pusan National University, Korea), and POAMA (BOM, Bureau of Meteorology, Australia) were selected to collect the longest hindcast period. Since MODIS VIs from the Aqua satellite were provided from 2002, we had to configure the model set so that we can have the longest period after 2002. A simple composite method (SCM), a simple ensemble method, was applied to these six GCMs. However, the resolution (2.5°×2.5°) MME hindcasts might not be adequate for an agricultural application in this study. A statistical downscaling method using EOFA (Empirical Orthogonal Function Analysis) and SDVA (Singular Value Decomposition Analysis) was applied to provide monthly precipitation, sunshine hours, and temperature at a station scale. The statistical downscaling method estimates station- scale meteorological predictands from large-scale atmospheric predictor variables including SLP (Sea-Level Pressure), T2m (temperature at 2m), T850 (850hPa temperature), u200 (200hPa zonal wind), u850 (850hPa zonal wind), v200 (200hpa meridional wind), v850 (850hpa meridional wind), z500 (500hpa geopotential height) in this study. These meteorological variables in Aug, Sept, Oct, and an average of these three months were used to cover ripening periods of mid-late maturing rice cultivars which are major cultivars in Korea. Rice yields in each provincial level as well as in the national level for the study period were also collected.

These collected datasets including MODIS VIs, landuse, rice yields, observed and estimated meteorological variables were reclassified into agro- climatic zones consisting of 9 climatic zones based on climatic conditions and agricultural cultivation and were used to develop a rice yield estimation model at national and agro-climatic zone levels. The agro-climatic zone used for this study consists of 9 classifications including Apline (ACZ1), Mountainous (ACZ2), Central Northern Inland (ACZ3), Central Northern West Coastal (ACZ4), Central Southern Coastal (ACZ5), Gyeongbuk Inland (ACZ6), Southern Inland (ACZ7), Southern Coastal (ACZ8), and Eastern Coastal (ACZ9). It is widely known that this 9 classifications of agro-climatic zone is more adequate agricultural applications than 19 classifications of agro-climatic zone, since it does not require too much detailed agricultural managements and meteorological variables. Furthermore, since any provinces in this 9 classifications of agro-climatic zone are not divided, rice yields in each agro-climatic zone can be easily aggregated from the those at a provincial level. However, it should be noted that only one ASOS and two ASOSs were located in ACZ1 and ACZ9, respectively. For the rice yield estimation model at a national level, the meteorological datasets from the 61 ASOS, while 57 out of 61 ASOS were used for the yield estimation model at an agro-climatic zone level. Since more than two independent variables including MODIS VIs and meteorological variables were used for the rice yield estimation model, a multiple linear regression approach was used. For this study, we fixed one of MODIS VIs and select other independent variables by the stepwise approach in the SAS software package so that the predictability of the rice yield estimation model using both MODIS VIs and meteorological variables were evaluated against estimates by a model using only MODIS VIs and observed rice yields.

MODIS VIs at rice paddy fields were two resolutions and three day-of-year (DOY): 250-m and 1-km resolutions and DOY 201, DOY 217, and DOY 233. For the national level of MODIS VIs, the correlation coefficients between both NDVI and LAI on DOY 201 and rice yields were about –0.2, those between NDVI and LAI on DOY 217 and rice yields were about 0.3. That between 1-km resolution NDVI on DOY 233 and rice yields was about 0.7 and higher than 250-km resolution NDVI on DOY233 (about 0.6). However, those between LAI on DOYs 217 and 233 and rice yields were approximately 0.3, while that between NDVI on DOY 233 and rice yields was higher than that between NDVI on DOY 217 and rice yields. For the agro-climatic zone level of MODIS VIs, similar patterns and ranges of correlation coefficients were observed for the NDVI. However, higher correlation coefficients between LAI on DOY 217 and rice yields were found in ACZ8 and ACZ9 than those between LAI on DOY 233 and rice yields. These findings observed from the both national and agro-climatic zone levels are in substantial agreement with those of previous research. It should be noted that the rice paddy fields in the rice paddy field mask map (30-m resolution) can be different rice cultivated fields in each year and the mask map does not reflect any changes in landuse over the study period. Therefore, it is suggested that more accurate rice paddy field mask map should be developed for the further study to improve the rice yield estimation model. The PROC MEANS procedure of the SAS software package were used to develop the rice yield estimation models, and the stepwise selection method was used for the selection of the independent variables. The developed rice yield estimation models were Yield=703.35NDVI 1km, 233 -34.82 (R2=0.3) and Yield=582.99NDVI 1km, 233 -0.11P 9 +13.18T max, 9 -269.71 (adjusted R2=0.75) for using NDVI only and using NDVI and the observed meteorological datasets, respectively. (where, yield is the rice yield, NDVI1km, 233 is the 1-km resolution NDVI on DOY 233, P9 is the observed monthly mean precipitation in Sept, and Tmax, 9 is the observed monthly mean maximum temperature in Sept.) With the models, NDVI and monthly mean maximum temperature in Sept. have a positive correlation with rice yields, while precipitation in Sept. has a negative correlation with rice yields. MAPE (Mean Absolute Percentage Error) was used to evaluate the developed models against the observed rice yields. The largest MAPE was approximately 3.53% when the estimated meteorological variables were used. The smallest MAPE (about 1.7%) was found in the rice yield estimation model with both VI and observed meteorological datasets. Interestingly, the errors of the estimated rice yields by the model using VI were relatively high in 2007 to 2009 when the errors of the 9.15 estimates of rice yield were high. The estimates of the rice yield estimation model using the estimated meteorological variables were better than those with VI only and the 9.15 estimates for the three years. Similar approaches were used for the rice yield estimation model at an agro-climatic zone level. However, very low adjusted R-squared value (only 0.1) was found from ACZ7 and there was no consistency in the selection of independent variables over the agro-climatic zones. A further study are suggested on the improvement of the estimation of the rice yields for those agro-climatic zones.

A statistical downscaling method was used to provide the two selected meteorological variables for the rice yield estimation model at a station-scale. EOFA (Empirical Orthogonal Function Analysis) and SDVA (Singular Value Decomposition Analysis) were applied for the meteorological variables. These estimated meteorological variables (precipitation and temperature in Sept. for this case) were used to assess applicability of the APCC MME datasets. The largest MAPE was approximately 3.53% when the estimated meteorological variables were used. The smallest MAPE (about 1.7%) was found in the rice yield estimation model with both VI and observed meteorological datasets. Interestingly, the errors of the estimated rice yields by the model using VI were relatively high in 2007 to 2009 when the errors of the 9.15 estimates of rice yield were high. For the three years, the estimates of the rice yield estimation model using the estimated meteorological variables (MAPE=2.23%) were better than those with VI only (MAPE=4.04%) and the 9.15 estimates (MAPE=4.80%).

Even though the rice yield estimation models should be improved to more accurately estimate rice yields, the suggested rice yield estimation models using MODIS VIs and APCC MME seasonal forecasts can be useful for the followings:

1) Policymakers in South Korea can project the budget for the Direct Payment Program by predicting rice prices at harvesting stages using accurate rice yield estimates. The Direct Payment Program is a government program that is utilized for farmers when rice prices fall below a targeted price. For example, the budget for the Direct Payment Program will need to be increased if rice prices fall significantly below the targeted price. If this increased expenditure for Direct Payment Program is larger than budget that policymakers originally prepared, additional funding for the program should be secured.

2) If rice yields are accurately predicted, it will be possible to reduce market variability through adjusting the timing and quantity of imported rice. A scheme to adjust imports and the domestic availability of rice for the table for shock absorbing in the domestic market is one of policy means that can be considered. If increases in rice production lead to decreases in the market price, it is necessary to delay the harvesting time and shipping time of imported rice. On the other hand, if there is the possibility that prices will rise due to decreases in rice production, it is necessary to advance the shipment time and harvesting time of imported rice.

3) When rice production is expected to be greater than normal, there is a need to prepare for the decline in the market price by isolating the amount required for the market. Currently, there are a few means that can be utilized for government: rice market quarantine, public stockpiling, the adjustment of the distribution amounts using the RPC (Rice Processing Complex), or foreign aid.

It is concluded that by providing additional information on climatic conditions of the ripening stage of rice crops for decision-makers, the proposed rice yield estimation model can be useful to improve the effectiveness of the policies for managing and demand, and price stabilization for the Korean government.

1. 서론

2. 연구 자료 및 방법

2.1 자료수집

2.1.1 관측기상자료

2.1.2 APCC MME Hindcast 자료

2.1.3 벼 수량 자료

2.1.4 MODIS 식생지수 자료

2.2 쌀단수예측모형개발

2.2.1 생산량 전망의 의미

2.2.2 쌀단수예측모형

3. 연구결과 및 토의

3.1 쌀단수예측모형개발

3.1.1 전국 쌀 단수 예측모형

3.1.2 농업기후지대별 쌀단수예측모형

3.2 쌀단수예측모형적용

3.2.1 APCC MME Hindcast 자료

3.2.2 APCC MME 활용 쌀단수추정

4. 정책 함의 및 활용 방안

4.1 모형 활용 방안

4.2 정책적 함의

4.2.1 쌀소득보전직불제 소요 예산에 반영

4.2.2 공공비축제 활용

4.2.3 쌀 수입 및 방출 시기 조절

4.2.4. 수입 관세 활용

4.2.5 RPC를 통한 유통 물량 조절

4.2.6 해외원조 물량 확대

5. 요약 및 결론

5.1 쌀단수예측모형의 개발

5.2 모형 활용 방안

5.3 정책적 함의