Due to global warming, bud-burst and flowering dates of fruit crops are occurring earlier in recent years and, as the abnormal climate increases the variabilities of temperature in spring, the risk of frost damage has increased. This abnormal climate contributes to the increase in the variability of fruit prices, creating a potentially damaging situation for farmers and consumers. Despite the broad impacts of this phenomenon, the full blooming date prediction model for peach trees, employed by the Rural Developmental Administration (RDA), was developed using only one cultivar (Youmyeong) and observations from a single station (Suwon). This model might not adequately reflect the characteristics of other peach cultivars or local orchards. Therefore, this report develops site- and cultivar-specific blooming date prediction models for major peach cultivation regions and cultivars, presents a framework for applications of the APEC Climate Center Multimodel Ensemble (APCC MME) seasonal datasets, develops frost occurrence prediction models using logistic regression and decision tree techniques, and presents applications of this study. In addition, the prediction models of frost occurrence in spring were developed to apply the "Ready-Set-Go" framework that different strategies for temporal scales from seasonal to short-range forecasts can be more effective.

Multilinear regression, developmental rate (DVR), and sequential dormancy models (chill day, new chill day, and fraction-time models) were used to develop the locally-tailored full blooming date prediction models for major peach cultivars. For the development of these models, bud-burst and full blooming dates of peach tree for 5 cultivars (Cheonhong, Youmyeong, Changbangjosaeng, Cheonjoongdo, and Janghowon) were collected from the 7 major peach cultivation sites: Chuncheon, Suwon, Yesan, Cheongwon, Cheongdo, Naju, and Jinju. Because the observed period of those dates in Yesan is too short to develop those models, those dates in Yesan were not used for this study. Although the weather datasets including daily maximum, minimum, and mean temperature were required to develop those prediction models, the orchards where those dates were collected from did not have weather stations. Therefore, these weather variables were collected from the nearest Automated Synoptic Observing System (ASOS) from the orchards (Chuncheon:101, Suwon:119, Cheongju:131, Daegu:143, Gwangju:156, and Jinju:192). APCC　MME seasonal datasets were used for the predicted temperature in April, which in turn was used to predict the yearly full blooming dates of peach tree at the end of March. The frost occurrence days and 8 meteorological variables (minimum temperature, grass minimum temperature, dewpoint, wind speed on frost occurrence days, mean relative humidity, minimum relative humidity, cloud amount on one day before the frost occurrence days, and difference between maximum temperature on one day before the frost occurrence days and minimum temperature on the frost occurrence days) were obtained from the Korea Meteorological Administration (KMA). These datasets were collected from 8 total stations including the 7 major observed sites for the full blooming dates of peach tree and Wonju (114), one of regions in which RDA predicts the full blooming dates of peach trees.

For the multilinear regression model, mean temperature in March and mean temperature in early, mid, and late April were used for the explanatory variables to the Julian days on the full blooming dates. For this model, observed full blooming dates for the two cultivars (Youmyeong and Changbangjosaeng) in the Suwon site were used since approximately 30 of full blooming dates have been collected since 1979. Both linear and exponential models for the DVR model were used and compared for the 6 major sites and 5 cultivars. It is assumed that only days when the daily mean temperature is above 5℃ contribute to the development of peach trees, as the National Academy of Agricultural Science (NAAS) proposed. Once the DVRs for each day were calculated, the DVRs were summed until those sum reached 1.0 to determine the full blooming dates. Three sequential dormancy models including chill day, new chill day, and fraction-time models were applied for the 6 major observed sites and the 5 cultivars. Notably, to the best of our knowledge, this is the first application of new chill day and fraction-time models to fruit tree crops in Korea. For these models, the threshold temperature (Tc) and chilling requirement (Cr) were estimated with an iteration method. Based on the literature review, we defined the ranges of Tc and Cr (5≤TC ≤10, -70≤Cr ≤-150). When ≥  rest is assumed to be broken and when ≥ , quiescence is assumed to be overcome (where Cd is chilling accumulation and Ca is heating accumulation). The TC and CR were determined where the Root Mean Square Error (RMSE) for the estimates against observed bud-burst dates was at a minimum. With those Tc and Cr, heating requirements (Hr) for each observed full blooming dates were calculated and averaged. The periods when both bud-burst and full blooming dates were observed were used for the training periods for the models. The periods with only full blooming dates were used to validate the models.

Among the full blooming dates of peach trees developed in this study, the model that predicted full blooming dates of peach trees most accurately (i.e., measures of goodness-of-fit were best) was selected for the applications of the APCC MME seasonal datasets for the predicted temperature in April were applied. However, the APCC MME seasonal datasets are too coarse (about 250 km of resolution) for the 6 major sites for the full blooming dates. Due to the constraints of the APCC MME seasonal datasets, we predicted required temperature variables (e.g., maximum and minimum temperature) using a statistical downscaling method for the 6 stations nearest from the 6 major sites for full blooming dates. However, the number of the observations of the full blooming dates and the accuracy of the predicted temperature by the statistical downscaling method were considered selecting a site and cultivars. For the prediction models of the full blooming dates of peach trees, daily temperature datasets are required. Unfortunately, the APCC MME seasonal datasets currently do not provide daily temperature. We proposed a framework adapted from the bottom-up procedure used by the Euro-Mediterranean Center on Climate Change (CMCC) in their operational activities for watershed management. The Maximum Entropy Bootstrap Weather Generator (MEBWG) was used for this framework. The synthetic daily temperature generated by MEBWG was aggregated at a monthly temporal scale. These aggregated synthetic temperatures were compared with the downscaled monthly temperature at the site, and those similar to the predicted monthly temperatures were selected.

Logistic regression and decision tree techniques were used to develop the prediction models for the frost occurrence events. Based on the literature review, a total of 8 meteorological variables were selected for the explanatory variables. The results predicted from these developed models were summarized in the 2×2 contingency table. With this table, we calculated Hit Rate (HR), Probability of Detection (POD), and False Alarm Rate (FAR) as the skill scores for the predictability of the models. We selected threshold values to maximize HR and POD and minimize FAR for each station. Because the smaller number of the selected explanatory variables are selected, the operational applicability of the frost occurrence prediction models can be the more useful. We compared the developed models using both logistic regression and decision tree techniques, and proposed a better technique for the operational use. This proposed technique may be useful to better support farmers by providing adequate strategies to reduce frost damages through a timely warning.

The multilinear regression model for the full blooming date prediction was not appropriate for this study as the fitted explanatory variables were not significant at the 0.05 probability level. We concluded that this model is not useful for the prediction model to reflect the local and cultivar characteristics. The required days for full blooming were fitted into both linear and exponential models and compared. No significance at the 0.05 probability level was found in the linear model for most of the sites, while the exponential model showed significance at the 0.05 probability level. Fitted coefficients were varied from locations and cultivars. This result implies that the prediction model for the full blooming dates of peach trees should reflect the local and cultivar characteristics to better predict the full blooming dates. Three measures of goodness-of-fit, Mean Absolute Percentage Error (MAPE), coefficient of determination (R2), and Root Mean Square Error (RMSE) were used to evaluate the model. The highest MAPE (5.2%) was found from the Cheonhong cultivar in the Cheongwon site and the highest R2 was 0.96 for Cheonhong cultiar in the Naju. The RMSE values varied with the range of 0.91 and 6.31 days. The overall measures for the dataset regardless of the location and cultivar were 2.58%, 0.75, and 3.65 day for MAPE, R2, and RMSE respectively.

For the sequential dormancy model, we evaluated the chill day, new chill day, and fraction-time models. TC and CR were successfully estimated through the iteration method, while the fraction-time models failed to estimate TC and CR. The ranges of TC were 5 to 9℃ and 5 to 7℃ for chillday and new chill day models respectively. The estimated CR values ranges were -148 to –71 and -140 to -75 for the chill day and new chill day models respectively. The ranges of HR were 134.0 to 277.4 and 153.4 to 272.1 for the chill day and new chill day models. These results showed that the ranges for the new chill day model were smaller than those of the chill day model. The highest RMSE values for the bud-burst dates were 8.67 and 7.12 days for the new chill day and chill day models respectively, while those for the full blooming dates were 3.89 and 3.96 days for the new chill day and chill day models respectively. Like the results of the DVR model, the goodness-of-fit measures were different from the location and cultivars. For the chill day model, the measures for the entire dataset regardless the location and cultivar were 2.31%, 0.79, and 3.36 days for MAPE, R2 , and RMSE respectively. For the new chill day model, the values were slightly better than those of the chill day model at 2.19%, 0.82, and 3.16 days for MAPE, R2, and RMSE respectively.

The model results showed that the new chill day model displayed a higher performance than the others. Based on the considerations for the predictability of the statistical downscaling method and the observed periods of the full blooming dates at each site, we determined that the APCC MME seasonal datasets were applied for the new chill day model for the Changbangjosaeng and Youmyeong cultivars at the Suwon site. The values of the goodness-of-fit measures using the selected synthetic daily maximum and minimum temperatures reflecting APCC　MME seasonal datasets were worse than those using those collected from the Suwon station (119). We surmise that this is partially due to the errors generated by each step in applying the APCC MME seasonal datasets for the prediction of the full blooming dates of fruit tree. It should be noted that this study does not address the question of how the uncertainties of the APCC MME seasonal datasets and the errors from the selection of the synthetic temperatures generated by MEBWG can contribute to the overall errors of the prediction of the full blooming dates. It is recommended that this question be addressed in a future study.

The prediction models for the frost occurrence events were developed using the logistic regression and decision tree techniques. The threshold values determined for the logistic regression model varied with the range of 0.42 (Jinju station) to 0.59 (Daegu station). The HR, POD, and FAR values resulted from these threshold values were approximately 0.9, 0.8, and 0.2 respectively. However, the FAR value for the Daegu station was as high as 0.725. This high value can be explained by the fact that the Daegu station has one of the fewest frost occurrence events in Korea. A total of 7 meteorological variables were selected for the prediction model. This result suggests that the logistic regression model may not be adequate for operational activities to prevent frost damages. On the other hand, only 2 or 3 explanatory variables were selected for the frost occurrence event predictions using the decision tree technique. In addition, the values of HR, POD, and FAR for the decision tree model were slightly better than those for the logistic regression model. These results imply that the decision tree model for the frost occurrence event prediction can be more useful in providing a timely warning for the prevention of the frost damages to orchards.

Prediction models for full blooming dates that reflect the characteristics of the location and cultivar were developed in this study. However, further work is recommended to develop an operational system for early predicting of the full blooming dates for stakeholders including farmers. For this system development, collaboration with RDA is suggested so that the system can properly function according to user (i.e., RDA) preference. Since these models can better predict the full blooming dates for each location and cultivar, these models can be useful for the preparations of the peach blossom festivals distributed in multiple regions in Korea. To provide enough time to prepare and advertise the peach blossom festivals, the proposed models should be improved to predict the full blooming dates earlier than the late March. These prosed models in this study can be also used to project the quality of fruits by investigating the impacts of climate changes on periods of breaking rest and flowering; it is widely known that these breaking rest and flowering periods largely influence the quality of fruits. The prediction models of frost occurrence events can be used to prevent frost damages to orchards in spring by providing timely interventions for frost damages. The models can be developed for fall to prevent first frost damages to agricultural crops. However, since the frost occurrence events are very local and are affected by local topological characteristics, the models should be improved to reflect these local topological characteristics. It is concluded that this study can be useful to prevent damages to agricultural farms and facilities.

1. 서론

2. 연구 자료 및 방법

2.1 자료수집

2.1.1 복숭아 발아기 및 만개기 자료

2.1.2 관측 기상자료

2.1.3 APCC MME 자료

2.1.4 서리 현상일자료

2.2 복숭아 만개기 예측모델개발

2.2.1 다중회귀모형

2.2.2 발육속도모형

2.2.3 순차휴면모형

2.3 APCC MME 자료 적용

2.3.1 통계적규모축소법

2.3.2 날씨생성기 (Weather generator)

2.4 봄철 서리현상일 예측모델개발

2.4.1 로지스틱 회귀 (Logistic Regression)

2.4.2 의사결정트리 (Decision Tree)

2.4.3 검증

3. 연구결과 및 토의

3.1 복숭아 만개기 예측모델개발

3.1.1 다중회귀모형

3.1.2 발육속도모형

3.1.3 순차휴면모형

3.2 APCC MME 자료 적용

3.2.1 통계적규모축소법

3.2.2 날씨생성기 (Weather Generator)

3.2.3 복숭아 만개기 예측

3.3 봄철 서리현상일 예측모델개발

3.3.1 로지스틱회귀

3.3.2 의사결정트리

4. 활용 방안

4.1 복숭아 만개기 예측모델

4.2 봄철서리현상 예측모델

5. 요약 및 결론

참고 문헌