과수원의 서리 예측을 위한 다중 시간스케일 인공지능 모델

초록

Climate change has led to a delayed onset of seasonal frost. However, the frequency of frost events in major apple-producing regions of Korea has increased, underscoring the necessity of reliable advance prediction. We propose an artificial intelligence(AI) model for open-field orchards that exploits multi-timescale for frost prediction. The dataset includes 97,758 hourly observations from the Andong weather station(STN=136) in Gyeongsangbuk-do from 2014 to 2025. Multi-window inputs of 6, 12, and 24 hours simultaneously capture short-term rapid changes (radiative-cooling), diurnal cycles, and long-term atmospheric circulation patterns. The modeling framework consisted of XGBoost, Convolutional Neural Networks(CNNs), and an XGB-CNN soft-voting ensemble. Data are split into training/validation/test sets at a ratio of 70/20/10%, respectively. Performance evaluation showed that XGB-24h achieved high discrimination and a low false alarm rate with an ROC-AUC of 0.977, a PR-AUC of 0.921, and an FPR of 0.039. CNN-24h obtained the highest recall of 0.941, which minimized missed events but resulted in a relatively higher FPR. The proposed ensemble balanced these factors, achieving an Accuracy of 0.932, a Recall of 0.859, an FPR of 0.046, an MCC of 0.809, a PR-AUC of Approximately 0.919, and the best probability calibration with a Brier score of 0.056. To optimize performance, a two-dimensional grid search was conducted on the soft-voting ensemble model's weight(ω) and the frost detection threshold(θ). This revealed that weight was a more critical parameter than threshold for adjusting ensemble performance. This study suggests that applying multi-timescale inputs, ensembling, and season-specific dynamic threshold policies can further enhance performance. Recognizing limitations in geographical generalization, future work aims for continuous improvement in Recall-centric performance through field validation studies under various regional and climatic conditions.

키워드