Estimating spring frost risk in apple farms from crop insurance data

In South Korea, spring frost events frequently deliver catastrophic damage to apple orchards, leading to destabilization of apple markets and causing substantial financial burdens to the crop insurers. The present study aims to establish an analytical framework for spring frost prediction based on a Machine Learning (ML) algorithm and a four-year record of spring frost crop insurance loss claims. A fine-scale, observation-based gridded weather dataset is paired with the insurance dataset to analyze the relationship between spring frost damage and meteorological, climactic, and socioeconomic conditions. The results show that the Random Forest (RF) algorithm outperforms all tested algorithms of classical binary outcome variable regression and discriminant analysis, with an accuracy of > 94 % , and this superiority is robust over different classification thresholds. Farmer socioeconomic information held by the insurers provides additional improvements in RF performance, while the transformation of temperature variable to double-threshold degree days is not significant. The algorithms are applied to the assessment of spring frost risk at insurance-covered farms. We find evidence suggesting “adverse selection,” where farmers purchasing lower deductible plans show higher spring frost risk. The results will help both crop insurers and farmers manage spring frost risk by providing a reliable predicting framework of disaster occurrence and shaping the optimal crop insurance contract.