Predicting meteorological-disaster-induced damage to melon crops in Taiwan by using decision-tree-based algorithms

Abstract

Agrometeorological disasters pose substantial risks to melon ( Cucumis melo L.) production. This study developed interpretable decision-tree-based models to predict both the occurrence and severity of crop damage caused by typhoon-induced rainfall (TIR), non-typhoon-induced rainfall (NTIR), and cold events—the primary hazards affecting melon cultivation in Taiwan. Classification and regression tree (CART) and M5′ regression tree models were used to predict damage occurrence and damage percentage, respectively, based on county-level disaster records (2003–2022) and meteorological data. The model-derived decision rules were consistent with Taiwan’s climatic characteristics and identified an average temperature of approximately 22.4 °C as an empirical threshold for melon cultivation based on the present dataset. The CART models achieved reliable classification performance, with recall, precision, accuracy, and kappa values exceeding 82.46%, 88.24%, 88.43%, and 76.65%, respectively. The M5′ model showed the highest accuracy for predicting damage percentage in cold events but performed less effectively for TIR and NTIR events. More complex models (random forest and Cubist) did not yield consistent improvements, suggesting that performance limitations are primarily due to missing key predictors rather than model complexity. Overall, this study highlights both the potential and limitations of machine learning for agricultural risk assessment under data constraints.