Prediction of frost episodes based in agrometeorological information and machine learning techniques

Frosts are one of the main risks faced by farmers during the winter and spring seasons. These events can cause significant damage to cultivations and crops. In Chile, these frost generates significant losses in the agricultural production sector, causing crop losses of an entire year and compromising the income of the following year, especially fruit and wine growers. In this work we developed a prediction model based on historical agrometeorological information able to predict efficiently and up to 12 hours earlier than the occurrence of a frost event. With a special focus on the central Chilean zone. Various algorithms and machine learning methods were evaluated, we found that the method “Random Forest” exhibits the best results overall. The results obtained in the frost prediction reach over (90%) of efficiency in most of the evaluated scenarios.