Frost prediction with machine learning techniques

Abstract

Frost is the condition that exists when the temperature of the earth's surface and earthbound objects falls below freezing (0°C). These events may have serious consequences on crop production, so actions must be taken to minimize damaging effects. In particular, temperature predictions are of much help in frost protection decisions by providing the hortoculturist with warnings of critical temperatures. Consequently, reliable temperature prediction methods would be of significant economic value. These methods are usually empirical formulae based on significant correlations between the minimum temperature observed towards the end of the night and one or several meteorological variables measured at least several hours before this minimum is reached. These empirical regressions employ antecedent air temperature, various humidity measures, wind speed and cloud cover values. In this work we explore the possibility of developing an empirical prediction system for frost protection of fruits and vegetables in the southern part of the Santa Fe province in Argentina, in the region covered by the agrometeorological station located at Zavalla (33°01′S, 60°53′W). To this end we consider a handful of Machine Learning techniques usually employed in regression and classification problems, including Artificial Neural Networks, Simple Bayes classifiers and k-Nearest Neighbors. The results obtained in this preliminary study reveal a very noisy structure of the data that allows for only a slight improvement in performance of some of these more sophisticated nonlinear techniques over the standard (linear) multivariate regression equations.