Frost prediction using a combinational model of supervised and unsupervised neural networks for crop management in vineyards

Abstract

Frost prediction models could contribute significantly towards the successful growth and production of quality crop yield in horticulture, especially in precision viticulture where the benefits are significant because frost damage is well-known for its potential leading to total harvest failure, with a follow-on regional or national economic impact outcome. This reality has increased interest among scientists and growers to advance their knowledge in relation to the inter-dependencies and possible correlations between meso-micro climate variables and associated plant and soil condition values. Included in this array of variables there are also site specific environmental factors such as pesticide saturation and carbon density. Recent interest in building computational models is focused on predicting frost events using both regional (metrological) and vineyard weather monitoring data gathered via remote low-cost sensors, in addition to vineyard-specific environmental attribute data. This paper outlines some earlier research and then describes the climate and atmospheric data analysis, together with vineyard elevation and wind data in order to determine the inter-dependencies of variable values that inform enhanced frost protection measures. Developing a model with supervised and unsupervised neural networks as a means for characterising the data being analysed, is the focus of the investigation described here, which is part of a wider project called Eno-Humanas. This wider project incorporates the modelling of interrelations between grape wine sensory properties, such as aroma, colour and taste, and climate in addition to environmental factors for crop management and prediction purposes. The Eno-Humanas prediction models are intended to contribute to the body of knowledge aimed at finding grape varieties that best suit different potential future climate scenarios.