Forecasting the yield of a 'Valencia' orange´s crop, by means of a generalized-regression neural networks

Abstract

The yield of a given crop is the result of a multiplicity of variables whose complex interactions make its prediction difficult to achieve by regular means. Generalized regression artificial neural networks represent a promising alternative for such a task, due to its ability to model non-linear relationships, without the need of knowing its explicit nature. The present work aimed at assessing such approximation for predicting the potential yield of a crop of ‘Valencia’ orange (Citrus sinensis L. Osbeck), using the concentration of nitrogen, phosphorus, potassium, calcium, and magnesium in the foliar tissue as predicting variables. Special emphasis was placed in the mathematical treatment of the input/output data, using conventional techniques (normalization, standardization, and principal components) as well as other less common (row-centered log ratios, and individual and global nutritional indices from the Compositional Nutrient Diagnosis System). The results showed that, among the ones studied, the individual nutrient indices/normalized yield combination (Prediction error= 0.98 kg·tree^-1), and the unrotated principal components/normalized yield combination (Prediction error= 0.51 kg·tree^-1) resulted in the development of the neural networks with the highest yield prediction capabilities, as evidenced by the previously indicated prediction errors.