Comparison of techniques used in the prediction of yield in banana plants

Abstract

Phytotechnical characters observed in field experimental are of phenotypic nature and most of the time its assessment is based only on the experience of the observer. The assessment of the correlations between variables allows the estimation of the changes in a character based on the changes in other characters. This study investigated the potential of using the culture's characteristics in predicting production responses by applying two techniques: artificial neural networks (ANNs) and multiple linear regression (MLR) in banana plants cv. Tropical. The experiment was a test for uniformity, using the cultivar Tropical (YB42-21), an AAAB tetraploid hybrid. The characteristics evaluated over two cycles of fruit production were the yield, bunch's weight, number and length of hands and fruits, diameter of the fruit, and number of living leaves at harvest. In the evaluations, each plant was considered as a basic unit (bu) occupying an area of 6 m2; therefore, 360 basic units (bu) were studied. According to the analyses, the neural network proved to be more accurate in forecasting the weight of the bunch in comparison to the multiple linear regressions in terms of the mean prediction-error (MPE = 1.40), mean square deviation (MSD = 2.29) and coefficient of determination (R2 = 91%).