Comparison of Methods for Harvest Prediction in ‘Gigante’ Cactus Pear

Abstract

Behavior analysis and plant expression are the answers the researcher needs to construct predictive models that minimize the effects of the uncertainties of field production. The objective of this study was to compare the simple and multiple linear regression methods and the artificial neural networks to allow the maximum security in the prediction of harvest in ‘Gigante’ cactus pear. The uniformity test was conducted at the Federal Institute of Bahia, Campus Guanambi, Bahia, Brazil, coordinates 14°13′30″ S, 42°46′53″ W and altitude of 525 m. At 930 days after planting, we evaluated 384 basic units, in which were measured the following variables: plant height (PH); cladode length (CL), width (CW) and thickness (CT); cladode number (CN); total cladode area (TCA); cladode area (CA) and cladode yield (Y). For the comparison between the artificial neural networks (ANN) and regression models (single and multiple-SLR and MLR), we considered the mean prediction error (MPE), the mean quadratic error (MQE), the mean square of deviation (MSD) and the coefficient of determination (R2).The values estimated by the ANN 7-5-1 showed the best proximity to the data obtained in field conditions, followed by ANN 6-2-1, MLR (TCA and CT), SLR (TCA) and SLR (CN). In this way, the ANN models with the topologies 7-2-1 and 6-2-1, MLR with the variables total cladode area and cladode thickness and SLR with the isolated descriptors total cladode area and cladode number, explain 85.1; 81.5; 76.3; 74.09 and 65.87%, respectively, of the yield variation. The ANNs were more efficient at predicting the yield of the ‘Gigante’ cactus pear when compared to the simple and multiple linear regression models.