Dummy Variable Regression and Artificial Neural Networks for Maize Grain Yields Prediction in Push-Pull Systems

Abstract

Abstract Crop yield prediction is essential for sustainable production planning. In agroecological systems, the traditional linear or non-linear regression models used for this purpose present limitations and robustness problems due to the number of variables generated by the complexity of these systems. Therefore, the present study was carried out with the objectives of 1) fitting multiple linear regression models with dummy variables using the Ordinary Least Squares method (OLS); 2) parameterizing and training Artificial Neural Networks (ANNs) with the backpropagation algorithm; and 3) comparing the performance of both approaches in maize yield prediction in push-pull systems established in Yautepec, Morelos, Mexico. In both modeling approaches, maize grain yield predictive variables were: edaphoclimatic (soil temperature and moisture), phytosanitary (incidence and severity of Spodoptera frugiperda), morphological (leaf area index), and categorical (Blocks, Management Systems) variables. The ANN of architecture MLP 18-13-1 (r = 0.95; RMSE = 12.19%), with hyperbolic tangent activation function in the hidden layer and linear function in the output layer, generated consistent and more accurate predictions than those obtained with the regression equation with dummy variables (r = 0.87; R2 = 0.75; RMSE = 20.38%).