Artificial neural networks and multiple linear regression as potential methods for modeling seed yield of safflower (Carthamus tinctorius L.)

Abstract

ABSTRACT Modeling relationships between safflower seed yield (SY) and its components would be helpful to understand the most important trait-(s) with significant effects on seed yield, which in turn helps safflower breeders in the indirect selection of high-performance cultivars. For this purpose, we evaluated performance of five different artificial neural network (ANN) models, including Generalized feed forward (GFF), multilayer perceptron (MLP), Jordan/Elman (JE), Principal component analysis (PCA), and Radial basis function (RBF) with different learning algorithms, transfer functions, hidden layers and neuron in each layer, along with multi-linear regression model to predict seed yield of safflower. A panel of 25 lines/cultivars was used during two years (2012-2014) to evaluate some of the important traits related to seed yield. PCA was used to perform feature extraction in the first stage of modeling. Based on PCA and first two components that incorporated approximately 79% of the total variability, five agro morphological and phenological traits, including plant height (PH), number of branches per plant (NBP), number of capsules per plant (NCP), thousand seed weight (TSW), and number of seeds per capsule (NSC), due to high values of Eigen vectors were selected as input. The performance of the models in predicting seed yield was determined using some of the statistical quality parameters, including coefficient of determination (R2), root mean square error (RMSE) and mean absolute error (MAE). The MLP model with the Levenberg-Marquardt learning algorithm, sigmoid axon transfer function, two hidden layers (i.e. topology 5-5-41) and 1500 training epochs was selected as the best model to predict seed yield of safflower. Sensitivity analysis indicated that the NCP is the most influential trait to predict SY in the MLP/ANN model.