A feedforward neural network model for predicting Avena fatua seedling emergence in the field

Abstract

Avena fatua is an invasive weed affecting winter cereal crops of the semiarid temperate region of Argentina. Seedling emergence patterns show great variability between years mainly due to a highly unpredictable annual precipitation regime and a fluctuating thermal environment. Artificial feedforward neural networks (AFNN) are powerful tools for modeling non-linear relationships between variables thus showing a high potential applicability in ecological systems [1]. The quality of AFNN models depends on a proper setting of neural network architecture and other influential parameters (i.e. learning algorithm, transfer functions, number of neurons in hidden layers, etc.) [2]. The objective of the present work was to compare different AFNN architectures in order to obtain a model with the best capability to predict the proportion of daily emerged seedlings in the field as a function of meteorological data. Three input variables were considered for the network configuration: time (Julian days), mean daily air temperatura (ºC) and accumulated precipitation per day (mm). A total number of 528 input/output data pairs corresponding to 11 years of data collection were divided into training (64%), validation (18%) and test (18%) subsets. The evaluated model scenarios resulted of varying: the number of neurons in a single hidden layer (1, 2, 3, 5, 8, 10, 12, 15, 20), the combination of transfer functions in the hidden-output layer (sigmoid-linear, sigmoid-sigmoid), the learning algorithm (Levenberg-Marquardt, Resilient Backpropagation, Bayesian Regularization) and input/output data scaling (normalized vs. nontransformed data). From the 108 models tested, an AFNN with input/output data normalization configured with 3 neurons in the input layer, 20 neurons in the hidden layer, 1 neuron in the output layer, sigmoid transfer functions in both hidden and output layers and a Bayesian Regularization training algorithm gave the best prediction of independent emergence data (RMSError=0.0487). Obtained results clearly point out the potential implementation of AFNN models as accurate predictive tools in leed management decision support systems. Further studies should be performed in order to compare the generalization performance and accuracy of AFNN models compared to traditional nonlinear regression techniques.