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

CHANTRE GUILERMO; LODOVICHI, MARIELA V.; BLANCO ANÍBAL M.; BANDONI J. ALBERTO; SABBATINI, M. RICARDO; LOPEZ RICARDO; VIGNA MARIO; GIGÓN RAMÓN

Cordoba

Congreso; 2do Congreso Argentino de Bioinformática y Biología Computcional; 2011

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 temperature (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.  non-transformed  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 non-linear regression techniques.