Regression and artificial neural regression (ANN) modeling approaches were combined to develop models to predict the severity of gray leaf spot of maize. Regression models were used as a preliminary step to select potentially useful variables to be used in ANN model development. A total of 329 cases were used for model development. These consisted of environmental, cultural, and location specific variables collected from 17 counties in Iowa between 1998 and 2002. All
subsets regression was performed, generating different models from different combinations of 11 input variables. The best nine of 80 preliminary models were selected based on Mallow's Cp criteria, and the variables selected in these models were used to develop ANN models. A three-layer, feed-forward, back-propagation network with three hidden nodes was used to model the data. A random sample of 60% of the cases was used to train the network, and 20% each for testing and validation. The networks with the highest predictive accuracies corresponded well to
the best subsets of variables selected by the regression models. The predictive accuracy of the top four networks ranged from 70 to 75%, with mean squared errors ranging form 174.7 to 202.8. Networks with seven and eight inputs generally performed better than those with nine inputs. The best predictors of gray leaf spot severity were longitude, surface residue, planting date, cumulative hours of daily temperatures between 22 and 30°C and nightly RH > 90% between growth stages V4 and V12, mean nightly air temperature between V12 and R2, and gray leaf spot
resistance rating. Using regression to select predictors prior to fitting ANN models
resulted in faster convergence of networks to a solution when the best subsets of
input variables were used. Four subsets of variables with good predictive accuracies
were identified, allowing for greater flexibility in the choice of variables to be used to
predict gray leaf severity.