Neural network classification of tan spot and Stagonospora blotch infection periods in a wheat field environment

Tan spot and Stagonospora blotch of hard red spring wheat served as a model system for evaluating disease forecasts by artificial neural networks. Pathogen infection periods on susceptible wheat plants were measured in the field from 1993 to 1998, and incidence data were merged with 24-h summaries of accumulated growing degree days, temperature, relative humidity, precipitation, and leaf wetness duration. The resulting data set of 202 discrete periods was randomly assigned to 10 model-development or -validation (n = 50) data sets. Backpropagation neural networks, general regression neural networks, logistic regression, and parametric and nonparametric methods of discriminant analysis were chosen for comparison. Mean validation classification of tan spot incidence was between 71% for logistic regression and 76% for backpropagation models. No significant difference was found between methods of modeling tan spot infection periods. Mean validation prediction accuracy of Stagonospora blotch incidence was 86 and 81% for backpropagation and logistic regression, respectively. Prediction accuracies of other modeling methods were less than or equal to 78% and were significantly different (P = 0.01) from backpropagation, but not logistic regression, results. The best back-propagation models of tan spot and Stagonospora blotch incidences correctly classified 82 and 84% of validation cases, respectively. High classification accuracy and consistently good performance demonstrate the applicability of neural network technology to plant disease forecasting.