Implementation of Crop Disease Risk Forecasts that Rely on Neural Networks and Medium-Range Numerical Weather Prediction Model Output

Abstract

Neural networks were combined with medium range weather prediction model output to produce five day crop disease risk forecasts for potato late blight in the Great Lakes region of the United States. Initial analysis used 2000-2004 data from 12 National Weather Service stations in the region as input. The neural network model achieved an average accuracy of 81 percent when predicting risk on a Boolean scale (0 = no risk, 1 = risk) according to a modified version of the Wallin potato late blight model. Neural networks resulted in more accurate predictions than did similarly constructed logistic regression, determinacy analysis and discriminant analysis models. The neural network model was more accurate on non-risk days than on risk days, but the accuracy of predictions on both day types was significantly higher than expected based on climatic norms. Risk days were predicted with higher accuracy in July and August, when risk was more frequent, while non-risk days were predicted most accurately at the beginning and end of the growing season. Potato late blight risk predictions will be implemented as part of an automated expert system that assists growers in the management of fungicide application rates and timing. Daily updates in recommendations at locations throughout the region are uploaded and disseminated through the Internet.