Comparative analysis of models integrating synoptic forecast data into potato late blight risk estimate systems

Determinacy analysis, logistic regression, discriminant analysis and neural network models were compared for their accuracy in 5-day (120 h) forecasts of daily potato late blight risk according to a modified-Wallin disease severity model. For 12 locations in Michigan, variables derived from extended forecast data (MEX) from the National Weather Service model output statistics (MOS) were compared with those similarly derived from Unedited Local Climatological Data (ULCD) for the growing seasons 2001–2004. The most effective model for late blight risk prediction based on comparison with risk estimated with ULCD was a resilient propagation (Rprop) neural network model with 49 variables and 10 hidden nodes. The neural network model had significantly higher overall accuracy than the other models, and was particularly successful at predicting risk values in June, July, and August when knowledge of potato late blight risk is most critical to growers making management decisions with regard to fungicide sprays and irrigation scheduling. The neural network model was also significantly more accurate than the regional average of days with no late blight risk (0.72%). For each of the four models, monthly accuracy at any single station was negatively correlated with the percentage of days per month classified as risk days for potato late blight (P = 0.01). Although no validation with disease data was conducted, such models are still useful in the context of advising growers of forecast conditions that may be favorable for late blight according to model values, such as Wallin style disease severity values, with which they are familiar.