Improved weather-based late blight risk management: comparing models with a ten year forecast archive

Abstract

SUMMARYAgroecosystem decision support systems typically rely on some types of weather data. Although many new digital weather and forecast datasets are gridded data, the current authors feel that evaluating previous methods with data of increased archive length is critical in aiding the transition to new datasets that lack extensive archives. To that end, the present paper reviews the improvements made to an artificial neural network for forecasting weather-based potato late blight (Phytophthora infestans) risk at 26 locations in the Great Lakes region. Accuracies of predictions made using an early model, developed in 2007, are compared with accuracies of predictions made using a new 10-year hourly optimized model. In nearly every comparison by month, forecast lead time and spatial region, the newly optimized model is more accurate, especially when the weather is conducive to high disease levels.