Point-trained models in a grid environment: Transforming a potato late blight risk forecast for use with the National Digital Forecast Database

• We compare methods for using gridded weather data in models developed at points.
• Results of direct replacement of point data with grid data were unpredictable.
• Results from simpler models were fairly consistent, but of lower quality.
• Retraining during the use of complex artificial neural network models was essential.
• Spatio-temporal patterns must be fully explored in retraining results.

As publicly available weather forecasting datasets advance in accuracy and spatial and temporal resolution, it is relatively simple to apply these established models to new datasets but the results may deviate from what users of decision support systems have come to expect. Potato late blight risk models were some of the earliest weather-based models. This analysis compares two types of potato late blight risk models that were originally trained on location specific (point) data in Michigan. A unique system using NoSQL was developed to train, validate and implement potato late blight risk modeling using a grid data format. Each model was tested two ways; it was first deployed directly with gridded weather forecasting data as a replacement for point data, and then retrained on the gridded data. Despite consistently lower overall accuracy, the grid trained artificial neural network model was deemed of better quality for use by stakeholders because of its accuracy on days with potato late blight risk. However, the success of the model was dependent upon its retraining using the newly available data source. In the direct implementation scenario without retraining, a simpler modified-Wallin model achieved better results than the neural network model.