Neural network model to predict deoxynivalenol (DON) in barley using historic and forecasted weather conditions

Fusarium head blight, caused by Gibberella zeae, produces trichothecene mycotoxins, primarily deoxynivalenol (DON) in the grain of barley. Severe price discounting can occur at the sale of the crop if the grain contains a DON concentration of greater than or equal to 0.5 mg/kg. An artificial neural network model was developed to predict the risk of DON accumulation greater than or equal to 0.5 mg/kg using forecasted data up to 120-hours (5-days) at 36 locations in the U.S. Northern Great Plains. Weather data were collected at 313 location-years for three months (May, June and July) from two sources, quality controlled local climatological data (QCLCD) and extended range forecast model output statistics (MOS). Estimated `true´ risk of economic DON accumulation (DON ≥ 0.5 mg/kg) was calculated for each date using the data from QCLCD based on a previously developed logistic regression model (n=19,093). Four input novel variables were constructed by summarizing the weather data from both QCLCD and MOS sources. A single hidden layer neural network model was developed using 75% of the total observations and was then validated using the remaining 25% of the observations. The overall accuracy of the model was around 90% on both training and validation data sets. The results of this study demonstrate that the risk of DON accumulation of greater than or equal to 0.5 mg/kg can be predicted 5-days in advance during the period leading up to full head emergence using forecasted weather conditions.