Forecasting the spread of aerially transmitted crop diseases with a binary classifier for inoculum survival

• Published in: Plant pathology 2018-05, Vol.67 (4), p.920-928
• Main Authors: Skelsey, P., Dancey, S. R., Preedy, K., Lees, A. K., Cooke, D. E. L.
• Format: Article
• Language: English
• Subjects: Blight; Case studies; classification model; Classifiers; Crop diseases; Crops; decision support system; disease risk; Epidemics; Forecasting; Health risks; Inoculum; Inspection; Late blight; Outbreaks; Phytophthora infestans; Plant diseases; Potatoes; Predictions; Solar radiation; Sporangia; spore survival; Statistical models; Survival; Transportation; Viability; Warning systems
• ISSN: 0032-0862; 1365-3059
• DOI: 10.1111/ppa.12808

The risk of between‐field spread of disease is typically omitted from crop disease warning systems, as it is difficult to know the number and location of inoculum sources and thus predict the abundance of inoculum arriving at healthy crops. This study explores the utility of a simple approach to predicting risk of between‐field spread, based on the estimated probability that inoculum will survive the transportation process. Using potato late blight as a case study, the effect of solar radiation on the viability of detached Phytophthora infestans sporangia was assessed. A model to estimate the probability of spore survival was derived using a binomial generalized linear mixed model (GLMM), and receiver operating characteristic (ROC) curve analysis and cross‐validation were used to evaluate the global performance of the model as a binary classifier for discriminating between viable and nonviable sporangia. The model yielded an area under the ROC curve of 0.92 (95% CI = 0.90–0.93), signifying an excellent classification algorithm. Inspection of the curve provided a number of suitable decision threshold (or cut‐off) probabilities for discriminating between viable and nonviable sporangia. The classifier was tested as a forecasting system for potato late blight outbreaks using 10 years of outbreak data from across Great Britain. There was a marked differentiation among the cut‐offs, but the best prediction outcome was an accuracy of 89% with an alert frequency of 1 in 7 days. This model can be easily modified or the methodology replicated for other pathosystems characterized by airborne inoculum.