Performance of simple logistic models for predicting wheat rust occurrence in Ethiopia.

(A-B) wheat stripe rust; (C-D) wheat stem rust; (E-F) wheat leaf rust. Two simple logistic models were used to predict wheat rust occurrence: a temporal model (model 1, see Eq 2) predicting wheat rust occurrence as a function of the time since the start of the main wheat season and a spatiotemporal model (model 2, see Eq 3), predicting wheat rust occurrence as a function of the time since the start of the main season and the location in Ethiopia (latitude, longitude, and altitude). Model performance was tested by fitting the models to training data from all but 1 year of surveys and then conducting a ROC analysis for testing the performance of the fitted model against the data from the year not used for fitting (repeated for every year). The upper row shows the resulting AUC score of both models for each year and all rusts. The bottom row shows the corresponding ROC curves of one exemplar year. For the analysis illustrated here all survey entries with non-zero disease incidence were classified as “diseased” and all surveys with zero incidence were classified as “healthy”. The testing procedure was also conducted using an alternative dichotomization scheme classifying all surveys with moderate or high incidence values as “diseased” and all surveys with zero or low incidence as “healthy” (see S13 Fig).