Artificial neural network models of wheat leaf wetness

Most bacterial and fungal plant pathogens require free moisture for infection and reproduction so a disease forecasting model needs to estimate wetness duration on plant surfaces if direct measurement is unavailable. Artificial neural network (ANN) models with a backpropagation architecture were developed to predict wetness on wheat flag leaves using environmental variables recorded at 0.5 h intervals by an electronic datalogger. For comparison, multivariate discriminant and stepwise logistic models also predicted wetness, and dew periods were predicted with a classification tree-discrimination approach, a relative humidity indicator, and a physical model. ANN and logistic models correctly classified 93% and 90% of the validation cases and were more accurate than discriminant models. Many of the ANN errors occurred during times of dew onset or evaporation and were less than or equal to 1 h per event. Model accuracy for all methods improved 3-4% when input data from a wetness sensor positioned above vegetation were included. The average absolute error per night for prediction of dew period was 0.6 h for the classification-discrimination model, 0.8-1.1 h for ANN models, 1.2 h for a logistic model, 1.4 h for the physical model, and 2.1 h for the relative humidity index. ANN and logistic models thus performed well compared to previously developed models for prediction of dew duration and these models predicted leaf wetness from both dew and rain, an advantage for disease forecasting. Further improvements in ANN performance are possible, making the technique a viable research tool. (C) 1997 Elsevier Science B.V.