Exploring RNN Model Implementations for Predicting Sunflower Disease Outbreaks

Abstract

The development of diseases in oilseed fields is influenced by meteorological conditions and the pathogen's predilection for susceptible hosts. This research examines forecasting models for two sunflower diseases: Alternaria leaf blight and powdery mildew. The disease percentage for the Alternaria leaf blight and powdery mildew is predicted for the Kharif and Rabi seasons, respectively. For prediction, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bi-directional Long Short-Term Memory (BILSTM), and Simple Recurrent Neural Network (SimpleRNN) models are employed. These models incorporated six parameters: precipitation, maximum temperature, minimum temperature, maximum and minimum relative humidity, and disease percentage. Meteorological data and disease percentages were obtained from India (Marathwada region). After the experimental study, the results indicate that the SimpleRNN model demonstrated superior performance for Alternaria Leaf Blight with a Mean Squared Error (MSE) value of 0.11%, while for Powdery Mildew, LSTM exhibited the best performance with an MSE value of 0.32%. Each model exhibits unique performance characteristics, and all models are evaluated using the same dataset. A non-parametric Friedman test is employed to statistically validate the differences in performance, followed by a Nemenyi Test as a post-hoc analysis. This approach enables a side-by-side comparison of the average performance across all models.