Understanding the Impact of Seasonal Weather Dynamics on Rice Disease Occurrence Using Neural Networks: A Case Study of Panicle Blast and Grain Rot

Panicle blast (PB) and grain rot (GR) are two major rice diseases that directly affect panicles and result in severe yield losses worldwide. This study introduces a novel data-driven approach to understanding the impact of seasonal weather dynamics on the occurrence of these diseases using neural networks. By relying solely on meteorological data, the proposed method demonstrates the potential to elucidate hidden relationships between meteorological conditions and disease occurrence. In this study, time-series data comprising seven meteorological variables over 180 days until the peak incidence dates of each disease were used to train a long short-term memory-based model. By applying the holdout method, the prediction model achieved maximum test accuracies of 64.9 and 68.0% for the PB and GR, respectively. Subsequently, a gradient-based analysis further reinforced the reliability of the resulting models by showing consistency with previous findings, in which rainfall and wind speed were frequently identified as critical variables for disease prediction. The temporal dynamics of individual meteorological variables, contributing to disease occurrence, were also revealed from the gradient-based analysis. Overall, our results emphasize the reliability of deep learning models when predicting disease occurrence using only meteorological data, making a substantial contribution to the crop disease prediction system development, and the scalability of applying the same method to other crop diseases when sufficient data are available.