Predicting Pest and Disease Occurrence Using Synthetic Data and Explainable Machine Learning Methods

Prediction of occurrence for pests and diseases is an essential problem for agriculture, as such events have a huge influence on the productivity of the crop with regard to the security of food production. Traditional methods lack datasets and tend not to incorporate domain knowledge, which leads to suboptimal performance with limited sets of interpretation. This study addresses such gaps by developing a systematic machine learning-based framework for combining synthetic data generation, robust predictive modeling, and explainability techniques to produce actionable insights in pest and disease dynamics. Synthetic datasets are first generated based on the domain-driven logic simulating the correlations between critical environmental and biological factors such as temperature, humidity, rainfall, pest lifecycle stage, and soil moisture and the incidence of pests or diseases. For interpretability, Local Interpretable Model-agnostic Explanations LIME with Random Forest provides localized, instance-level insights on feature contributions to individual predictions. For complement, permutation importance calculates the global relevance of every feature by assessing its effect on model performance. Both of these techniques ensure that fine-grained and holistic understanding is achieved regarding the model's behavior. This integrated approach therefore addresses the limitations of traditional methods by improving the predictive accuracy and enhancing interpretability. The findings have tremendous implications for precision agriculture in order to allow stakeholders to put into action data-driven strategies for pest and disease management. This framework is reproducible and therefore adaptable to different contexts in agriculture sets.