Specialized stacking ensemble with optimized base models for the automated detection of pests and diseases in maize

Abstract Maize pests and diseases significantly reduce crop yields and cause substantial economic losses. Manual monitoring is slow and difficult to scale, highlighting the need for automated diagnostic systems. This study presents a specialized two-stage stacking ensemble model for the automated detection of ten classes of maize pests and diseases from digital leaf images. The proposed method integrates optimized base models derived from seven transfer learning architectures under both feature extraction and fine-tuning schemes, combined through a stacking ensemble strategy optimized via Bayesian optimization. All possible combinations of base models were systematically evaluated to identify the most effective ensemble configuration. Furthermore, a specialized two-stage stacking model was implemented to enhance discrimination between visually similar classes, particularly between plants infested with fall armyworm and healthy plants. The final model achieved outstanding accuracy, surpassing previously reported results, demonstrating the effectiveness of the two-stage design and the optimized stacking ensemble approach. These findings indicate that integrating transfer learning models within a specialized stacking ensemble provides a robust, accurate, and scalable solution for the automated detection of maize pests and diseases.