Application of mixture of experts models for the recognition of pests and diseases in maize

Abstract

Manual monitoring of pests and diseases in maize crops requires considerable time and resources, significantly increasing production costs. Artificial intelligence (AI)-based studies have explored their automated detection, primarily through transfer learning architectures, although with limited success. This study evaluated and compared four AI approaches: convolutional neural networks (CNN), a hybrid CNN with support vector machines (CNN-SVM), mixture of experts (MoE) models, and transfer learning architectures. Eighteen CNN models were developed and optimized using a factorial design, and the best-performing model was used as the foundation for constructing the hybrid CNN-SVM and CNN-SVM-MoE models. The CNN-SVM-MoE model achieved the highest accuracy (99.14 %) and demonstrated strong generalization capabilities, even with data collected under field conditions. In contrast, transfer learning architectures showed lower performance. Statistical analysis revealed significant differences among the models, highlighting the superiority of the CNN-SVM-MoE approach. The results confirm that MoE models enhance performance in classifying maize pests and diseases and offer strong potential for integration into mobile or embedded devices, enabling their direct application in the field.