An Efficient Deep Learning Framework for Agricultural Pest Classification via ResNet50V2 Transfer Learning and Edge Deployment

Agricultural pest infestations significantly contribute to global crop yield losses, with the Food and Agriculture Organization estimating annual reductions of 20-40% in major cereal production. Therefore, early, accurate, and automated pest identification is crucial for sustainable crop management and food security. This study introduces an efficient deep learning framework based on the ResNet50V2 pre-activation residual architecture for the automated classification of 13 agricultural pest species. A balanced dataset of 14,872 images was assembled by combining two diverse sources and applying targeted augmentation to ensure uniform class distribution. The model was trained using a structured two-phase transfer learning approach. During initial training, validation accuracy reached 96.33%, while full fine-tuning enhanced performance to a maximum validation accuracy of 98.69% with a validation loss of 0.0558. Per-class F1-scores surpassed 0.95 across all categories, resulting in a macro-average F1-score of 0.982. A systematic ablation study verified the contribution of each architectural and training component. The trained model was converted to TensorFlow Lite through post-training dynamic range quantization, reducing the model size by approximately 67% and enabling real-time on-device inference without cloud connection. These findings demonstrate that the proposed framework offers a lightweight, high-accuracy solution ideal for practical edge-based agricultural pest monitoring systems.