FPGA-based Mb2DKDNet for field disease and pest diagnosis

• Proposed Mb2DKDNet model for intelligent diagnosis of rice pests and diseases in the field. • Constructed MIRD dataset covering diverse regions and rice growth stages. • Deployed on FPGA to enable efficient real-time diagnosis in agricultural machinery. • Achieved 98.07 % accuracy, supporting precise pest control and reduced pesticide use. Accurate and efficient identification of rice diseases and pests is crucial for ensuring stable crop yields and sustainable agricultural development. It is necessary for crop protection to realize the in-time identification in the relatively short control window for some plant diseases and pests. In this study, we propose Mb2DKDNet, a lightweight convolutional neural network (CNN) optimized through Decoupled Knowledge Distillation (DKD) to enhance classification performance while maintaining computational efficiency. To further improve real-time inference capability, we implement a field-programmable gate array (FPGA)-based acceleration scheme, incorporating matrix blocking and pipelined vector multiplication techniques to optimize computational efficiency. As an essential part of this study, we collected and constructed a high-quality dataset, named Multi-scale Image Dataset for Rice Disease and Pest Diagnosis (MIRD), which spans multiple regions and time periods. Experimental results demonstrate that Mb2DKDNet achieves a classification accuracy of 98.07 %, representing a 1.25 % improvement over the baseline model, with only 2.23 M parameters. The FPGA implementation attains a computational throughput of 21.3 GFLOPs and an inference speed of 79 milliseconds per frame, ensuring efficient real-time processing in resource-constrained environments. The proposed method provides a promising solution for intelligent pest and disease monitoring in precision agriculture, contributing to improved crop protection and reduced reliance on chemical pesticides.