Decentralized Federated Learning for Agricultural Plant Diseases Identification on Edge Devices

Agricultural plant diseases identification during harvesting is highly important for ensuring food quality and quantity. The integration of edge devices facilitates real-time identification in agricultural settings since early addressing of plant diseases enables effective containment and prevents their further spread. With the increasing computational power of edge devices, complex machine learning models can be directly deployed and trained on edge devices. Federated learning enables training models collaboratively across distributed edge devices. In this paper, we proposed a resource-efficient model to identify agricultural plant diseases for limited-resources devices. As a result, the memory usage of the proposed model is reduced by 67.58% compared to MobileNet, the state-of-the-art lightweight model. By eliminating the reliance on centralized servers in centralized federated learning, fault tolerance can be enhanced. This paper introduces a decentralized federated learning mechanism on edge devices. Compared to traditional decentralized federated learning, our decentralized mechanism significantly reduces communication costs while preserving model accuracy.