Optimizing Federated Graph Attention Networks for Crop Disease Detection in Low-Resource Agricultural Environments

Abstract

The growing integration of Artificial Intelligence of Things (AIoT) in agriculture is revolutionizing how crop diseases are detected and managed. While centralized deep learning models have shown promising results in disease detection, their feasibility in low-resource agricultural environments is often limited by high communication overhead and data privacy risks. Federated Learning (FL) offers a decentralized solution to this challenge, though existing FL models struggle to handle heterogeneous data distributions that are common in real-world farm settings. This study introduces FeGAN, a Federated Graph Attention Network framework, optimized to improve wheat disease detection in resource-constrained agricultural systems. FeGAN combines adaptive attention mechanisms with communication-efficient aggregation techniques to enhance classification accuracy while minimizing bandwidth consumption. The model was evaluated using datasets from PlantVillage and PlantPAD Wheat Collections, containing diverse disease samples such as Brown Rust, Fusarium Head Blight, and Powdery Mildew. Experimental results demonstrate that FeGAN achieved 94% classification accuracy, outperforming traditional FL models which averaged 88–93%. Moreover, FeGAN reduced communication costs by 30% compared to baseline FL models while   converging 40% faster. The model also demonstrated a 25% reduction in energy consumption, making it a suitable solution for deployment on edge devices in remote agricultural environments. FeGAN's scalability, improved efficiency, and privacy-preserving design offer a viable solution for AI-driven smart farming, ensuring accurate disease detection without compromising resource constraints. This study provides insights for developing sustainable agricultural intelligence systems that address the unique challenges of smallholder farmers and low-resource communities.