FedXAI-AgriNet: A Federated Explainable AI Framework for Real-Time Crop Disease Detection on Edge Devices

Abstract

Plant pathogens pose a significant threat to global food security, with the World Food Programme estimating that crop diseases account for 20–40% of annual agricultural losses, directly affecting the livelihoods of more than 500 million smallholder farmers worldwide. However, existing deep learning-based crop disease detection models often suffer from limited generalisation under real-world conditions, lack interpretability, and rely on centralised training approaches that raise data privacy concerns. To address these challenges, this paper proposes FedXAI-AgriNet, a lightweight federated explainable artificial intelligence (XAI) framework for real-time crop disease detection on resource-constrained edge devices. The proposed framework integrates a customised MobileNetV3 architecture with Federated Learning and Grad-CAM++ to enable privacy-preserving, interpretable, and efficient disease diagnosis. The framework is evaluated on a curated dataset comprising 18,160 PlantVillage images and 1,850 field images collected from five geographically diverse farms. Experimental results demonstrate 97.8% classification accuracy on real-world tomato disease images, outperforming centralised baseline models, including ResNet50 (94.2%), MobileNetV2 (92.8%), and HPDC-Net (89.7%). Furthermore, the proposed model exhibits only a 3.2% accuracy drop on unseen field data with varying illumination, occlusions, and mixed disease stages, compared with the 26–30% performance degradation observed in existing deep learning models. The framework also achieves real-time inference at 28 FPS on a Snapdragon 888 smartphone CPU while providing visual explanations that increase farmer trust ratings from 3.1/5.0 to 4.6/5.0 (p < 0.001) in user studies involving 15 agricultural practitioners. Overall, FedXAI-AgriNet provides a scalable, interpretable, and privacy-preserving framework for precision agriculture, enabling collaborative model learning without sharing raw agricultural data and supporting reliable deployment of AI-driven crop disease diagnosis in real-world farming environments