RECONCILING PRIVACY, EXPLAINABILITY, AND FEDERATED LEARNING IN DECENTRALIZED PRECISION AGRICULTURE

Abstract

Precision agriculture is decentralized, relying on heterogeneous datasets provided by farms, cooperatives, and sensing platforms; nevertheless, these datasets cannot be centrally aggregated owing to privacy, regulatory, and economic limitations. Although deep learning has achieved significant performance increases in crop monitoring and disease detection, it not only uses centralized training, which conflicts with the distributed nature of agricultural data, but also leads to issues of confidentiality, accountability, and trust.  Previous studies have mostly studied federated learning, privacy protection, and explainable artificial intelligence individually, with no evaluations of their complexity in situations that are defined by non-IID data, sparse connectivity, and edge-computational limitations. We provide an analytical synthesis in this study, where we consider federated learning, privacy, and explainability as design requirements that are mutually dependent. We propose a common taxonomy of architectures, federated learning frameworks, privacy preservation methods, explainability methods, data formats, and deployment platforms. Through comparative analysis, we reveal trade-offs between predictive accuracy, interpretability fidelity, communication overhead, and privacy robustness, and new challenges, especially the instability of explanations, lack of auditable and decentralized benchmarks, and  trade-off between privacy and utility.