Early detection of plant and crop diseases is vital for achieving sustainable agriculture and global food security. Traditional inspection methods are often slow and subjective, whereas Artificial Intelligence (AI) techniques offer fast, scalable, and objective alternatives. Following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, this systematic review synthesizes 145 studies published between 2023 and 2025 that employ Machine Learning (ML), Deep Learning (DL), Explainable AI (XAI), and Federated Learning (FL) for plant and crop disease classification. The studies are organized by crop species, imaging modality, and model architecture to evaluate performance in terms of accuracy, robustness, interpretability, and privacy preservation. Results reveal that XAI techniques such as Grad-CAM, LIME, and SHAP enhance transparency and trust, while FL enables decentralized and privacy-aware collaboration across distributed agricultural datasets with only minimal reductions in model accuracy compared to centralized training. Despite strong results in controlled conditions, many models struggle to generalize under real-field variability due to data imbalance and environmental factors. Emerging directions include lightweight edge architectures, domain adaptation, and unified explainable-federated frameworks. Overall, this review identifies FL and XAI as complementary drivers of transparent, privacy-preserving, and scalable AI systems for sustainable precision agriculture.

The rapid expansion of digital agriculture driven by sensors, drones, IoT devices, and remote sensing technologies has generated unprecedented volumes of farm-level data, intensifying concerns surrounding privacy, ownership, and secure data sharing. Federated Learning (FL) has emerged as a promising paradigm that enables distributed model training without exposing raw data, making it well suited for precision agriculture. This systematic review examines 90 peer-reviewed studies published between 2022 and 2025 across major academic databases to evaluate how FL is being applied to enhance agricultural intelligence while reducing raw-data sharing and, in fewer studies, using explicit privacy mechanisms. This review distinguishes between (i) baseline FL, where privacy is primarily achieved by keeping raw data local and (ii) FL augmented with explicit privacy-preserving mechanisms such as differential privacy, homomorphic encryption and secure multi-party computation. Besides, it highlights that horizontal and cross-silo FL architectures dominate current research, particularly in crop disease detection, yield prediction, smart irrigation, and IoT-based farm monitoring. A key finding is that most agricultural FL studies still rely on baseline FL without formal privacy guarantees, and the adoption of DP/HE/SMPC remains limited highlighting a major research gap. Although FL often matches centralized deep-learning accuracy, real-world adoption is limited by non-IID data, communication costs, and constrained compute/connectivity in rural settings. This review summarizes FL architectures, agricultural applications, and the limited use of explicit privacy mechanisms, highlighting key gaps and research directions (personalized/hierarchical FL, federated reinforcement learning, explainable and energy-efficient models, and quantum-enhanced FL).