Abstract
Crop diseases constrain global agricultural productivity and food security, particularly for smallholder farmers who lack timely, affordable diagnostic tools. Artificial intelligence and machine learning have transformed automated crop disease detection, but much reported performance rests on controlled laboratory datasets that may not reflect field conditions. This systematic review, following PRISMA 2020 guidelines, synthesizes 35 peer-reviewed studies drawn from 1,210 records across six databases, published 2015–2025 and covering 22 crop species across 17 countries. Convolutional neural networks remain dominant, although Vision Transformers, lightweight architectures, explainable AI, federated learning, and hybrid models are emerging alternatives. Most studies relied on laboratory datasets, reported accuracy alone, and lacked external validation, producing a consistent laboratory-to-field performance gap, particularly in African smallholder contexts. The review identifies gaps in dataset diversity, reporting transparency, and deployment feasibility, and proposes a four-pillar research agenda spanning data, algorithms, deployment, and policy to guide field-eady AI development.