Machine Learning and Deep Learning in Agriculture: A PRISMA Systematic Review of Architectures, Applications, and Open Science Practices (2019–2026)

Agriculture faces compounding pressures from food insecurity, climate change, and resource scarcity, creating urgent demand for scalable analytical tools. This PRISMA 2020-compliant systematic review synthesises 582 peer-reviewed studies on machine learning (ML) and deep learning (DL) applications in agriculture, drawn from Scopus for the period January 2019 to March 2026. The 2026 data cover only the first quarter (January–March) and are therefore not directly comparable to full-year counts. Publication volume grew exponentially — from 6 papers in 2019 to 251 in 2025 — driven by the adoption of convolutional neural networks (CNNs), Vision Transformers (ViT), and YOLO-based object detectors. Plant disease detection (27.0%) and crop yield prediction (13.7%) dominated the application landscape. South Asia and East Asia together contributed 59.3% of the corpus, while Sub-Saharan Africa and Latin America each accounted for only 1.4%, revealing a profound mismatch between research output and global food insecurity burden. Median reported classification accuracy was 98.1% for disease detection, largely reflecting controlled laboratory datasets rather than field conditions. Median R² was 0.823 for yield prediction, based on 22 of 80 yield studies reporting this metric. Unit heterogeneity, dataset artefacts, and inconsistent evaluation practices limit cross-study comparability and the real-world interpretability of these figures. Open science practices remain critically low: only 7.7% of papers shared code and 14.1% shared data openly. Explainable AI, federated learning, and physics-informed modelling represent emerging frontiers. The review identifies benchmark standardisation, smallholder-relevant design, and geographic equity as the field's most pressing unresolved challenges.