A PRISMA-guided systematic review synthesized 83 studies from 2015-2025 on image-based plant-variety classification. Deep learning, especially convolutional neural networks, dominated and often reported high accuracy, while handcrafted features and optimization methods could still improve robustness.
Key findings
- CNN variants were the predominant approach and often outperformed traditional methods. Hybrid use of conventional machine learning, optimization, and handcrafted features may improve robustness. Public datasets and smartphone images were the main data sources, and cereal varieties were the most studied.
Why this matters globally
Accurate variety identification could support seed certification, biodiversity conservation, field monitoring, and farmer-facing mobile tools. The review highlights needs for open datasets, benchmark standards, and cross-region validation.
Thai researcher contribution
The Asian Institute of Technology team led the global evidence synthesis, linking agricultural systems engineering, AI, and biodiversity policy from a Thailand-based institution.
Limitations to consider
The abstract does not detail databases searched, exclusions, risk-of-bias appraisal, or metric comparability. Accuracy may be inflated by small controlled datasets, simple backgrounds, or non-independent image splits, while publication bias may favor successful models.