The Algorithm-to-Action Gap in Computer Vision-Based Medicinal Plant Identification: A Systematic Review (2015 – October 2026)

Deep-learning models identify medicinal plants from images with a median reported accuracy of 96.4%. Yet only 4 of 136 full-text studies tested their model on field images, and only 2 describe a deployed application. Misidentified herbal material is a safety and quality risk, and image-based deep learning is widely proposed to prevent it. Whether reported accuracy translates into reliable use outside the laboratory has not been examined systematically. We searched PubMed, OpenAlex, Europe PMC, Semantic Scholar and arXiv for studies published from 2015 to October 2026, with the search run on 13 October 2026 (2,169 records; 1,239 after de-duplication). Titles and abstracts were screened with a written rubric by two independent passes of a large language model, with adjudication of all 127 disagreements. Of 752 records advanced, 150 open-access full texts were retrievable and extracted. For each study, we recorded data provenance, data splitting, leakage risk, external and field testing, look-alike and unknown-class handling, and deployment evidence. Each study was then graded on a six-level Deployment Readiness Level (DRL 0–5) scale and an eight-item reporting-quality score. 136 studies met the criteria; 128 were published from 2022 onwards. The median study used 15 species (IQR 6–50) and 4,000 images (IQR 1,330–7,590). 119 studies (88%) reached only DRL 0–1, meaning a single-dataset laboratory benchmark. 134 (99%) used a random or unreported split, and leakage risk was judged high in 17 and possible in 89. External testing (7 studies, 5%), field testing (4, 3%), look-alike or toxic-species testing (3, 2%), unknown-class handling (3, 2%) and user studies (7, 5%) were rare. Code was available for 5 studies (4%). Forty studies built an application and two reported deployment. High accuracy in this literature mostly reflects within-dataset performance, not readiness for use. Median reported accuracy did not rise with readiness level — the few studies at higher levels reported lower figures — but three to four studies per level make this comparison descriptive only. Studies should split by plant or site, test on independent field images and look-alike species, handle unknown classes, and release code and data. These findings come from the 150 of 752 candidate papers whose full text was openly available, so they describe the open-access literature and may not generalize to all of it.