The history of botany is fundamentally a history of its visual culture, shaped by a contested global encounter between diverse indigenous traditions and the classificatory systems of European empires. Historically, the botanical archive functioned as a “visualization machine” designed to facilitate imperial administration through remote observation (Bleichmar, 2012). Within this visual epistemology, illustrations often served as mobile proxies for physical specimens, effectively erasing the local environmental context and geographic origins of plants to fit them into the reductive, universalizing logic of Linnaean taxonomy. These imperial practices frequently marginalized or subsumed sophisticated epistemic traditions, such as those found in the Chinese Bencao systems, Mesoamerican codices, and Ayurvedic herbals, which prioritized different modes of relating to the natural world.

To address these historical erasures, we introduce Phyto-Vision, a scalable Digital Humanities workflow situated within the emerging field of Digital Plant Humanities (DPH). This discipline brings together environmental and digital frameworks to recover marginalized narratives of botanical life and human-flora relations (Arthur & Ryan, 2024). Phyto-Vision is designed to transform fragmented historical imagery into a quantifiable corpus by treating the digitized page not merely as a carrier of information, but as a visual territory for critical analysis.

To ensure technical interoperability and sustainability, Phyto-Vision implements a multi-stage pipeline focused on reproducible research practices. We first executed a metadata unification phase, adapted from the ZuantuSet method (Mei et al., 2025), normalizing data into machine-actionable schemas. To maintain archival integrity across disparate digital libraries, we performed archival deduplication, identifying and merging overlapping editions through image collation techniques (Kaoua et al., 2021). This infrastructure ensures that the resulting dataset follows FAIR principles, making the archive discoverable for longitudinal study and cross-cultural comparison.

The extraction core utilizes the YOLOv11 architecture (Jocher & Qiu, 2024). We curated a seed corpus of 781 manually annotated botanical illustrations representing various global traditions, sourced from the Library of Congress, Bibliothèque nationale de France, Royal Horticultural Society, Yale University Library, and the National Archives of Japan. To ensure model robustness against material degradation and varying visual grammars, we applied a rigorous data augmentation strategy involving horizontal flipping, random cropping, and noise injection, which expanded the training set to 1,875 examples. The fine-tuned model demonstrated exceptional efficacy for cultural heritage data, achieving a peak mean Average Precision (mAP@50) of 96.9%, with precision and recall reaching 95.5% and 90.3%, respectively.

Beyond detection, Phyto-Vision implements a domain-specific taxonomy to categorize iconography by form, specifically morphological diagrams, habit illustrations, and decorative borders. This allows for a visual–syntactic classification that measures the persistence of indigenous knowledge versus colonial classificatory systems. By automating the curation of these visual series, we provide a pathway for decolonizing visual epistemologies in the plant humanities. This poster showcases the end-to-end workflow of the project, and to support community reuse, the Phyto-Vision pipeline and normalized metadata will be made available as an Open Access resource.