Phenobase: a harmonized, AI-ready knowledge base of global plant phenology
Robert Guralnick · From Mobilizing Data to AI-Ready Knowledge: Infrastructure for Multimodal Biodiversity Data
Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.
The short versionCombining ontologies with machine learning lets Phenobase close global phenology data gaps, but making the result AI-ready requires per-record model provenance, quality and citation.
What this was about
Robert Guralnick presented Phenobase, an open, integrated database of global in situ plant phenology. It rests on the Plant Phenology Ontology, which harmonises monitoring-programme terms and supports inference, with extensions for observations of only part of a plant. To fill large geographic and taxonomic gaps, the team trained models on 1.5–2 million iNaturalist phenology annotations (masked-autoencoder pre-training) and a separate ensemble for herbarium sheets. They reached about 98% accuracy for flowers and 95% for fruits, and machine-labelled records now make up about 97% of records and fill many empty grid cells. He said making Phenobase AI-ready was the hard part: per-record model URIs and quality scores, provenance links back to sources, data dictionaries, update strategies and citation policies.
Why it matters. Phenology is a key climate-change indicator. Phenobase shows a working pattern for turning images from citizen science and herbaria into trustworthy, reusable trait data with provenance attached to each record.
In the room
- Phenology monitoring programmes use incompatible, bespoke terminology; the Plant Phenology Ontology provides URIs and logical axioms for integration and inference (e.g. count of leaf buds implies presence).
- Images show only part of a plant, so absence inferences differ from whole-plant monitoring; the ontology was adapted accordingly.
- Integrated sources include USA-NPN, herbarium specimens, NEON and the Pan-European Phenology dataset.
- iNaturalist has about 110 million angiosperm observations (150,000 in one recent weekend) and 1.5–2 million human phenology annotations.
- Model uses masked autoencoder pre-training with threshold tuning and humans in the loop: about 98% for flowers and 95% for fruits (present), close to human performance; a separate ensemble model scores GBIF-imaged herbarium sheets.
- About 97% of Phenobase records are machine-annotated; new data fill many previously empty grid cells.
- AI-readiness requires per-record model URI (Hugging Face), per-record quality scores from family-level accuracy, traceable links to source records, data dictionaries, a portal plus Zenodo deposit, and clear data use and citation policies.
Notable moments
Transcript
Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.
Read the transcript ↓
Related talks.
BHL and the Planetary Knowledge Base: Structuring Biodiversity Literature for the Next Generation
NHM is re-OCRing all of BHL with modern models and extracting entities into a knowledge graph, with confidence gating to protect BHL data quality.
Machine-Ready Botanical Data: Mapping Integrated Collections and Traits with Darwin Core Data Package
Darwin Core Data Package can serve as a semantic integration layer for a botanical garden's diverse collections and trait data, and MBG's draft mapping is offered for community refinement.
Are plant image datasets ready for AI-enabled community assessment? A systematic review of bias, representativeness and annotation quality
Current plant image datasets are not yet ready for AI-enabled ecological monitoring, especially for spatially explicit tasks and annotation quality.