Mass Georeferencing Pipeline for Natural History Collections

Natural history collections hold millions of specimens whose only location information is a text-based locality description — something like "5 mi S of Springfield" — with no coordinates attached. That's a major barrier to using this data for biodiversity, climate, and conservation research, where geographic precision matters.

Existing tools for this problem, GEOLocate chief among them, have real gaps worth naming honestly: uncertainty estimates that don't clearly distinguish confidence levels, limited handling for historical or defunct place names, and little flexibility for institution-specific quirks in how locality data was originally recorded. This isn't a one-museum problem — it's a shared bottleneck across the entire field of digitized natural history collections.

The core of the project is a deterministic/semantic hybrid framework, and it rests on one plainly stated insight: LLMs alone tend to produce confident-sounding but unreliable uncertainty estimates. So the pipeline is built the other way around — deterministic gazetteer matching does the anchoring work, and LLM assistance is reserved for genuinely ambiguous cases, never used as a first resort.

Two principles follow from that:

• Every match is auditable. No black-box guessing — every resolved location traces back to a specific, inspectable decision.
• The system is built to fail visibly rather than silently. Ambiguity gets flagged for review; it's never quietly guessed at.

• Layered resolution — country → state/province → county → free-text locality, with each level fully and measurably resolved before the next depends on it.
• PostgreSQL/PostGIS-backed, with GeoNames as the primary gazetteer.
• Deduplication by locality string — a relatively small number of truly unique localities does the resolution work for a much larger number of specimen records, meaningfully cutting both compute cost and manual review burden.
• Historical vs. current administrative boundaries are handled as genuinely separate resolution paths — correctly resolving specimens collected in defunct provinces or counties without conflating them with their modern successors.

Real numbers, stated directly rather than hedged:

Fully deterministic for the vast majority of records — no AI dependency required to get there.

• Honest uncertainty representation — geographic uncertainty is expressed as a wedge, polygon, or donut shape depending on what's actually known, instead of collapsing everything into a single misleading point.
• An empirical comparison of full-LLM vs. hybrid resolution approaches, testing the core design premise directly rather than assuming it.
• A departmental tier analysis framework for prioritizing which collections departments to run through the pipeline next, and in what order.
• Darwin Core compliant output, so results integrate cleanly with the standards the biodiversity data community already relies on.

• A structured extraction system for free-text locality descriptions that tracks exactly which characters of the original text were interpreted, and how — preserving full traceability back to the source.
• Retroactive narrowing: evidence resolved at one geographic level can retroactively narrow ambiguity discovered at another, without ever forcing a guess.
• Cost-conscious, tiered design — cheap deterministic passes run first at every stage, with more expensive disambiguation reserved for whatever's left over.

The project includes an uncertainty explorer map and a review UI for manually checking ambiguous cases, plus departmental-level statistics views. Screenshots and links to come.

The Anthropic API is used narrowly, for bounded disambiguation only — not as a general-purpose part of the resolution pipeline.

• Smithsonian NMNH collections departments
• Data & Digitization Program Office (DPO)
• GeoLocate / Yale Peabody Museum

Accepted talk, FOSS4G-NA 2026 — "Honest Uncertainty: A Hybrid LLM Approach to Mass Georeferencing."

Actively expanding across additional collections departments. Free-text locality parsing — the hardest layer, handling compound descriptions and spatial-relationship language like "5 mi S of X" — is in active development.