TDWG2026
Home / Talks / SAL A
talk · Tuesday 22 September · 11:51–⁠12:03 · SAL A

AI-assisted georeferencing a posteriori of herbarium specimens: a case study on Italian herbaria

Matteo Conti · AI for Biodiversity Data

Recording time 3:23:17–3:34:12Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionCombining LLM parsing and LLM judging with gazetteers more than doubles correct georeferences for historical herbarium localities compared with Google Maps, even with a local model.

Overview

What this was about

Matteo Conti described a workflow for retrospective georeferencing of specimens from a national project digitising 12 large Italian herbaria (about four million specimens). Standard tools performed poorly on historical, often Latin, handwritten localities (Nominatim georeferenced under 10% and Google Maps about 50%). The six-step workflow uses an LLM parser, with dynamically injected few-shot examples matched by language and spatial relation, to extract toponyms, habitats, elevations and relations; matches toponyms against three gazetteers; ranks and filters candidates; uses an LLM judge to accept, flag for review or reject; and adjusts coordinates and uncertainty for relational localities. Run locally with Gemma it produced more than double Google Maps' correct georeferences with fewer wrong ones, and Gemma and Gemini Flash Lite flagged all coarser and wrong results for human review.

Why it matters. Coordinates are essential for analyses like species distribution models, and historical collections are hard to georeference with standard geocoders.

Key ideas

In the room

  • Italian project phases: imaging (Picturae), transcription, cleaning and Darwin Core alignment, publication.
  • About 20% of records are in Latin with historical toponyms and handwritten labels causing transcription errors.
  • Prompt: system instructions, dynamically inserted examples by pattern, then batches of localities.
  • Up to 10 candidates per toponym are retrieved and ranked with threshold filtering.
  • LLM judge returns accepted match, uncertain match flagged for review, or rejection with an explanation.
  • Results categorised as correct, coarser or wrong.
  • Gemini Flash Lite performed well at half the cost of Gemini Flash; Gemma 4 runs locally at almost no cost.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
talk · 25 Sept · 11:30–⁠11:42 · SAL A

Enhancing Biodiversity Georeferencing Using Large Language Models and Knowledge Graphs

LLM preprocessing plus knowledge-base matching fills in locality hierarchies and proposes coordinates, with a review tool to keep humans in control.

Roselyn Gabud
talk · 22 Sept · 11:27–⁠11:39 · SAL A

HerbAudit: Validating AI-Driven Herbarium Transcriptions

HerbAudit shows AI herbarium transcription reaching ~94% accuracy and provides a fair, field-aware way to benchmark it.

Dilara Ağacık
talk · 22 Sept · 12:00–⁠12:15 · SAL B

Geonomia: Metadata and Community for Georeferencing

Clustering specimens into collecting trips makes georeferencing and AI enrichment more efficient and should be done as a community effort.

Nicky Nicolson
Programme abstract

AI-assisted georeferentiation a-posteriori of herbarium specimens: a case study on Italian herbaria

Matteo Conti

Italy’s largest herbaria have recently undergone a large-scale digitization, making more than 4 million specimens available online through the national portal PlantData and the GBIF. Most specimens only had textual locality descriptions without geographic coordinates, limiting their use in many ecological analyses such as species distribution models. To address this limitation, we started a process of georeferentiation a-posteriori.

The data presented different challenges: locality labels were often in Latin or partially illegible, mixing locality names with habitat notes, altitude and collector annotations. Initial georeferencing approaches using national gazetteers and OpenStreetMap Nominatim achieved success rates below 10%, as geocoding services were frequently confused by excessive contextual details in locality strings. Workflows based on Google Maps APIs improved results (~50%) but remained highly sensitive to spelling errors and historical orthographic variants.

Thus, we developed pipelines that combine lightweight large language models (LLMs) with geocoding services. The models use the full label context to separate locality names from ancillary text and correct spelling inconsistencies before georeferencing. Small LLMs combined with Nominatim achieved ~71% successful georeferentiations at very low cost. Small LLMs paired with Google Maps improved this result to ~78%, while using Frontier LLMs reached up to 84% accuracy, but they remain too expensive for high volume data.

This approach shows that AI can assist and accelerate the process of georeferencing a-posteriori herbarium specimens. Small LLMs can achieve strong performance and, along with Nominatim, can be fully self-hosted, reducing reliance on external APIs and services.

Submitted to the conference programme before the talk, so the talk itself may differ.