From PDFs to places: A semi-automated, auditable workflow to map study localities for geoheritage inventories

Abstract

The scientific relevance of geological sites can be reflected in the intensity and diversity of research conducted there and in the publications that document those places. For geoconservation, however, this literature signal is useful only when the underlying localities can be identified, mapped, and checked against source evidence. Locality evidence relevant to geoconservation is often dispersed across heterogeneous portable document format (PDF) articles and is rarely available as geocoded, source-linked data. We present a semi-automated workflow that converts geoscience PDFs into a map-ready layer of candidate terrestrial study localities for expert review. The workflow uses a large language model (LLM) to extract candidate localities, assigns an article-specific salience score, filters marginal mentions, resolves retained place names to coordinates, and preserves links to supporting passages. We evaluated the workflow on a manually curated benchmark of 99 articles and 241 terrestrial article-locality records. The archived reference workflow output recovered 135 reference localities, produced 216 additional candidates not accepted as reference localities, and missed 106 reference localities, giving precision 0.385, recall 0.560, and 𝑭𝑭1 = 0.456. A baseline without LLMs achieved precision 0.082, recall 0.174, and 𝑭𝑭1 = 0.112. A repeated-run check under a fixed stability-check configuration on a stratified subset of 20 articles showed measurable variability, with mean 𝑭𝑭1 = 0.359 and standard deviation 0.042. The method is intended as an auditable screening tool for finding, inspecting, and prioritising literature-derived locality evidence before formal geoheritage assessment.