Correcting and georeferencing natural history occurrence records with calibrated uncertainty

Abstract

1. Occurrence records underpin biodiversity analyses, but their coordinates are often erroneous or missing. Misplaced coordinates distort environmental predictors and range estimates used in conservation assessment. Existing tools flag suspect records or georeference locality text, but rarely correct coordinates, provide calibrated uncertainty, or quantify downstream consequences. 2. We present GeoMend, a deterministic, rule-based method that flags and corrects suspect coordinates and georeferences records lacking coordinates from locality text, assigning each a precision tier or uncertainty radius. It uses no machine learning or artificial intelligence. We evaluate it against expert coordinates for 16,913 records and 736 species on Hispaniola, compare georeferencers using 5,000 localities from 72 countries, and measure changes in environmental predictors and the extent of occurrence. 3. On held-out Hispaniola records, GeoMend flags 92% of coordinates more than 5 km from expert truth, versus 27% for CoordinateCleaner. Among the accepted corrections with paired coordinates, the median error falls from 2.19 to 0.82 km, and 68% of the complete held-out pairs improve. On records placed by both GeoMend and GEOLocate that also have original coordinates, the median error is 0.67 km for GeoMend versus 10.69 km for GEOLocate. For records lacking coordinates, evaluated uncertainty radii contain expert truth about 92% of the time in well-supported gazetteer classes. Correction reduces the mean standardized environmental distance from expert truth by 38% and moves the extent of occurrence toward expert truth by a median of 8% (mean of 79%). Corrections and newly georeferenced records give 17 Hispaniola species their first range estimate. 4. On the global benchmark, GeoMend is more accurate than GPT-4o for two-thirds of the records placed by all four comparators, with median errors of 4.4 and 10.8 km, respectively. Applied to Global Biodiversity Information Facility (GBIF) records lacking coordinates, it places records from 82,180 previously unmapped species across 233 countries, ranging from precise locality interpretations to administrative centroids. Automated placement makes the global backlog more approachable, while explicit tiers distinguish uses supported by precise and coarse coordinates. Further georeferencing requires regional expertise to combine automated placement with local knowledge.