Learning from mistakes: Understanding and refining machine learning predictions of extinction risk for single‐country endemics
Emma S. Ericksen, Steven P. Bachman, Amy Barker, James S. Borrell, Andrew P. Budden, Shambel Alemu Chengere, Iain Darbyshire, Becca Davis, Kelda F. V. A. Elliott, Tesfanesh Feseha, Bezawit Genanaw, Hanny Lidetu, Lidet Mehari, Efrata Mekebib, Henry Miller, Timothy R. Pearce, Jack F. Plummer, Alexandra Roberts, Harry Smith, Emily Terry, Carolina Tovar, Jenna C. Willis, Sebsebe Demissew, Ermias Lulekal, Sileshi Nemomissa, Vida J. Svahnström, Eimear Nic Lughadha, Matilda J. M. BrownSocietal Impact Statement
As of 2026, fewer than 23% of vascular plants have International Union for the Conservation of Nature (IUCN) Red List assessments, despite accelerating threats to biodiversity. Predictive models of extinction risk complement assessments and help prioritize conservation action but face a trade‐off between data resolution and species coverage. Using Ethiopian endemic flowering plants, we found that medium‐scale spatial data—similar to provinces or counties—strike a balance between resolution and coverage, improving predictive performance. Because these data can be obtained from multiple sources, including textual records, our approach can be implemented in future prediction models, helping experts direct limited resources toward species most at risk.
Summary
Extinction risk predictions complement conservation assessments but are limited by available data, creating a trade‐off between resolution and taxonomic completeness. Fine‐scale predictors offer greater resolution but exclude data‐limited species, whereas coarse‐scale predictors maximize species coverage but lack resolution—especially at smaller scales. Recent comprehensive extinction risk assessment of Ethiopia's endemic angiosperms provides an opportunity to investigate the robustness of extinction risk predictions and identify patterns or systematic biases at a national scale. Here, we interrogate the performance of coarse‐scale predictions from Bachman et al. (2024) for Ethiopian endemics, using patterns in misclassifications to identify additional (protected area) and refined (geographic range size and human footprint) predictors that could improve future models. We found that number of floristic regions (similar in scale to provinces) significantly outperformed point‐based range size metrics (extent of occurrence and area of occupancy), and that recalculating human footprint at floristic region‐scale improved predictive performance. Protected area coverage did not improve model performance. While coordinates can be impossible to pinpoint for many records, administrative units can be obtained from textual data, unlocking more possibilities for georeferencing. By leveraging medium‐scale predictors, our approach balances resolution and taxonomic completeness, offering a scalable path to substantially improve global extinction risk predictions.