Evidence-based guidelines for automated conservation assessments of plant species

Assessing species' extinction risk is vital to setting conservation priorities. However, assessment endeavors, such as those used to produce the IUCN Red List of Threatened Species, have significant gaps in taxonomic coverage. Automated assessment (AA) methods are gaining popularity to fill these gaps. Choices made in developing, using, and reporting results of AA methods could hinder their successful adoption or lead to poor allocation of conservation resources. We explored how choice of data cleaning type and level, taxonomic group, training sample, and automation method affect performance of threat status predictions for plant species. We used occurrences from the Global Biodiversity Information Facility (GBIF) to generate assessments for species in 3 taxonomic groups based on 6 different occurrence-based AA methods. We measured each method's performance and coverage following increasingly stringent occurrence cleaning. Automatically cleaned data from GBIF performed comparably to occurrence records cleaned manually by experts. However, all types of data cleaning limited the coverage of AAs. Overall, machine-learning-based methods performed well across taxa, even with minimal data cleaning. Results suggest a machine-learning-based method applied to minimally cleaned data offers the best compromise between performance and species coverage. However, optimal data cleaning, training sample, and automation methods depend on the study group, intended applications, and expertise.