Barnaby E. Walker, Tarciso C.C. Leão, Steven P. Bachman, Eve Lucas, Eimear Nic Lughadha

Assessing species’ extinction risk is vital to setting conservation priorities. However,
assessment endeavours like the IUCN Red List of Threatened Species have significant gaps
in coverage of some taxonomic groups. Automated assessment (AA) methods are gaining
popularity to fill these gaps, leveraging improvements in computing and digitally-available
information. Choices made in developing, using, and reporting AA methods could hinder
successful adoption or lead to poor allocation of conservation resources.
We explored how choice of data-cleaning, taxonomic group, training sample, and
automation method affected performance of threat status predictions for groups of plant
species. We used occurrence records from GBIF to generate assessments for species in
three taxonomic groups using six different occurrence-based AA methods. We measured
each method’s performance and coverage after applying increasingly stringent cleaning to
occurrence data.
Automatically cleaned data from GBIF yielded comparable performance to occurrence
records cleaned manually by experts. However, all types of data-cleaning removed species
and limited the coverage of automated assessments. Overall, machine-learning-based
methods performed well on all taxonomic groups, even with minimal data-cleaning.
Results suggest a machine-learning-based method on minimally cleaned data offers the
best compromise between performance and species coverage. However, optimal datacleaning, training sample, and automation methods depend on the study group, intended
applications and expertise. We recommend evaluating new AA methods across multiple
groups and providing additional context with extinction risk predictions so users can make
informed decisions.