Global Conservation Prioritisation Approach Provides Credible Results at a Regional Scale

Abstract

Aim

Conservationists and managers must direct resources and enact measures to protect species, despite uncertainty about their present status. One approach to covering the data gap is borrowing information from data-rich species or populations to guide decisions about data-poor ones via machine learning. Recent efforts demonstrated proof-of-concept at the global scale, leaving unclear whether similar approaches are feasible at the local and regional scales at which conservation actions most typically occur. To address this gap, we tested a global-scale predictive approach at a regional scale, using two groups of taxa.

Location

State of Maryland, USA.

Taxa

Vascular land plants and lepidopterans.

Methods

Using publicly available occurrence and biogeographic data, we trained random forest classifiers to predict the state-level conservation status of species in each of the two focal taxa. We assessed model performance with cross-validation, and explored trends in the predictions.

Results

Our models had strong discriminatory ability, accurately predicting status for species with existing status assessments. They predict that the northwestern part of Maryland, USA, which overlaps the Appalachian Mountains, harbours a higher concentration of unassessed, but likely threatened plants and lepidopterans. Our predictions track known biogeographic patterns, and unassessed species predicted as most likely threatened in Maryland were often recognised as also needing conservation in nearby jurisdictions, providing external validation to our results.

Main Conclusions

We demonstrate that a modelling approach developed for global analysis can be downscaled and credible when applied at a regional scale that is smaller than typical species ranges. We identified select unassessed plant and lepidopteran species, and the western, montane region of Maryland as priority targets for additional monitoring, assessment and conservation. By rapidly aggregating disparate data and integrating information across taxa, models like those we used can complement traditional assessment tools and assist in prioritisation for formal assessments, as well as protection.