Species distribution model predictability doesn't always decline under
novel temperature conditions
Abstract
Despite the rapid development and application of species distribution
models (SDMs) to predict species responses to climate-driven ecosystem
changes, we have a limited understanding of model predictive performance
under novel environmental conditions. We aimed to address this gap using
a simulation experiment to evaluate how novel ecosystem conditions and
species movement influence SDM predictability. We leveraged observed sea
surface temperature responses in the California Current and Northeast
U.S. Shelf large marine ecosystems (LMEs) and prescribed
species-response curves to simulate the distribution of a resident but
mobile ectotherm, and a seasonally migrating ectotherm in each LME. For
each LME and species archetype, we fitted boosted regression tree SDMs
using data from 1985-2004 and then predicted the monthly probability of
presence from 2005-2020 and calculated the environmental novelty of
prediction month conditions. Generally, climate-driven ocean warming
resulted in increasing environmental novelty over time, though patterns
varied seasonally as warming caused novel conditions to increase over
time in the summer and fall and decrease in the winter and spring as
novel, cool conditions became more rare. Overall, predictive performance
declined as novelty increased and occurred before prediction conditions
became distinguishable from observation conditions. There were also
unexpected increases in performance under novel environmental conditions
when these novel conditions occurred at optimum species-response curve
temperatures. These results highlight that environmental novelty may not
always pose prediction challenges and will depend on where novel
conditions map onto species-response curves. As SDM applications expand,
there will be an ongoing need to maximize data quantity and quality to
more fully characterize a species’ fundamental niche, explore
environmental novelty relative to species-response curves, and continue
to improve methods for quantifying and communicating model uncertainty.
These efforts will open opportunities for model improvement and support
stakeholders’ capacity to understand and integrate predictions into
decision-making processes.