Finding what you don’t know: testing SDMs methods for poorly-known species

Aim: A limitation of species distribution models (SDMs) is that species
with low sample sizes are difficult to model. Yet it is often important to
know the habitat associations of poorly known species to guide
conservation efforts. Techniques have been proposed for modeling species’
distributions from few records, but their performance relative to one
another has not been compared. Because these models are built and
evaluated with small datasets, sampling error could cause severely biased
sampling in environmental space. As a result, SDMs are likely to
underpredict geographic distributions given small sample sizes. We perform
the first comparison of methods explicitly promoted or developed for
predicting the geographic ranges of species with very low sample sizes.
Location: North Carolina, USA Taxon: South Mountains Gray-cheeked
Salamander (Plethodon meridianus) Methods: Using the sparse, existing
georeferenced records of P. meridianus, we built SDMs using a range of
methods that previous researchers have argued should work for low sample
sizes. We then tested each SDM’s ability to accurately predict independent
survey data that were not georeferenced prior to our study. We compared
SDMs using omission error and AUC. Results: Roughly half of the models
successfully predicted survey records in the range center, and all models
had high omission error rates in the range exterior. In the range interior
or exterior, the ‘ensemble of small models’ technique produced SDMs with
high omission error rates. Spatial filtering had negligible impact on
model performance. Most, but not all, models outperformed predictions
using distance from known populations. Using one of the best-performing
methods, we developed an improved range map of P. meridianus. Main
Conclusions: Geographically peripheral populations were difficult to
predict for all SDMs, though some methods were clearly inferior for our
dataset. We recommend that when sample sizes are low, researchers use
Maxent with species-specific model settings.