Testing spatial transferability of species distribution models reveals differing habitat preferences for an endangered delphinid (Cephalorhynchus hectori) in Aotearoa, New Zealand

TL;DR

This study shows that habitat preferences of Hector's dolphins vary by region, so conservation models need local data to be accurate.

Contribution

The study demonstrates that a one-size-fits-all model for species distribution is not suitable for Hector's dolphins due to differing habitat preferences.

Findings

SDMs for two dolphin populations performed well locally but poorly when transferred to other areas.

The model for the Timaru population performed poorly both locally and when transferred.

Combining data from different areas improved model performance only when low and high density regions were combined.

Abstract

Species distribution models (SDMs) can be used to predict distributions in novel times or space (termed transferability) and fill knowledge gaps for areas that are data poor. In conservation, this can be used to determine the extent of spatial protection required. To understand how well a model transfers spatially, it needs to be independently tested, using data from novel habitats. Here, we test the transferability of SDMs for Hector's dolphin (Cephalorhynchus hectori), a culturally important (taonga) and endangered, coastal delphinid, endemic to Aotearoa New Zealand. We collected summer distribution data from three populations from 2021 to 2023. Using Generalised Additive Models, we built presence/absence SDMs for each population and validated the predictive ability of the top models (with TSS and AUC). Then, we tested the transferability of each top model by predicting the distribution of the remaining two populations. SDMs for two populations showed useful performance within their respective areas (Banks Peninsula and Otago), but when used to predict the two areas outside the models' source data, performance declined markedly. SDMs from the third area (Timaru) performed poorly, both for prediction within the source area and when transferred spatially. When data for model building were combined from two areas, results were mixed. Model interpolation was better when presence/absence data from Otago, an area of low density, were combined with data from areas of higher density, but was otherwise poor. The overall poor transferability of SDMs suggests that habitat preferences of Hector's dolphins vary between areas. For these dolphins, population‐specific distribution data should be used for conservation planning. More generally, we demonstrate that a one model fits all approach is not always suitable. When SDMs are used to predict distribution in data‐poor areas an assessment of performance in the new habitat is required, and results should be interpreted with caution.

Species distribution models (SDMs) are commonly used to predict habitat and distribution for species in data poor areas, under the assumption that habitat use is the same across space. We tested the transferability of SDMs for Hector's dolphin, revealing differences in habitat use between populations. These results indicate that to understand habitat use of Hector's dolphin, local data is required for the model building process.

Figures