Limited transferability of stream‐fish distribution models among river catchments: reasons and implications

Abstract

Summary Spatial transfers of species distribution models (SDMs) are often applied in the study of land use and climate‐change effects, spread of invasive species and conservation planning. However, model transferability and risk of error are rarely evaluated prior to predicting species distribution to different regions. We aim to assess spatial transferability of SDMs for stream fish and to evaluate the effect of model types and habitat heterogeneity on transferability. We developed SDMs for 21 fish species and made predictions of occurrence of these species among five pairs of river catchments (i.e. model transfers). Forty eight transfers were made for each of three modelling approaches (Lasso‐regularised logistic regression (LLR), boosted regression trees (BRT) and MaxEnt) after incorporating spatial autocorrelation. In addition to internal and external evaluation of discrimination power, we assessed the cross‐catchment consistency of variable selections, fish‐habitat relationships and predicted probabilities of species presences. Approximately half of 144 spatial transfers of SDMs had moderate to high discrimination power. Discrimination power was low for the rest of the models. Incorporating spatial autocorrelation could not improve discrimination power in the model transfers. Friedman test showed that BRT did not differ significantly from LLR but it outperformed MaxEnt in terms of AUC in the model transfers. BRT and LLR models tended to have high overall accuracy and specificity, whereas MaxEnt tended to have high sensitivity in the model transfers. The degree of model transferability varied among species, and was asymmetric when reciprocal transfers were made between paired catchments. Ranks of variable importance in BRT models differed among catchments for most species. Temperature, base flow index, altitude and habitat condition index were more often ranked as the most important predictors in the five study river catchments, although the functional forms of their effects on fish presence were sometimes inconsistent between paired catchments. Compared to published transferability of some terrestrial taxa SDMs, spatial transferability of stream‐fish distribution models was limited, reflecting the natural barriers to dispersal among catchments, and the necessity of evaluating model transferability in conservation applications (e.g. evaluation of climate‐ or landscape‐change effects, invasive species risk assessments and species reintroduction planning). We suggest the following strategies to enhance spatial transferability: (i) match the range and location of the habitat predictors between the model region and prediction region, or alternatively choose a model training region with a large extent and a wide range of environmental gradients; (ii) use presence–absence models such as BRT and LLR and (iii) include habitat features with a sound ecological basis such as temperature and hydrology.