Transferability of predictive models of coral reef fish species richness

Understanding biodiversity patterns depends on data collection, which in marine environments can be prohibitively expensive. Transferable predictive models could therefore provide time- and cost-effective tools for understanding biodiversity-environment relationships. We used fish species counts and spatial and environmental predictors to develop predictive models of fish species richness (S) for two major coral reefs located in separate ocean basins: Australia's Great Barrier Reef (GBR; Queensland) and Ningaloo Reef (NR; Western Australia). We tested the ability of the GBR model to predict S at NR (its transferability) under various scenarios using different sampling durations, years sampled and transect sizes. Based on R², the GBR model poorly predicted S at NR (R² < 16%) with few predicted values strongly correlated with observations. However, comparable spatial patterns in S across NR were predicted by both the NR and the GBR models when calibrated at similar spatio-temporal scales. This result suggests that poor validation of the transferred models may indicate low deviance explained by the predictors in the new system (where other predictors not included might have a more direct effect on the response) and that in some situations, model transferability may be considerably improved by using data sets of similar spatio-temporal scales. Therefore, data filtering by time and space may be required prior to transferring models. Policy implications. Transferable models can provide initial estimates of fish species richness patterns in poorly sampled systems, and thereby guide the design of better and more efficient sampling programs. Further improvements in model transferability will increase their predictive power and utility in conservation planning and management.