Transferability of predictive fish distribution models in two coastal systems

Species distribution modelling has emerged as a tool both for exploring niche theory and for producing distribution maps for management. To understand and predict potential effects of large scale habitat change there is a need for proper model validation and applicability also in unstudied areas. However, knowledge about factors influencing the transferability of distribution models, i.e. the accuracy of the models when applying them in a new geographical area, is limited. We have successfully modelled the larval distribution of two fish species, northern pike (Esox lucius L.) and roach (Rutilus rutilus L.), on a regional scale in the Baltic Sea using a few and easily measured environmental variables. When models were transferred from the training area to the testing area the models showed reasonable to very good discrimination (ROC 0.75 and 0.93) based on external validation using independent data separated also in time (1–2 years). The predicted larval distribution also overlapped with the distribution of young-of-the-year fish later in the season. Performance when reversing the transfer, by constructing the models in the testing area and predicting back to the original training area, was less successful. This discrepancy was species-specific and could be explained by differences in the species presence ranges along the predictor variables in the testing area compared to the training area. Our results illustrate how transferability success can be influenced by area-specific differences in the range of the predictor variables and show the necessity of validating model predictions properly.