Temporal transferability of stream fish distribution models: can uncalibrated SDMs predict distribution shifts over time?

Aim We aim to assess the temporary transferability of species distribution models (SDMs) for stream fish species in terms of discrimination power and calibration. Location New River basin, eastern United States. Methods In this study, we used Lasso-regularized logistic regression (LLR), boosted regression trees (BRT), MaxEnt, and ensemble model (ENS) to evaluate the habitat suitability of 16 fish species with different rarity and temperature preference based on historical species occurrences obtained during 1950-1990. These SDMs were used to make probabilistic predictions of species presence in the independent datasets sampled during 2012-2014. We evaluated the temporal transferability of these SDMs in terms of discrimination power and calibration with the temporarily independent datasets. Results The area under the receiver-operating-characteristic curve (AUC) was over 0.6 for 13 (81%) of the species in the evaluations of ENS models with the independent datasets. Cool-water species and species with small local population size traits tended to have good temporal transferability. With observed species prevalence as the discrimination cut-off, LLR had the highest overall accuracy for 13 of the 16 species and highest specificity for 10 species, whereas MaxEnt had the highest sensitivity for 14 species. Biases, under-or over-fitting problems were common in the temporal model transfers, among the modelling approaches used here. Main conclusions SDMs developed with historical data generally had moderate to good discrimination power but they tended to systematically underestimate current probability of species presence. To predict species distribution shifts over time, SDMs should be well-calibrated with high discrimination power. We suggest reclassifying predicted probability of species occurrence to ordinal ranks to deal with (under-and over-estimation) bias, and fine-tuning variable selection with regularization or cross validation to remedy under-and over-fitting.