Assessing predictive performance and transferability of species distribution models for freshwater fish in the United States

Rigorous modeling of the spatial species distributions is critical in biogeography, conservation, resource management, and assessment of climate change. The goal of chapter 2 of this dissertation was to evaluate the potential of using historical samples to develop high-resolution species distribution models (SDMs) of stream fishes of the United States. I explored the spatial transferability and temporal transferability of stream–fish distribution models in chapter 3 and chapter 4 respectively. Chapter 2 showed that the discrimination power of SDMs for 76 non-game fish species depended on data quality, species' rarity, statistical modeling technique, and incorporation of spatial autocorrelation. The area under the Receiver-Operating-Characteristic curve (AUC) in the cross validation tended to be higher in the logistic regression and boosted regression trees (BRT) than the presence-only MaxEnt models. AUC in the cross validation was also higher for species with large geographic ranges and small local populations. Species prevalence affected discrimination power in the model training but not in the validation. In chapter 3, spatial transferability of SDMs was low for over 70% of the 21 species examined. Only 24% of logistic regression, 12% of BRT, and 16% of MaxEnt had AUC > 0.6 in the spatial transfers. Friedman's rank sum test showed that there was no significant difference in the performance of the three modeling techniques. Spatial transferability could be improved by using spatial logistic regression under Lasso regularization in the training of SDMs and by matching the range and location of predictor variables between training and transfer regions. In chapter 4, testing of temporal SDM transfer on independent samples resulted in discrimination power of the moderate to good range, with AUC > 0.6 for 80% of species in all three types of models. Most cool water species had good temporal transferability. However, biases and misspecified spread occurred frequently in the temporal model transfers. To reduce under- or over-estimation bias, I suggest rescaling the predicted probability of species presence to ordinal ranks. To mitigate inappropriate spread of predictions in the climate change scenarios, I recommended to use large training datasets with good coverage of environmental gradients, and fine-tune predictor variables with regularization and cross validation.