Modeling Optimal Habitat for Stream Fish Species with Derived Historical Metacommunity Samples

Rigorous modeling of species-habitat relationships and the spatial pattern of species distributions is critical in conservation and resource management. Such models have wide applications in invasion risk assessments, projection of climate change effects, bioassessment, and general conservation planning for freshwater ecosystems. Ideally, species distribution models (SDM) require species presence and absence data for adequate estimation of model parameters and derivation of classification rules for decision-making. However, lack of observations of true species absences has characterized many studies that have generated data now used in SDM, leading to a growing pursuit of models based on presence-only data. It is gradually being realized that presence-only models yield less accurate predictions of species-habitat associations and distributions, and are also difficult to evaluate for accuracy via both threshold-dependent and -independent measures. In this study, we propose and test a SDM framework using historical presence observations of fish species recorded in atlases, and absences inferred from locations where historical presences have been recorded for non-game fish species.  We used the confluence-to-confluence stream segment at 1:100,000 scale of the NHD+ database as the spatial habitat unit. Species presences were accumulated in each habitat unit for all available, readily accessible atlas records up to approximately 1980. Members in a habitat unit are more appropriately described as a metacommunity sample because while species present in the unit belong to the same regional pool, they may not all have co-existed in that unit at any point in time. To test this framework, hierarchical Bayesian models were used to examine the effects of climatic and landscape-scale habitat factors on the distribution of two endemic fish species, Notropis scabriceps, and Percina gymnocephala, in the New River Basin of southeastern US. Models were validated with independent samples, and field-testing has been planned based on predicted optimal habitat not previously sampled. From preliminary results, the Bayesian models using presence/absence are preferred to Genetic Algorithm for Rule-set Production (GARP) and Maximum Entropy (MaxEnt) in terms of prediction errors and area under the ROC curve (AUC). By combining various sources of public data, the metacommunity-based SDM framework broadens our capability to model fish distributions by innovatively removing the constraint of lack of absence data.