Prediction of Large Whale Distributions: A Comparison of Presence–Absence and Presence-Only Modeling Techniques

Species distribution models that predict species occurrence or density by quantifying relationships with environmental variables are used for a variety of scientific investigations and management applications. For endangered species, such as large whales, models help to understand the ecological factors influencing variability in distributions and to assess potential risk from shipping, fishing, and other human activities. Systematic surveys record species presence and absence, as well as the associated search effort, but are very expensive. Presence-only data consisting only of sightings can increase sample size, but may be biased in both geographical and niche space. We built generalized additive models (GAM) using presence-absence sightings data and maximum entropy models (Maxent) using the same presence-absence sightings data, and also using presence-only sightings data, for four large whale species in the eastern tropical Pacific Ocean: humpback (Megaptera novaeangliae), blue (Balaenoptera musculus), Bryde’s (Balaenoptera edeni), and sperm whales (Physeter macrocephalus). Environmental variables were surface temperature, surface salinity, thermocline depth, stratification index, and seafloor depth. We compared predicted distributions and environmental response curves from each of the two model types. Maxent and GAM model predictions based on systematic survey data are very similar, when Maxent absences are selected from the survey trackline data. However, we show that spatial bias in presence-only Maxent predictions can be caused by using pseudo-absences instead of observed absences and by the sampling biases of both opportunistic data and stratified systematic survey data with uneven coverage between strata. Predictions of uncommon large whale distributions from Maxent or other presence-only techniques may be useful for science or management, but only if spatial bias in the observations is addressed in the derivation and interpretation of model predictions.