Presence-only Species Distribution Modeling for King Mackerel (Scomberomorus cavalla) and its 31 Prey Species in the Gulf of Mexico

Ecosystem based fisheries management has been broadly recognized throughout the world as a way to achieve better conservation. Therefore, there is a strong need for mapping of multi-species interactions or spatial distributions. Species distribution models are widely applied since information regarding the presence of species is usually only available for limited locations due to the high cost of fisheries surveys. Instead of regular presence and absence records, a large proportion of the fisheries survey data have only presence records. This makes the modeling problem one of one-class classification (presence only), which is much more complex than the regular two-class classification (presence/absence). In this study, four different presence-only species distribution algorithms (Bioclim, Domain, Mahal and Maxent) were applied using 13 environmental parameters (e.g., depth, DO, bottom types) as predictors to model the distribution of king mackerel (Scomberomorus cavalla) and its 31 prey species in the Gulf of Mexico (a total of 13625 georeferenced presence records from OBIS and GBIF were used). Five-fold cross validations were applied for each of the 128 (4 algorithms × 32 species) models. Area under curve (AUC) and correlation coefficient (R) were used to evaluate the model performances. The AUC of the models based on these four algorithms were 0.83±0.14, 0.77±0.16, 0.94±0.06 and 0.94±0.06, respectively; while R for the models were 0.47±0.27, 0.43±0.24, 0.27±0.16 and 0.76±0.16, respectively. Post hoc with Tukey’s test showed that AUC for the Maxent-based models were significantly (p<0.05) higher than those for Bioclim and Domain based models, but insignificantly different from those for Mahal-based models (p=0.955); while R for the Maxent-based models were significantly higher than those for all the other three types of models (p<0.05). Thus, we concluded that the Maxent-based models had the best performance. High AUC and R also indicated that Maxent-based models could provide robust and reliable results to model target species distributions, and they can be further used to model the king mackerel food web in the Gulf of Mexico to help managers better manage related fisheries resources.