Improving Species Distribution Models of Continental Shelf Fishes off California

Species distribution models (SDMs) relate species occurrences, or abundances, with oceanographic or bathymetric data to predict species distributions. SDMs are used to forecast the effects of climate change on species distributions, evaluate existing protected areas and identify locations for future protected areas, estimate the effects of invasive species on a community, and determine how community assemblages change spatially or temporally. Articles describing marine SDMs are becoming increasingly common in the scientific literature; however, their efficacy for predicting the distribution and relative abundance of continental shelf fishes off California is not universally accepted. SDM performance can be affected by the inclusion of seascape variables, different spatial resolutions, the number of years of data, and the choice of statistical model. I examined how SDM performances changed with inclusion of these other factors. I combined multibeam bathymetry maps with remotely operated vehicle surveys to obtain remotely-sensed habitat data and the abundances and presence/absence of four continental shelf fishes: Blue Rockfish (Sebastes mystinus), Gopher Rockfish (Sebastes carnatus), Lingcod (Ophiodon elongatus), and Vermilion Rockfish (Sebastes miniatus). Using this information, I examined the strength of species-habitat associations and the spatial extent to which different SDM models were able to accurately predict presence/absence. Species-habitat associations for each species were determined from generalized linear models (GLMs) at three widely-spaced locations: a Northern California location (Bodega Bay or the Farallon Islands), a Central location (Point Sur or Point Buchon), and a Southern location (Harris Point). Model performance was evaluated using information about the overall accuracy, area under the curve (AUC), and Cohen’s Kappa statistics. Seascape characteristics (habitat patch area, perimeter:area ratio, and distance to habitat edge) at Point Sur were used in GLMs to determine if those variables increased model performance. Remotely-sensed habitat variables were created at three different resolutions (2, 5, and 10 m) at Bodega Bay and used in GLMs to evaluate the effect of spatial resolution on model performance. The effect of the number of years of data on model performance was evaluated at Harris Point, with GLMs and independent years of data. Additionally, at Harris Point, the choice of statistical model on model performance was evaluated with GLMs, generalized additive models, random forest, and boosted regression tree models. A coastwide model that optimized model performance was generated and tested against an independent site, Point Lobos. Lastly, an optimized model was created at Point Sur and tested against the independent sites of Point Lobos and Big Creek. Each species exhibited differences in habitat associations among the three regions. For example, Blue Rockfish distributions were positively associated with eastness at the Farallon Islands and Harris Point but negatively associated with eastness at Point Sur. Additionally, Gopher Rockfish distributions were positively associated with broad-scale BPI at each of the three locations, but exhibited differences in associations with other metrics, such as vector ruggedness measure and eastness. The inclusion of seascape models and the different spatial resolutions either decreased or did not affect model performance. In general, model performance increased with more years of data, but was dependent on the number of observations in a test year. Furthermore, model performance increased with machine learning methods such as random forest or boosted regression tree models. However, the coastwide models that were tested at Point Lobos exhibited poor performance metrics when predicting presence/absence, except for Lingcod. Additionally, models created at Point Sur and tested at Point Lobos or Big Creek were only successful for Lingcod at Point Lobos and Gopher Rockfish at Big Creek. These results suggest that SDMs for continental shelf fishes can be improved through different methods, but their efficacy in predicting presence/absence in unsurveyed areas of the coast remains poor