The skipjack tuna fishery in the west-central Pacific Ocean: applying neural networks to detect habitat preferences

Spatial models for habitat selection were developed using neural networks. The model specifications were elucidated from model construction, training, validating, testing, and interpretation, and applied to skipjack tuna in the west-central Pacific Ocean. The model was created using commercial data from the Oceanic Fisheries Programme of the South Pacific Fisheries Commission and oceanic environmental data include sea surface temperature, horizontal gradient of sea surface temperature calculated from sea surface temperature, sea surface height, and chlorophyll-a. Local abundance indices for skipjack tuna were compiled using catch per unit effort, catch or effort. The optimal neural network models for each abundance index were selected by mean square errors and average relative variances. The predictive ability for optimal neural network models was evaluated by the R² value using a cross-validation approach. The accuracy and stability of the optimal models, the contribution of independent variables, and the distribution of spatial sensitivity analyses were shown to vary with the abundance index chosen as the response variable. Chlorophyll-a was the most significant oceanographic factor in habitat selection. These results improve our understanding of how best to apply neural networks for modeling habitat selection by skipjack tuna.