Abstract
Yellowfin tuna (
Thunnus albacares
) is a highly migratory and economically important species in the Pacific Ocean, and its spatial distribution is closely associated with marine environmental conditions. To investigate the environmental associations underlying the spatial patterns of yellowfin tuna nominal catch per unit effort (CPUE), quarterly spatial prediction models were developed using Pacific longline fishery data from 2004 to 2023 together with multi-source marine environmental datasets. Candidate environmental variables were first screened using permutation importance derived from Random Forest and XGBoost models in combination with variance inflation factor (VIF) analysis to identify the key environmental drivers. Multiple models, including machine learning techniques (Random Forest, XGBoost, and graph neural networks), spatial statistical models (NNGP), and geographic neural networks (GeoSpaNN), were then compared. The findings indicated that: (1) sea surface temperature (temp0), temperature at 150 m depth (temp150), velocity component at 5 m depth (v5), mixed layer depth (mld), sea surface height anomaly (sla), dissolved oxygen (do), and chlorophyll concentration (chl) are seven important environmental variables; (2) GeoSpaNN achieved overall test-set values of 0.86 for R², 0.66 for root mean squared error(RMSE), and 0.40 for mean absolute error (MAE), indicating consistent predictive performance across quarters. In addition, it more accurately reproduced the observed spatial distribution patterns; (3) Environmental interpretation of the model results indicated that temperature variables were the most important environmental factors associated with yellowfin tuna CPUE with the strongest effects observed in the western Pacific warm pool and its extension into the central Pacific. Factors such as do, mld, sla, chl, and flow velocity predominantly exert a regulatory influence in certain marine regions and during seasonal variations. Research indicates that GeoSpaNN can concurrently delineate the nonlinear impacts and spatial dependence structure of environmental variables.