Table 1_Predicting Pacific saury fishing sites using machine learning and spatial environmental variables reflecting recent eastward shifts.docx

In recent years, the Northwest Pacific has seen a decline in Pacific saury (Cololabis saira) catch and an eastward shift of fishing grounds, both of which have posed increasing challenges for effective resource management. To identify environmental drivers underlying the formation of Pacific saury fishing grounds, we developed machine learning-based prediction models using spatial environmental variables. Our models combined fishing site and pseudo-absence data with high-resolution oceanographic data from the Japan Fisheries Research and Education Agency Regional Ocean Modeling System (FRA-ROMS). We employed three machine learning methods to evaluate three types of explanatory variable representations: averaged, vectorized, and spatially structured. The results demonstrated that preserving spatial structure using a two-dimensional grid layout improved model performance. Our prediction results reflected the recent eastward shifting fishing grounds, suggesting a strong influence of environmental factors, particularly water temperature derived from the ocean circulation model. The convolutional neural network model, which best replicated the eastward shift of fishing sites, achieved a recall of 45.0% and a precision of 95.4%, although its performance declined under higher environmental novelty, which was associated with low-catch years (2020-2022). By evaluating how different spatial representations of environmental variables affect model performance, this study demonstrates that incorporating spatial structure improves predictive ability and enables models to capture recent eastward shifts in fishing activity under changing ocean conditions.