Chub Mackerel (Scomber japonicus) is an important pelagic fishery resource in the Northwest Pacific, and its fishing-ground distribution is strongly influenced by dynamic marine environmental conditions. This study aimed to evaluate how environmental information at different temporal scales affects the prediction of high-catch fishing grounds and to identify environmental-variable contributions. Fishery logbook data from Chinese light purse seine vessels during 2014–2022 were combined with marine environmental variables to construct four feature sets: instantaneous features (E1), multi-temporal-scale fusion features (E2), short-term features with 7-day rolling means (E3), and long-term features with 30-day rolling means (E4). Deep Forest, random forest, XGBoost, LightGBM, and CatBoost were evaluated using nested spatial group cross-validation, and SHapley Additive exPlanations (SHAP) was applied to interpret model predictions. The results showed that, after historical environmental information was added, AUC values increased for most models, and the multi-temporal-scale fusion features performed better in metrics related to high-catch sample identification; therefore, the hypothesis proposed in this study was supported in the overall trend. Model comparisons showed that Deep Forest performed relatively stably under E2, E3, and E4, whereas RF performed relatively well under E1. Short-term environmental features helped improve overall fishing-ground discrimination, whereas multi-temporal-scale fusion was more favorable for identifying high-catch samples. Time-lag correlation and SHAP analyses indicated that short-term environmental changes, longer-term background conditions, and seasonal signals jointly provided information for model prediction. This study may provide a reference for real-time fishing-ground prediction and fishery management.
Zhang et al. (Sun,) studied this question.