Revealing the effects of environmental and spatio-temporal variables on changes in Japanese sardine (Sardinops melanostictus) high abundance fishing grounds based on interpretable machine learning approach

The construction of accurate and interpretable predictive model for high abundance fishing ground is conducive to better sustainable fisheries production and carbon reduction. This article used refined statistical maps to visualize the spatial and temporal patterns of catch changes based on the 2014-2021 fishery statistics of the Japanese sardine Sardinops melanostictus fishery in the Northwest Pacific Ocean. Three models (XGBoost, LightGBM, and CatBoost) and two variable importance visualization methods (model built-in (split) and SHAP methods) were used for comparative analysis to determine the optimal modeling and visualization strategies. Results: 1) From 2014 to 2021, the annual catch showed an overall increasing trend and peaked at 220,009.063 tons in 2021; the total monthly catch increased and then decreased, with a peak of 76, 033.4944 tons (July), and the catch was mainly concentrated in the regions of 39.5°-43°N and 146.75°-155.75°E; 2) Catboost model predicted better than LightGBM and XGBoost models, with the highest values of accuracy and F1-score, 73.8% and 75.31%, respectively; 3) the overall importance ranking of the model’s built-in method differed significantly from that in the SHAP method, and the overall importance ranking of the spatial variables in the SHAP method increased. Compared to the built-in method, the SHAP method informs the magnitude and direction of the influence of each variable at the global and local levels. The results of the research help us to select the optimal model and the optimal visualization method to construct a prediction model for the Japanese sardine fishing grounds in the Northwest Pacific Ocean, which will provide a scientific basis for the Japanese sardine fishery to achieve environmental and economically sustainable fishery development.