Evaluating the importance of vertical environmental variables for albacore fishing grounds in tropical Atlantic Ocean using machine learning and Shapley additive explanations (SHAP) approach

This study aims to find reliable vertical environmental variables for modeling the fishing grounds of albacore (Thunnus alalunga) in the tropical waters of the Atlantic Ocean. Logbook data of 13 Chinese longliners operating in the high seas of the Atlantic Ocean from 2016 to 2019 were collected and matched with vertical environmental variables including dissolved oxygen, temperature, and salinity from 0 to 500 m at 50‐m depth intervals. Then four machine learning (ML) models: decision tree (DT), random forest (RF), light gradient boosting (LGB) and categorical boosting (CGB) were constructed and compared with generalized additive models (GAMs) within spatial resolutions of .5° × .5°, 1° × 1°, and 2° × 2° grids to find the significant features. The importance of each variable was ranked and compared based on Shapley additive explanations (SHAP) approach across five ML models at three resolutions. Results showed that (1) the vertical environmental variables—temperature at the depth of 100 m and dissolved oxygen concentration at the depth of 100 and 150 m—were the significant features that contributed most to all the ML models at three spatial resolutions; (2) the models with a spatial resolution of 2° × 2° grid exhibited higher accuracy compared to the models with .5° × .5° and 1° × 1° grids; (3) the RF model had the best prediction performance among all the models tested. Our results suggested that significant vertical environmental variables showed similar importance across different ML models at different resolutions, and these specific variables can be relied upon for accurately predicting the fishing grounds of albacore in the tropical waters of the Atlantic Ocean.