Exploring the Habitat Distribution of Decapterus macarellus in the South China Sea Under Varying Spatial Resolutions:...

The selection of environmental variables with different spatial resolutions is a critical factor affecting the accuracy of machine learning-based fishery forecasting. In this study, spring-season survey data of Decapterus macarellus in the South China Sea from 2016 to 2024 were used to construct six machine learning models-decision tree (DT), extra trees (ETs), K-Nearest Neighbors (KNN), light gradient boosting machine (LGBM), random forest (RF), and extreme gradient boosting (XGB)-based on seven environmental variables (e.g., sea surface temperature (SST), chlorophyll-a concentration (CHL)) at four spatial resolutions (0.083°, 0.25°, 0.5°, and 1°), filtered using Pearson correlation analysis. Optimal models were selected under each resolution through performance comparison. SHapley Additive exPlanations (SHAP) values were employed to interpret the contribution of environmental predictors, and the maximum entropy (MaxEnt) model was used to perform habitat suitability mapping. Results showed that the XGB model at 0.083° resolution achieved the best performance, with the area under the receiver operating characteristic curve (ROC_AUC) = 0.836, accuracy = 0.793, and negative predictive value = 0.862, outperforming models at coarser resolutions. CHL was identified as the most influential variable, showing high importance in both the SHAP distribution and the cumulative area under the curve contribution. Predicted suitable habitats were mainly located in the northern and central-southern South China Sea, with the latter covering a broader area. This study is the first to systematically evaluate the impact of spatial resolution on environmental variable selection in machine learning models, integrating SHAP-based interpretability with MaxEnt modeling to achieve reliable habitat suitability prediction, offering valuable insights for fishery forecasting in the South China Sea.