Marine fishery resource dynamic prediction based on CNN-XGBoost fusion model

Marine fishery resource prediction is crucial for sustainable fishery management and ecosystem conservation, yet traditional statistical methods face limitations in capturing the complex non-linear relationships and multi-scale temporal dependencies inherent in marine environmental systems. This study proposes a novel CNN-XGBoost fusion model that integrates convolutional neural networks' temporal pattern recognition capabilities with extreme gradient boosting's ensemble learning strengths for enhanced marine fishery resource forecasting. The fusion architecture employs a hierarchical two-stage framework where CNN components extract high-level temporal features from multi-source marine environmental data, while XGBoost modules process both extracted features and engineered variables to generate final predictions. Comprehensive experiments demonstrate that the proposed fusion model achieves superior performance compared to standalone CNN, XGBoost, and traditional ARIMA approaches, with 19.1% improvement in RMSE and statistically significant enhancements across all evaluation metrics. The optimal fusion weight analysis reveals that CNN-extracted features and XGBoost-processed features are weighted at 40 and 60% respectively in the final prediction fusion, achieving RMSE of 2.847, MAE of 2.184, and R² of 0.846. These percentages represent fusion weight allocation rather than prediction accuracy values. Time series analysis confirms robust performance across seasonal variations and exceptional capability in predicting extreme abundance events critical for adaptive fishery management. The results provide valuable insights for sustainable marine resource management and offer practical tools for fishery policymakers and resource managers.