基于时间滑动窗口加权融合的东南太平洋茎柔鱼渔场深度学习预测OA

Dosidicus gigas is an important commercial species in the Southeast Pacific Ocean and a key target of China's distant-water fisheries.To address the challenges of fine-scale fishing ground forecasting,particularly the low accuracy and instability under short temporal scales(e.g.,3 d and 6 d),this study proposes a deep learning-based fishing ground prediction method integrated with a temporal sliding-window weighted fusion strategy.The proposed method improves short-term prediction performance by integrating multiple overlapping window-based predictions rather than directly generating short-term dynamic information.Using catch per unit effort(CPUE)from fishery production statistics and sea surface temperature(SST)from satellite remote sensing data during 2012-2021,a U-Net-based deep learning framework was developed to build fishing ground prediction models at five temporal scales(3 d,6 d,10 d,15 d,30 d).Based on this framework,a temporal sliding-window approach was designed to generate multiple overlapping 30 d-scale samples.The trained 30 d model was then applied for prediction,and multiple outputs corresponding to the same 6 d target period were aggregated using a weighted fusion scheme to obtain the final 6 d-scale prediction results.The results showed that the proposed temporal sliding-window weighted fusion method outperformed the directly trained 6 d model,improving both prediction accuracy and temporal continuity.From the perspective of artificial intelligence-based fishery forecasting and temporal sliding-window weighted fusion,this study enriched multi-scale fishing ground prediction methodologies.The proposed method effectively integrated long-term predictive information,suppressed short-term random fluctuations,and enhanced the continuity and robustness of predictions,providing a more accurate tool for fishing ground forecasting.