军事知识服务平台
Benchmarking architectures of deep learning-based models for sea ice concentration and thickness spatiotemporal prediction
摘要 Accurate spatiotemporal prediction of Arctic sea ice will contribute to revealing the mechanisms of sea ice variability within the atmosphere-ocean-ice system.In recent years,deep learning-based spatiotemporal prediction models have diversified significantly and found applications across numerous domains.Their rapid development offers new alternatives for Arctic sea ice spatiotemporal forecasting.However,current deep learning models applied to sea ice prediction primarily follow the recurrent neural network(RNN)-based and convolutional neural network(CNN)-based paradigm,while the applicability of other architectures,such as transformer-based models,remains relatively unexplored in the context of sea ice prediction.Additionally,the capability of these models to specifically characterize different sea ice fields[sea ice concentration(SIC)and sea ice thickness(SIT)]warrants further investigation.To address these gaps,this study selects eight classical deep learning-based spatiotemporal prediction models(four recurrent and four non-recurrent architectures)to conduct weather-scale forecasting using the Global Sea Ice Concentration Climate Data Record(GSICCDR)SIC data and CryoSat-2/SMOS Merged Sea Ice Thickness(CS2SMOS)data.A systematic evaluation and comprehensive analysis are performed to assess model performance across key dimensions,including overall accuracy,computational efficiency,spatiotemporal variability,and adaptability to anomalous conditions.From a structural perspective,this study elucidates the predictive disparities among models,providing insights for designing sea ice-specific forecasting architectures.Furthermore,the research delineates the divergent responses of models to SIC and SIT,thereby clarifying their respective suitability for different sea ice fields.
刊 名 海洋学报(英文版) ISTICSCI
年,卷(期): 2026, 45(7)
机标分类号: P445 F713.36 TP18
在线出版日期: 2026年10月20日