<b>Daily high-resolution Arctic sea ice 3D structure inference based on a satellite-driven physics-aware machine learning framework</b>

This study proposes a satellite-driven machine learning (ML) framework (ICE-3D) that performs real-time inference of pan-Arctic sea ice thickness (SIT) based on ocean-atmosphere-ice coupling physical information. ICE-3D establishes a physics-aware indirect retrieval paradigm that generates pan-Arctic daily SIT at a 25 km resolution from 1979 to 2023, without relying on any empirical parameters or SIT observations. We design an innovative training strategy called spatiotemporal multiscale fusion window (STMFW) for fusing physical evolution at different scales in real time to improve targeted inference, especially for summer SIT. ICE-3D employs five decision tree-based ML models that are trained by the reanalysis and then applied to satellite observations through transfer learning. The evaluation results show that AdaBoost outperforms the other four models (MAE=0.11m, PCC=0.97). Compared to the reanalysis products (PIOMAS and TOPAZ4), ICE-3D provides a more accurate summer SIT and a more reasonable estimate of the long-term trend in sea ice volume (SIV). ICE-3D is highly consistent with SIT observations from satellite and in-situ, especially in capturing high-frequency signals and seasonal SIV. In addition, the analysis results show that approximately 38% of the Arctic sea ice has disappeared from 1991 to 2020. Over the past decade, the rate of Arctic sea ice melting has stabilized, indicating that Arctic sea ice reserves may be critically insufficient and the loss is irreversible. ICE-3D provide an opportunity to understand the pan-Arctic climate across different timescales and can play a crucial role during peak periods of the Arctic shipping season.本研究提出了一种卫星驱动的机器学习（Machine Learning，ML）框架ICE-3D，该框架基于海-气-冰耦合物理信息，对泛北极海冰厚度（Sea Ice Thickness，SIT）开展实时反演。ICE-3D构建了具备物理感知性的间接反演范式，可生成1979年至2023年期间分辨率为25公里的泛北极逐日海冰厚度数据，且无需依赖任何经验参数或海冰厚度观测资料。本研究设计了一种名为时空多尺度融合窗口（Spatiotemporal Multiscale Fusion Window，STMFW）的创新训练策略，可实时融合不同尺度的物理演化过程，以提升针对性反演效果，尤其针对夏季海冰厚度反演。ICE-3D采用了五种基于决策树的机器学习模型，这些模型先通过再分析数据进行训练，随后通过迁移学习应用于卫星观测资料。评估结果显示，AdaBoost模型的表现优于其余四种模型（平均绝对误差（Mean Absolute Error，MAE）=0.11米，皮尔逊相关系数（Pearson Correlation Coefficient，PCC）=0.97）。相较于再分析产品PIOMAS与TOPAZ4，ICE-3D可提供更为精准的夏季海冰厚度数据，以及对海冰体积（Sea Ice Volume，SIV）长期变化趋势更为合理的估算结果。ICE-3D与卫星及原位海冰厚度观测数据具有高度一致性，尤其在捕捉高频信号与季节性海冰体积变化方面表现优异。此外，分析结果显示，1991年至2020年间，北极地区约38%的海冰已消失。过去十年间，北极海冰融化速率趋于稳定，这表明北极海冰储量可能已处于严重不足的状态，且海冰流失已不可逆。ICE-3D为人们在不同时间尺度上理解泛北极气候提供了契机，且可在北极航运旺季的峰值时段发挥关键作用。