Deep Learning Applications in Arctic Sea Ice Remote Sensing: A review

Arctic sea ice is a crucial component of Earth’s system and is experiencing dramatic variations. Satellite remote sensing is one of the most significant tools for monitoring Arctic sea ice. Efficiently extracting sea ice information from satellite-observed materials is a considerable challenge. In recent years, deep learning (DL) has been widely applied to Arctic sea ice remote sensing, enabling advances in classification, parameter retrieval, object detection, superresolution, and prediction. Beyond summarizing existing studies, this review article identifies a clear methodological evolution from early image-based feature learning to multilevel feature integration and, more recently, toward unified representation learning. Compared with existing reviews, this article contributes a stage-based synthesis of methodological evolution, a cross-task comparison of representative models, and practical guidance for framework selection in major Arctic sea ice applications. Through a critical synthesis of the literature, we highlight persistent challenges shared across applications, including limited generalization across regions and sensors, insufficient physical constraints, scarcity in high-quality labeled data, and limited treatment of uncertainty and interpretability. We further discuss emerging directions in DL-based sea ice remote sensing, including sea ice foundation models, physics-informed learning, alternative supervision strategies, uncertainty modeling, and explainable analysis specific to sea ice processes. These insights provide a consolidated perspective on advancing reliable and scientifically meaningful Arctic sea ice remote sensing.