Arctic sea ice concentration (SIC) serves as a key geophysical parameter for monitoring sea ice variability and constraining numerical weather and climate models. Although deep learning approaches that integrate active and passive microwave data have recently gained attention, most simply concatenate features from different sensors, ignoring the physical dependencies and spatial–temporal co-structures among radiative, scattering, and environmental driving features, which limits the physical consistency and spatial generalization ability of the retrieval. To address these challenges, this paper proposes a framework for SIC estimation via physical information-guided multi-source data fusion and spatial continuity preservation (PIMS-Net). PIMS-Net designs a physically decoupled three-branch encoder structure to independently extract layered features of passive microwave brightness temperatures, active microwave backscattering signatures, and meteorological reanalysis data. Subsequently, a physical-driven residual correction module (PRCM) is introduced at the multi-scale feature level to explicitly guide the physical correction of passive microwave features with active microwave and meteorological features in latent feature space. Further, a spatial continuity modeling module (SCMM) is embedded in the encoding stage to capture the spatially continuous changes in the sea ice field, enhancing the structural consistency and detail resolution ability of the model in the ice edge and thin ice areas. Experiments on the AI4Arctic Challenge dataset indicate that PIMS-Net achieves better results than the compared baseline models, with a spatial resolution of 80 m, with a mean square error of 0.0138 and a coefficient of determination of 93.5%. • A continuity-aware deep framework is proposed for Arctic SIC retrieval. • PRCM enables physics-guided fusion of multi-source observations. • SCMM captures spatial continuity and regional transitions in SIC fields. • The method improves SIC accuracy and physical consistency over baselines.
Liu et al. (Mon,) studied this question.