An Arctic Sea Ice Thickness Inversion Method Based on Deep Learning Two-Branch Architecture and Multi-Source Remote Sensing Data Fusion

Polar navigation and Arctic research are two fields in which sea ice thickness (SIT) information with medium resolution, high accuracy, and near-real-time availability is of the utmost importance. At present, remote sensing retrieval of Arctic SIT continues to encounter the challenge of balancing spatial and temporal resolution. The inherent limitations of the physical mechanisms of individual sensors result in considerable uncertainties in the retrieval results. The present study proposes a methodology for addressing the aforementioned issue. This methodology is based on a Multi-modal Hierarchical Feature cross-fusion Network (MHFNet), which integrates Synthetic Aperture Radar (SAR), passive microwave, and optical multi-source remote sensing data in a synergistic manner to construct a regression model suitable for Arctic SIT. The network employs a hybrid fusion strategy to achieve pixel-level fusion of image and sequence features at the feature level. The findings suggest that snow depth (SD) is a pivotal factor influencing the precision of SIT retrieval. This study proposes two complementary data fusion schemes. The first of these is the active-passive microwave fusion scheme, which offers all-weather monitoring capability, thereby reducing the mean absolute error (MAE) for first-year ice (0.3–2 m) to 0.2 m under shallow snow (<10 cm) conditions. The second is the optical fusion scheme, which significantly improves the retrieval accuracy for thin ice (0–0.3 m), with the MAE reduced by 14.5%, and achieves an MAE of 6.1 cm for ice thickness in the 0.3–0.8 m range, representing a 50.4% performance improvement. Following the introduction of MHFNet for feature fusion, the coefficient of determination (R²) of the model increased by 12%, and all machine learning models demonstrated significant enhancements across various ice thickness ranges. A comparative validation with independent products such as CryoSat-2 and Soil Moisture and Ocean Salinity (SMOS) demonstrated good consistency in the retrieval results, with the mean absolute error (MAE) generally below 0.5 m, although systematic underestimation remains under deep snow conditions. This study proposes a large-scale, near-real-time system for SIT monitoring in polar regions, offering practical value for engineering applications.