Multi-Source Deep-Learning Reconstruction of Monthly Arctic Sea Ice Thickness Constrained by Satellite and In Situ Observations

Abstract

Monthly Arctic sea ice thickness (SIT) fields are needed for sea ice volume estimation and climate diagnostics, but continuous reconstruction remains challenging because in situ observations are sparse and satellite retrievals are uncertain. We present ArcticSIT-ReconNet, a multi-source deep-learning framework that combines ERA5 and OSI SAF environmental predictors, a U-Net gridded branch, a GraphSAGE buoy-topology branch, sparse Ice Mass Balance Buoy (IMB) constraints, and CryoSat-2 monthly supervision; CryoSat-2 is not used as a predictor input. For the held-out 2023–2024 winter test period, the model achieved an RMSE of 0.442 m against the CryoSat-2 reference field. Independent evaluation against ICESat-2 yielded an RMSE of 0.650 m, compared with 0.667 m for a matched No-IMB control, supporting a modest contribution from sparse in situ constraints. Regional diagnostics indicate that reconstruction errors are largest in complex coastal and thick-ice regimes. The reconstructed record also captures a coherent winter thickening cycle and modest interannual variability in basin-mean SIT. ArcticSIT-ReconNet therefore provides a multi-source framework for basin-scale Arctic SIT reconstruction and variability analysis.