Deep Learning for Super-resolution of Sea Ice Concentration: Case Study of OSISAF Downscaler

Julia Borisova, Dmitry Morozov, Dmitry Gilemkhanov, Nikolay Nikitin

TL;DR: We built a deep learning model that turns low-resolution (25 km) Arctic sea ice maps into high-resolution (4 km) ones. It makes the ice edges and patterns more realistic by learning from a high-quality reference dataset

Abstract: This paper presents an approach for Arctic sea ice concentration correction and downscaling based on multi-sensor fusion that is applied only during the training stage. While conventional downscaling methods primarily enhance spatial resolution, the proposed approach also improves physical consistency by fusing the widely used OSISAF sea ice concentration dataset with the high-resolution multi-sensor-corrected MASAM2 product. This synthesis not only increases spatial resolution from 25 km to 4 km, but also yields more realistic ice patterns, as validated by a mean absolute error (MAE) of 0.009 and high structural similarity (SSIM) scores. Comparisons with classical downscaling techniques and manual product hybridization confirm the superior performance of our model. The proposed solution has strong potential to improve sea ice forecasting and generate high-resolution historical datasets. Code and weights are provided on GitHub - https://github.com/ITMO-NSS-team/ML4RS_2026_arctic_downscale.

Submission Number: 49