Improved Arctic Sea Ice Forecasts in the Melting Season by Retrieving Sea Ice Thickness From GNSS‐R Based on DNN

Abstract Accurate sea ice forecasting is essential for safe navigation in the Arctic Ocean, but it remains particularly challenging during the summer melting season. This difficulty mainly arises from incomplete sea ice initialization due to the scarcity of sea ice thickness (SIT) observations from May to September. To address this issue, we develop a method to retrieve SIT using Deep Neural Network. The model integrates reflection data from Global Navigation Satellite System‐Reflectometry (GNSS‐R) with satellite‐derived sea ice concentration (SIC) and relevant atmospheric and oceanic parameters. The retrieved SIT and observed SIC are assimilated into a mesoscale‐permitting ice‐ocean coupled model to generate a continuous sea ice “reanalysis” data set for the entire melting season. Evaluation shows that our reanalysis outperforms existing data sets, reducing errors in both SIC and SIT by over 10%. It provides a notably improved sea ice distribution in key marginal zones and serves as a superior initial field for synoptic‐scale sea ice forecasts. SIT forecast errors are substantially reduced in regions with historically large prediction biases, such as the Beaufort Sea, where the 7th‐day forecast error decreases by more than 80% compared to forecasts that do not use our retrieved SIT. This study confirms the value of combining GNSS‐R technology with atmospheric and oceanic data for retrieving SIT, and highlights its importance for improving short‐term Arctic sea ice forecasts.