Estimating Sea Ice Concentration From Microwave Radiometric Data for Arctic Summer Conditions Using Machine Learning

The Arctic region is sensitive to climate change, and polar sea ice is a crucial indicator of global climate change. Microwave radiometry has been applied to retrieve Arctic sea ice concentration (SIC) for over 50 years. During summer, the retrieval algorithm for SIC based on microwave brightness temperature is affected by the melt ponds or wet sea ice, which may cause bias in the estimated SIC. In this study, a machine learning (ML) model is constructed using special sensor microwave imager (SSM/I)-special sensor microwave imager/sounder (SSMIS) brightness temperature data as input variables and the 2001-2020 moderate resolution imaging spectroradiometer (MODIS) SIC as a reference dataset to retrieve SIC, followed by a validation analysis using Landsat SIC, ship-based visual observations of SIC, and MODIS SIC. The comparison results show that the precision of SIC retrieved by the ML model is higher and effectively improves the bias of SIC by using microwave radiometry in summer conditions. The results show that, whether in the entire Arctic or localized regions with intense melt ponds, the ML-based SIC is superior to the four canonical microwave SIC products [Arctic Radiation and Turbulence Interaction Study Sea Ice (ASI), Ocean and Sea Ice Satellite Application Facility (OSI), National Aeronautics and Space Administration (NASA) Team (NT), and Bootstrap (BT)]. Based on the Arctic sea ice dataset obtained from 1988 to 2020 in this study, the spatiotemporal trends of Arctic SIC are analyzed. The results indicate a significant declining trend of SIC in the Arctic, which agrees with classic SIC products. The most pronounced ice reduction is observed in the Barents Sea, Chukchi Sea, East Siberian Sea, Kara Sea, Laptev Sea, and Beaufort Sea.