Retrieval of pan-Arctic sea ice concentration using deep learning

Abstract

Due to the importance and popularity of sea ice concentration (SIC) in polar research, many retrieval algorithms have been proposed to generate SICs from passive microwave data. However, the most SIC retrieval algorithms employ linear combinations of brightness temperatures in different frequencies and polarizations to identify open water, first- and multi-year ice because of the large emissivity differences between water and ice. Additionally they often require tie-point selection, weather filters and land-ocean spillover masks. To handle these limitations, in this research, deep learning (DL), which has recently received increased attentions in diverse fields of study, is incorporated into passive microwave and optical remote sensing data to retrieve more accurate SIC information than the past. To create true SIC labels which is the most critical part in DL model training and evaluation, we first propose a spectral unmixing based true SIC calculation algorithm. The true SIC labels are then used to train a DL-based Arctic SIC retrieval model. Therefore, we obtained visually and statistically improved Arctic SIC maps, and the results outperformed popular Bootstrap and ASI SIC retrieval algorithms at global and local scales. Our proposed method especially captured more detailed SIC representations and variability in difficult-to-estimate thin and melting ice zones in summer than other algorithms. The consistency in time and space of the proposed retrieval model enables it to be a new operational SIC retrieval algorithm in practice. Further, more accurate SIC products as initial conditions can allow capability to improve climate models