Deep learning based retrieval algorithm for Arctic sea ice concentration from AMSR2 passive microwave and MODIS optical data

Abstract

This study applies deep learning (DL) to retrieve Arctic sea ice concentration (SIC) from AMSR2 data. MODIS-derived SICs are calculated based on spectral unmixing with a new ice/water endmember extraction algorithm that exploits global/local representatives, and then used to train a DL network with AMSR2 data. The resulting SIC maps outperform popular SIC products both regionally and globally. The RMSE of the proposed DL model is 5.19, whereas those of the widely used Bootstrap and ASI-based SIC images are 6.54 and 7.38, respectively, with respect to MODIS-derived SICs at global scale. In particular, our proposed method better describes regions of low-SIC and melting ice in summer, which are generally difficult-to-estimate. As the DL-based model consistently generates accurate SIC values that are not time- or region-dependent, it is considered to be an operational system. Additionally, our SICs can be used to generate initial conditions facilitating development of more accurate climate models.