Experiments with estimation of sea ice concentration using a convolutional neural network and microwave data

Using an existing convolutional neural network (CNN) framework that uses both synthetic aperture radar (SAR) data and passive microwave data to learn sea ice concentration (SIC), we test the inclusion of different passive microwave channels and also perturbing the labels used in the loss function. Improved results are found if the low frequency channels from the passive microwave data are included. We also find perturbing the labels may be a reasonable approach to improved SIC estimation in marginal ice zones.