Accounting for Label Errors When Training a Convolutional Neural Network to Estimate Sea Ice Concentration Using Operational Ice Charts

Convolutional neural networks (CNNs) are being increasingly investigated as a means to extract sea ice concentration from synthetic aperture radar (SAR) in an automated manner. This is often done using ice charts as training data. However, in these charts, an ice concentration label is given to a large region, which may not have a spatially uniform sea ice concentration distribution at the prediction scale of the CNN. This leads to representativity errors, which can be more pronounced at intermediate sea ice concentrations. In this study, we first investigate ways to perturb the ice chart labels to obtain improved predictions to account for the label uncertainty for intermediate ice concentrations. We then propose a method to augment the ice chart data by rescaling the information in the SAR imagery. The method is found to lead to improved accuracy in comparison to using the ice chart labels alone, with accuracy improving from 0.919 to 0.966. The sea ice concentration maps with the augmented labels also have much finer detail than the other approaches evaluated. These details are visually in agreement with expected sea ice concentration from the SAR data.