LAUSR

ABSTRACT Convolutional neural networks (CNNs) are being used increasingly for classification and regression tasks in remote sensing. However, a challenge in the remote sensing field is having a sufficient quantity of data to train a CNN, in particular, for conditions that are observed less frequently. In this study, a CNN is trained and tested to estimate sea ice concentration from synthetic aperture radar imagery (SAR). We first investigate the importance of including samples of intermediate ice concentration in the training data, which correspond to samples from the marginal ice zone (MIZ). In the present study, ice concentration from image analysis charts is used to provide training data labels. The MIZ for these ice charts are believed to be less accurate than those of the high and low concentrations, but nevertheless, our results support including these samples in the training data set. Additional experiments are then carried out increasing the number of MIZ sample, both from similar regions and a different ice region. It is found the MIZ is represented best in the test data when more samples from a similar region are included. Overall the results improve upon earlier studies, increasing the classification accuracy of the MIZ from 0.66 to 0.74.