Study on Improvement of Binary Classification Performance for Ships and Icebergs Using Polarimetric SAR and Deep Learning Approach

In this study, a method for improving the binary classification performance of ships and icebergs using polarimetric synthetic aperture radar (SAR) and deep-learning techniques is proposed. A linear transformation is applied to the polarimetric SAR data to generate hybrid quadrature polarimetric SAR data, which are then incorporated into the existing polarimetric SAR dataset to improve the classification performance. The experiments employed various convolutional neural network models for training, and the performances of the models initialized without pretraining were compared. Using a publicly available SAR dataset from Kaggle, the results demonstrated that the transformation techniques utilizing polarimetric SAR data effectively enhanced the classification performance of ships and icebergs.