Application-Specific Synthetic Aperture Radar Image Clustering with Convolutional Autoencoders on Example of Sentinel-1 Scenes Acquired Over Sea Ice

Sea ice is an important component of the Arctic climate system. With the decrease in sea ice extend and sea ice thickness, the human activity in the Arctic is increasing. Better knowledge of present sea ice conditions in the Arctic is required for better planning for the activities as well as for safer and more fuel-efficient shipping routes calculation. One of the bottlenecks in semantic segmentation algorithm developments for satellite image analysis over sea ice is the lack of ground truth information, which is often very hard to acquire. To tackle the problem of limited labeled data, a semi-supervised clustering algorithm is suggested. The algorithm is based on a convolutional neural network autoencoder, which is able to learn descriptors specific to the application. Four autoencoders are trained with 998 scenes in the training dataset and 222 scenes in the test dataset. Two of them include batch normalization and two do not. Two autoencoder are trained on data with open ocean masked out, two - with open ocean included. Out of four autoencoders trained, the best results are shown by the autoencoder with batch normalization included and open ocean masked out in the training dataset. The clustering is then done based on the learned encoder descriptors and a small set of seven manually labeled images acquired over sea ice.