Arctic Sea ICE Mapping Using Sentinel-1 SAR Scenes with a Convolutional Neural Network

In this paper we focus on automated sea ice mapping based on satellite images with deep learning methods. We adjust the UNET++ convolutional neural network architecture for sea ice type semantic segmentation and create a per-pixel classifier for Sentinel-1 SAR scenes. The classification is done tile-wise, i.e. a Sentinel-1 scene is divided into tiles, classified, and then the results are joined back to form the classified scene. We address to the border effect appearing when a tiled classification is applied.