Improved lead detection with Sentinel-1 SAR images in the Arctic

Abstract

The Sentinel-1 synthetic aperture radar satellite constellation provides a valuable source of information on sea ice conditions over the Arctic region. It is able to cover a major part of the Arctic Ocean every 3 days with two satellites, which provides an opportunity for fine scale object monitoring on a global scale. Lead distributions in the Arctic are of interest both for shipping as well as climate process like energy and gas exchange between the ocean and atmosphere in ice covered regions. We have previously introduced an algorithm for automatic lead detection with Sentinel-1 SAR images based on grey level co-occurrence matrix and random forest classification. Here we introduce a new method for lead detection based on UNET convolutional neural network. An important part influencing classification result quality is a new preprocessing procedure that provides more consistency between Sentinel-1 scene subswaths, which is especially important for the cross-polarization HV images. We present the results of the new classification on fine scale and Arctic-wide, and compare the new results with those produced with the previously suggested algorithm.