Remote sensing of sea ice leads with Sentinel-1 C-band synthetic aperture radar

Abstract

The presence of leads with open water or thin ice is an important feature of the Arctic sea ice cover. Leads regulate the heat, gas, and moisture fluxes between the ocean and atmosphere and are areas of high ice growth rates during periods of freezing conditions. In the present study an algorithm providing an automatic lead detection based on Synthetic Aperture Radar (SAR) images is developed using traditional machine learning techniques and deep learning methods. The algorithm is applied to a wide range of Sentinel-1 scenes taken over the Arctic Ocean. Distribution of the detected leads in the Arctic during winter seasons 2016--2021 is then analyzed.

An important part of the algorithm development is the data preprocessing as the classification quality depends on the quality of the input images. An advanced data preparation technique improves consistency of the cross-polarization channel and enables the use of dual-polarization SAR images. By using both the HH and the HV channels instead of single co-polarized observations the algorithm is able to detect more leads compared to the use of the HH polarization only.

First, a traditional machine learning approach is described. It is based on polarimetric features and texture features derived from the grey level co-occurrence matrix. The Random Forest classifier is used to investigate the individual feature importance on the lead detection. The precision-recall curve representing the quality of the classification is assessed to define a threshold for the binary lead/sea ice classification. The algorithm produces a lead classification with more than 90% precision with 60% of all leads classified, as evaluated on the test data. The precision can be increased by the cost of the amount of leads detected. Classification quality is improved by introducing an advanced binarization method based on watershed segmentation. Further improvements include object shape analysis resulting in a shape-based filter, which efficiently removes objects appearing due to noise patterns over young ice.

Second, an algorithm based on a convolutional neural network is developed. It shows more robust results compared to the algorithm based on the gray level co-occurrence matrix with Random Forest classification and is applicable to the entire Arctic Ocean. Classification results are evaluated against the dataset which does not include training or testing data, and are also compared to Sentinel-2 optical satellite images.

Finally, the lead detection algorithm is applied to all Sentinel-1 EW GRDM scenes taken in five winter seasons, 1 November - 30 April of 2016-2021 years. 3-day composite pan-Arctic lead maps with the native Sentinel-1 40~meters pixel spacing are produces. The frequency of lead occurrence derived from these maps is compared with MODIS thermal infrared lead detection results. The lead area fraction is compared with the AMSR2 passive microwave observations. The lead area distribution, lead length, and lead width distributions, as well as the lead orientation distributions, are analyzed in the following regions of the Arctic Ocean: Fram Strait, Barents Sea, Kara Sea, Laptev Sea, East Siberian Sea, Chukchi Sea, Beaufort Sea, Central Arctic. Each region shows the presence of regularity in lead orientation, the preferred orientation has little variation from year to year and during season. The lead width distribution is found to follow the power low with the exponent of 1.86 with 0.16 standard deviation. The yearly mean lead area fraction derived from Sentinel-1 images varies from 2.5% to 3.7% during winter seasons 2016-2021.