Application of Convolutional Neural Networks for Detecting Sea Ice Leads in the Laptev Sea with Landsat-8 Satellite Imagery

Abstract

A method for detecting leads in the ice of the Arctic seas from satellite images of the visible range is presented. It is shown that sea ice leads are formed under the influence of dynamic processes in the ice cover, such as convergence, drift, and deformation of sea ice, as well as during the interaction of drifting ice with icebergs that have gone aground. The method for identifying sea ice leads is based on the use of artificial intelligence. To analyze the Landsat-8 satellite imagery, a convolutional neural network (U-Net architecture) was used. The method was tested using the satellite images of the visible spectral range that were obtained for the Laptev Sea. The results showed that the lead detection accuracy was above 80%. The method of the minimum rotated rectangle surrounding the polygon was used to determine the geometric parameters of the leads (length, width, inflection points).