Creating a semantic segmentation machine learning model for sea ice detection on radar images to study the Thwaites region

• Main Author: Fuentes Soria, Carmen
• Other Authors: Strong, Shay, Sähkötekniikan korkeakoulu, Praks, Jaan, Aalto-yliopisto, Aalto University
• Format: Master Thesis
• Language: English
• Published: 2022
• Subjects: synthetic aperture radar (SAR); machine learning (ML); semantic segmentation; convolutional neural networks (CNN); thwaites glacier; Thwaites Glacier; Antarc*; Antarctica; Sea ice
• Online Access: https://aaltodoc.aalto.fi/handle/123456789/117392

This thesis presents a deep learning tool able to identify ice in radar images from the sea-ice environment of the Thwaites glacier outlet. The project is motivated by the threatening situation of the Thwaites glacier that has been increasing its mass loss rate during the last decade. This is of concern considering the large mass of ice held by the glacier, that in case of melting, could increase the mean sea level by more than +65 cm [1]. The algorithm generated along this work is intended to help in the generation of navigation charts and identification of icebergs in future stages of the project, outside of the scope of this thesis. The data used for this task are ICEYE’s X-band radar images from the Thwaites sea-ice environment, the target area to be studied. The corresponding ground truth for each of the samples has been manually generated identifying the ice and icebergs present in each image. Additional data processing includes tiling, to increment the number of samples, and augmentation, done by horizontal and vertical flips of a random number of tiles. The proposed tool performs semantic segmentation on radar images classifying the class "Ice". It is developed by a deep learning Convolutional Neural Network (CNN) model, trained with prepared ICEYE’s radar images. The model reaches values of F1 metric higher than 89% in the images of the target area (Thwaites sea-ice environment) and is able to generalize to different regions of Antarctica, reaching values of F 1 = 80 %. A potential alternative version of the algorithm is proposed and discussed. This alternative score F 1 values higher than F 1 > 95 % for images of the target environment and F 1 = 87 % for the image of the different region. However, it must not be confirmed as the final algorithm due to the need for further verification.