Toward Enhanced Sea Ice Parameter Estimation: Fusing Ice Surface Temperature with the AI4Arctic Dataset using Convolutional Neural Networks

Loading...
Thumbnail Image

Date

2025-04-14

Advisor

Scott, Andrea
Clausi, David

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

Arctic sea ice mapping is essential for supporting several key applications. These include facilitating safe marine navigation, providing accurate data for climate monitoring, and assisting efforts by remote northern communities to adapt to variable ice conditions. Automated mapping approaches can leverage an abundance of freely accessible satellite data, with the potential to supplement navigational ice charts, improve operational forecasting, and produce high-resolution estimates of sea ice parameters. However, current approaches rely on synthetic aperture radar (SAR) and passive microwave (PM) data, which can struggle to distinguish ice features due to ambiguous textures, atmospheric effects, and sensor limitations. This thesis explores the potential for thermal-infrared data to improve estimates of sea ice concentration, stage of development, and floe size produced by multi-task deep learning architectures. Work builds on the recent AI4Arctic dataset, which combines Sentinel-1 SAR, AMSR2 brightness temperature, ERA-5 reanalysis data, and ice charts to enhance deep learning-based mapping approaches. VIIRS ice surface temperature (IST) is investigated for its potential to improve predictions in regions where SAR and PM measurements are challenging to interpret. A VIIRS-AI4Arctic dataset is developed, which consists of 84 scenes, and demonstrates overlap between VIIRS, Sentinel-1, and AMSR2 products. Three variations on the U-Net architecture are introduced, which incorporate IST features at the input- and feature-levels. These models are evaluated against the winning AI4EO AutoICE Challenge architecture, which acts as an AI4Arctic baseline. A SIC accuracy metric is introduced to provide an additional assessment of model performance. Results demonstrate that models incorporating IST consistently reduce classification errors across all three tasks, particularly when identifying open water under conditions with low-incidence angle (SAR), high atmospheric moisture (PM), and wind roughening (SAR and PM). A single, shared decoder improves contextual awareness, although multi-decoder architectures effectively reconstruct task-specific features. The DEU-Net-V architecture, which learns IST features separately from AI4Arctic channels, is most effective at mitigating ambiguity introduced by SAR and PM data. Finally, estimation of aleatoric uncertainty yields heightened variance in marginal ice zones, highlighting potential discrepancies between ice chart labels and pixel-level conditions, and demonstrating the value of quantifying uncertainty from observation noise. IST ultimately enhances sea ice classification, but is limited by cloud contamination and the resolution of current products. These findings support the continued development of deep learning approaches incorporating IST, and highlight the potential for next-generation thermal-infrared instruments to further improve automated sea ice mapping.

Description

Keywords

sea ice, machine learning, deep learning, remote sensing, uncertainty quantification, ice surface temperature, synthetic aperture radar, passive microwave, data fusion, sea ice concentration, stage of development, floe size, AI4Arctic, VIIRS

LC Subject Headings

Citation