Towards Reliable Sea Ice Drift Estimation in the Arctic: Deep Learning Optical Flow on RADARSAT-2

Daniela Martin, Joseph Gallego · 2025

Read Towards Reliable Sea Ice Drift Estimation in the Arctic: Deep Learning Optical Flow on RADARSAT-2 by Daniela Martin, Joseph Gallego (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-26653.txt.

Introduction

Daniela Martin and Joseph Gallego test deep learning for tracking sea ice drift in the Arctic. Knowing how ice moves matters for navigation, climate research, and forecasting. Optical flow, a vision method that estimates motion between images pixel by pixel, is standard in computer vision but little tried on satellite radar images of ice.

Martin and Gallego run the first large benchmark of forty-eight deep learning optical flow models on RADARSAT-2 radar imagery, scored against GNSS-tracked buoys. Several models reached sub-kilometre accuracy, with errors of a few hundred metres. That is small next to the scales of ice motion and the needs of Arctic navigation. Unlike sparse buoy readings, flow gives a drift estimate for every pixel, which opens new options for navigation and climate modelling.

Catalog record

CREATORDaniela Martin, Joseph Gallego
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-26653

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Towards Reliable Sea Ice Drift Estimation in the Arctic: Deep Learning Optical Flow on RADARSAT-2 Daniela Martin, Joseph Gallego Abstract Accurate estimation of sea ice drift is critical for Arctic navigation, climate research, and operational forecasting. While optical flow —a computer vision technique for estimating pixel-wise motion between consecutive images—has advanced rapidly in computer vision, its applicability to geophysical problems, and more specifically, to satellite SAR imagery, remains underexplored. Classical optical flow methods rely on mathematical models and strong assumptions about motion, which limit their accuracy in complex scenarios; recent deep learning-based approaches have substantially improved performance and are now the standard in computer vision, motivating their application to sea ice drift estimation. We present the first large-scale benchmark of 48 deep learning optical flow models on RADARSAT-2 ScanSAR sea ice imagery, evaluated with endpoint error (EPE) and Fl-all metrics against GNSS-tracked buoys. Several models achieve sub-kilometer accuracy (EPE 6-8 pixels, 300-400 m), a small error relative to the spatial scales of sea ice motion and typical navigation requirements in the Arctic. Our results demonstrate that the models are capable of capturing consistent regional drift patterns and that recent deep learning-based optical flow methods, which have substantially improved motion estimation accuracy compared to classical methods, can be effectively transferred to polar remote sensing. Optical flow produces spatially continuous drift fields, providing motion estimates for every image pixel rather than at sparse buoy locations, offering new opportunities for navigation and climate modeling. 1 Introduction Arctic sea ice is a key component of the climate system, influencing global circulation patterns li2022impacts ; willmes2023patterns , ecosystems steiner2021climate ; campbell2022monitoring , and human activity through its role in navigation and forecasting yastrebova2021positioning ; bayler2024satellite . Accurate estimation of ice drift is therefore essential for both climate research and operational decision-making wang2022intercomparison . Classical approaches for sea ice motion estimation include correlation-based methods such as maximum cross-correlation and phase correlation lavergne2010sea ; hollands2012motion , as well as feature tracking strategies karvonen2012operational . While effective in certain conditions, these methods often degrade in low-texture regions, at floe boundaries, or under strong deformation roberts2019variational ; lei2021autonomous . Optical flow is a computer vision technique that estimates motion at the level of individual pixels between two consecutive images. Each pixel is assigned a vector indicating the direction and magnitude of its movement from the first to the second image, providing a dense motion field across the entire scene alfarano2024estimating . In the context of sea ice, optical flow has the potential to capture detailed drift patterns over large areas, including subtle deformations and floe interactions, which are difficult to resolve with sparse tracking methods such as drifting buoys, whose limited spatial coverage prevents the observation of fine-scale drift patterns and interactions between individual ice floes. Classical optical flow approaches rely on mathematical models, such as brightness constancy constraints combined with spatial regularization, solved through variational or local optimization methods weickert2006survey . While effective in certain contexts, these methods are limited in handling large displacements, complex motion patterns, or noisy data. In recent years, deep learning has become the standard for computer vision tasks, including optical flow, enabling models to learn motion patterns from data and achieve state-of-the-art accuracy and efficiency shah2021traditional . Recent developments in optical flow estimation have focused on improving both accuracy and efficiency, extending the RAFT teed2020raft family of methods into more scalable and generalizable architectures. WAFT wang2025waft challenges the need for cost volumes by relying on high-resolution warping, achieving competitive accuracy with lower memory usage and faster inference. MEMFOF bargatin2025memfof revisits RAFT-like designs with reduced correlation volumes and multi-frame estimation, striking a balance between runtime efficiency and state-of-the-art accuracy, particularly at FullHD resolution. DPFlow morimitsu2025dpflow addresses scalability by generalizing to ultra-high-resolution inputs (up to 8K) without requiring high-resolution training, while also introducing a new benchmark to systematically evaluate methods at large scales. Complementarily, SEA-RAFT wang2024sea improves RAFT’s convergence and generalization through a simplified architecture and novel training strategies, achieving strong cross-dataset performance with reduced computational cost. These approaches consistently achieve state-of-the-art results across standard optical flow benchmarks mehl2023spring ; Butler:ECCV:2012 ; Geiger2012CVPR ; Menze2015CVPR . Collectively, they highlight a shift toward architectures that not only improve accuracy but also address the growing demands of resolution, efficiency, and generalization in optical flow. Progress in this direction has been limited in part by the scarcity of large-scale labeled datasets for Arctic sea ice motion. The 2021 Sea Ice Dynamics Experiment (SIDEx) campaign hutchings2023sidex ; elosegui2023sidex provides a rare opportunity by deploying a set of Global Navigation Satellite System (GNNS) grewal2011global tracked buoys in the Arctic pack ice, delivering high-quality ground-truth trajectories that enable rigorous benchmarking of sea ice motion estimation at scale. In this work, we present the first large-scale benchmark of 48 pretrained deep learning optical flow models, covering convolutional, recurrent, and transformer-based architectures, applied to RADARSAT-2 ScanSAR imagery scheuchl2004potential over the Beaufort Sea in the Alaskan Arctic. Models performance is evaluated against the buoy positions using widely adopted metrics in the computer vision community for optical flow assessment, specifically endpoint error (EPE) and Fl-all metrics. Several achieve sub-kilometer accuracy, highlighting the promise of pretrained vision models for polar geophysical applications. Beyond quantitative scores, we report qualitative patterns and discuss implications for Arctic navigation and climate science. 2 Experimental Setup We evaluate the capability of deep learning optical flow methods to estimate sea ice drift from SAR imagery. Our experiments combine a dedicated RADARSAT-2 dataset with GNSS buoy observations, testing a wide range of pretrained models under consistent preprocessing and standardized evaluation metrics. This section outlines the dataset preparation, model selection, evaluation criteria, and quantitative and qualitative results. 2.1 Dataset Preprocessing We use a dataset of 54 RADARSAT-2 ScanSAR images collected during the 2021 SIDEx campaign in the Beaufort Sea (March–May). The data provide wide coverage of Arctic sea ice at 20–50 m spatial resolution, encompassing diverse conditions such as compact ice, leads, and fractures. The temporal resolution of the dataset is \approx 1.3 days. For validation, we compare the estimated motion fields against ground-truth displacements from 43 buoys hutchings2023sidex , including 12 SATICE buoys equipped with geodetic-quality GNSS tracking elosegui2023sidex . The SATICE buoys recorded positions every 5 minutes, while the other buoys provided updates every 10 minutes, enabling high-temporal-resolution tracking. To align with the SAR acquisitions, we linearly interpolated buoy positions, yielding 1880 displacement observations for evaluation. To ensure consistent model input, all images were resampled to a uniform 50 m spatial resolution. Each image was then cropped to a 40~\text{km}^{2} region centered on the position of the GNSS buoy OSU-IT-23 hutchings2023sidex . This Lagrangian frame of reference flotho2023lagrangian allows the image patches to move with the buoy, retaining fine-scale deformation patterns. 2.2 Optical Flow Models We benchmarked 48 pretrained optical flow models, spanning convolutional, recurrent, and transformer-based architectures. The selection was guided by both code availability and reproducibility of state-of-the-art optical flow methods reported in standard benchmarks mehl2023spring ; Butler:ECCV:2012 ; Geiger2012CVPR ; Menze2015CVPR . While some top-performing models have been reported in the literature, not all had accessible or verifiable implementations. To enable consistent evaluation with our sea ice data, we included only those models for which a reliable implementation could be obtained, using pretrained weights and interfaces provided by the

Read the complete text