Comment on egusphere-2023-2648

Andreas Stokholm, Jørgen Buus-Hinkler, Tore Wulf, Anton Korosov, Roberto Saldo, Leif Toudal Pedersen, David Arthurs, Ionut Dragan, Iacopo Modica, Juan Pedro, Annekatrien Debien, Xinwei Chen, Muhammed Patel, Fernando J. Pena Cantu, Javier Noa Turnes, Jinman Park, Linlin Xu, Andrea K. Scott, David A. Clausi, Yuan Fang, Mingzhe Jiang, Saeid Taleghanidoozdoozan, Neil C. Brubacher, Armina Soleymani, Zacharie Gousseau, Michał Smaczny, Patryk Kowalski, Jacek Komorowski, David Rijlaarsdam, Jan N. van Rijn, Jens Jakobsen, Martin S. J. Rogers, Nick Hughes, Tom Zagon, Rune Solberg, Nicolas Longépé, Matilde Brandt KreinerOpenalex Unknown 2024 · DOI · OpenAlex

Abstract

Mapping sea ice in the Arctic is essential for maritime navigation, and growing vessel traffic highlights the necessity of timeliness and accuracy of sea ice charts. In addition, with the increased availability of satellite imagery, automation is becoming more important. The aim of the AutoICE Challenge was to encourage the creation of models capable of mapping sea ice automatically from spaceborne Synthetic Aperture Radar (SAR) imagery using deep learning while inspiring participants to move towards multiple sea ice parameter model retrieval instead of the current focus on a single sea ice parameter, such as concentration. Participants were tasked with the development of machine learning algorithms mapping the total sea ice concentration, stage of development and floe size using a state-of-the-art sea ice dataset with dual-polarised Sentinel-1 SAR images and 22 other relevant variables while using professionally labelled sea ice charts from multiple national ice services as reference data. The challenge had 129 teams representing a total of 179 participants, with 34 teams delivering 494 submissions, resulting in a participation rate of 26.4 %, and was won by a team from the University of Waterloo. Participants were successful in training models capable of retrieving multiple ice parameters with convolutional neural network and vision transformer models. The top participants scored best on the total sea ice concentration and stage of development, while the floe size was more difficult. Furthermore, participants offered intriguing approaches and ideas that could help propel future research within automatic sea ice mapping, such as applying high downsampling of SAR data to improve model efficiency and produce better results.

Introduction

As Arctic ice thins and shipping traffic increases, the need for timely, accurate sea ice charts has never been greater. Traditionally, these charts are hand-drawn by professional ice analysts interpreting satellite imagery—a process that is slow, labor-intensive, and subjective. While previous research has successfully used machine learning to automate parts of this process, most models focus on a single variable: sea ice concentration (SIC). However, for safe maritime navigation, sailors also need to know the "Stage of Development" (SOD)—a proxy for ice thickness—and "Floe Size" (FLOE), which describes how broken up the ice is. The AutoICE Challenge was designed to move the field toward multi-parameter retrieval, forcing models to handle the complex, multi-modal nature of real-world ice charting.

Methodology and Model Architectures

The challenge invited 129 teams to develop models capable of pixel-wise segmentation of SAR imagery. Key methodological highlights include:

  • Multi-task Learning: Instead of separate models, top teams used multi-task architectures (like modified U-Nets) with multiple "heads" to predict SIC, SOD, and FLOE simultaneously.
  • Data Fusion: Models fused high-resolution Synthetic Aperture Radar (SAR) data with low-resolution passive microwave data (AMSR2), atmospheric reanalysis (ERA5 wind and temperature), and geographical metadata (latitude, longitude, and distance to land).
  • Context over Resolution: In a surprising move, the winning team from the University of Waterloo (UW) applied 10x downsampling to the SAR data. This increased the model's "field of view," allowing it to see larger spatial structures and patterns that are more indicative of ice types than fine-grained pixel details.
  • Architectures: While U-Net variants remained the baseline, some teams successfully experimented with Vision Transformers (ViT) and DeepLabV3 to capture multi-scale contextual information.

The AI4Arctic Sea Ice Challenge Dataset (ASID)

The challenge utilized the AI4Arctic Sea Ice Challenge Dataset (ASID), a massive compilation of co-located satellite and environmental data:

  • Input Data: Dual-polarized Sentinel-1 SAR images (HH and HV), AMSR2 brightness temperatures (6.9 to 89 GHz), and ERA5 numerical weather prediction parameters.
  • Ground Truth: Professionally labeled sea ice charts from the Danish Meteorological Institute (DMI) and the Canadian Ice Service (CIS), following the international SIGRID-3 standard.
  • Scale: The training set included 513 scenes spanning 2018–2021, with a hidden test set of 20 scenes used for final ranking.

Key Results and Insights

The University of Waterloo team took first place with a combined score of 86.39%. Performance varied significantly across the three target parameters:

  • SIC (Concentration): The most successful parameter, with the winner achieving 92.02% accuracy.
  • SOD (Stage of Development): Strong results at 88.61%.
  • FLOE (Floe Size): Significantly lower performance, averaging around 74.63% across top teams.

Several key insights emerged from the challenge:

  • The "Big Picture" Wins: The success of the 10x downsampling strategy suggests that for sea ice classification, spatial context is more important than raw resolution. By seeing a wider area, the model can better distinguish between ambiguous backscatter signatures (e.g., distinguishing open water from smooth young ice).
  • The FLOE Bottleneck: Floe size was consistently the hardest parameter to predict. This is likely because floe boundaries are often smaller than the resolution used by the models, and the labeling of floe size in human charts is notoriously inconsistent.
  • Ancillary Data is Essential: SAR data alone is often ambiguous. Fusing it with passive microwave and temperature data provides the physical constraints necessary to disambiguate signatures.

Limitations and Future Directions

Despite the progress, challenges remain. Human analysts can disagree on ice concentration by as much as 20% to 60%, introducing "label noise" into the training data. Future work may need to move toward more objective ground truth, such as high-resolution optical or altimetry data. Additionally, while the challenge evaluated models pixel-by-pixel, operational ice charts are produced as polygons; future benchmarks might evaluate at the polygon level to better reflect real-world needs. The upcoming ASID-v3 dataset, which will be 16 times larger, promises to provide even more diversity for training the next generation of automated mapping models.

Summary updated

Abstract

<strong class="journal-contentHeaderColor">Abstract.</strong> Mapping sea ice in the Arctic is essential for maritime navigation, and growing vessel traffic highlights the necessity of timeliness and accuracy of sea ice charts. In addition, with the increased availability of satellite imagery, automation is becoming more important. The aim of the AutoICE Challenge was to encourage the creation of models capable of mapping sea ice automatically from spaceborne Synthetic Aperture Radar (SAR) imagery using deep learning while inspiring participants to move towards multiple sea ice parameter model retrieval instead of the current focus on a single sea ice parameter, such as concentration. Participants were tasked with the development of machine learning algorithms mapping the total sea ice concentration, stage of development and floe size using a state-of-the-art sea ice dataset with dual-polarised Sentinel-1 SAR images and 22 other relevant variables while using professionally labelled sea ice charts from multiple national ice services as reference data. The challenge had 129 teams representing a total of 179 participants, with 34 teams delivering 494 submissions, resulting in a participation rate of 26.4 %, and was won by a team from the University of Waterloo. Participants were successful in training models capable of retrieving multiple ice parameters with convolutional neural network and vision transformer models. The top participants scored best on the total sea ice concentration and stage of development, while the floe size was more difficult. Furthermore, participants offered intriguing approaches and ideas that could help propel future research within automatic sea ice mapping, such as applying high downsampling of SAR data to improve model efficiency and produce better results.
BibTeX
@inproceedings{stokholm2024comment,
  author = {Andreas Stokholm and Jørgen Buus-Hinkler and Tore Wulf and Anton Korosov and Roberto Saldo and Leif Toudal Pedersen and David Arthurs and Ionut Dragan and Iacopo Modica and Juan Pedro and Annekatrien Debien and Xinwei Chen and Muhammed Patel and Fernando J. Pena Cantu and Javier Noa Turnes and Jinman Park and Linlin Xu and Andrea K. Scott and David A. Clausi and Yuan Fang and Mingzhe Jiang and Saeid Taleghanidoozdoozan and Neil C. Brubacher and Armina Soleymani and Zacharie Gousseau and Michał Smaczny and Patryk Kowalski and Jacek Komorowski and David Rijlaarsdam and Jan N. van Rijn and Jens Jakobsen and Martin S. J. Rogers and Nick Hughes and Tom Zagon and Rune Solberg and Nicolas Longépé and Matilde Brandt Kreiner},
  title = {{Comment on egusphere-2023-2648}},
  booktitle = {Openalex Unknown},
  year = {2024},
  doi = {10.5194/egusphere-2023-2648-rc2},
}
Figure