Geographically-Weighted Weakly Supervised Bayesian High-Resolution Transformer for 200m Resolution Pan-Arctic Sea Ice Concentration Mapping and Uncertainty Estimation using Sentinel-1, RCM, and AMSR2 Data
Authors: Mabel Heffring, Lincoln Linlin Xu
Organizations: Department of Geomatics Engineering, Schulich School of Engineering, University of Calgary, 2500 University Dr NW, Calgary, T2N 1N4, AB, Canada
Although high-resolution mapping of pan-Arctic sea ice with reliable corresponding uncertainty is essential for operational sea ice concentration (SIC) charting, it is a difficult task due to key challenges, such as the subtle nature of ice signature features, inexact SIC labels, model uncertainty, and data heterogeneity. This study presents a novel Bayesian High-Resolution Transformer approach for 200 meter resolution pan-Arctic SIC mapping and uncertainty quantification using Sentinel-1, RADARSAT Constellation Mission (RCM), and Advanced Microwave Scanning Radiometer 2 (AMSR2) data. First, to improve small and subtle sea ice feature (e.g., cracks/leads, ponds, and ice floes) extraction, we design a novel high-resolution Transformer model with both global and local modules that can better discern the subtle differences in sea ice patterns. Second, to address low-resolution and inexact SIC labels, we design a geographically-weighted weakly supervised loss function to supervise the model at region level instead of pixel level, and to prioritize pure open water and ice pack signatures while mitigating the impact of ambiguity in the marginal ice zone (MIZ). Third, to improve uncertainty quantification, we design a Bayesian extension of the proposed Transformer model, treating its parameters as random variables to more effectively capture uncertainties. Fourth, to address data heterogeneity, we fuse three different data types (Sentinel-1, RCM, and AMSR2) at decision-level to improve both SIC mapping and uncertainty quantification. The proposed approach is evaluated under pan-Arctic minimum-extent conditions in 2021 and 2025. Results demonstrate that the proposed model achieves 0.70 overall feature detection accuracy using Sentinel-1 data, while also preserving pan-Arctic SIC patterns (Sentinel-1 R\textsuperscript{2} = 0.90 relative to the ARTIST Sea Ice product).
Figures & tables
Sensor
Sentinel-1
RCM
AMSR2
Product Level
Level-1 GRD
Level-1 GRD
Level-1B
Acquisition Mode(s)
Extra Wide (EW)
Low Resolution, Medium Resolution, Low Noise
Conical scanning
Frequency Band(s)
C-Band
C-Band
18.7, 23.8, 36.5, 89.0 GHz
Polarizations
HH, HV
HH, HV
H, V
Ground Resolution
20 x 40 m
50 m - 100 m
14 x 22 km, 15 x 26 km, 7 x 12 km, 3 x 5 km
Daily Raw Data
∼ 30 GB
∼ 200 GB
∼ 1 GB
Table 1: Overview of Sentinel-1, RCM, and AMSR2 Data Specifications.
Figure 1: Overview of our key innovations for pan-Arctic SIC mapping with corresponding uncertainty quantification. The proposed model utilizes a (a) High-Resolution Transformer architecture, fω(.) , with an among-token Transformer block for modeling global context (GloFormer) and a within-token Transformer block for modeling local detail (LoFormer). The high-resolution Transformer becomes a (b) Bayesian Neural Network (BNN) by assuming the model parameters in the attention mechanisms are probabilistic, as highlighted in red. The primary training strategy is a (c) geographically-weighted weakly supervised LL1−GW loss , where the difference between the predicted SIC and NASA Team ground truth is evaluated at the cluster level rather than the high-resolution pixel level. Geographical weights dynamically scale the importance of each sample according to the NIC ice chart. Individual Bayesian Transformer models are trained using Sentinel-1, RCM, and AMSR2 89.0GHz in parallel. After variational inference, (d) decision-level data fusion is used to combine the SIC and corresponding uncertainty maps.
Figure 2: Theoretical Comparison of Epistemic Uncertainty Quantification Approaches. Figure (a) illustrates the Bayesian Neural Network (BNN) where the model parameters are assumed to be probabilistic. Figure (b) illustrates MC dropout where red denotes masked nodes and green denotes masked connections. With each variational inference the dropout configuration changes, introducing randomness into the neural network. Figure (c) illustrates Epoch ensemble generation where model parameters at each epoch are treated as independent ensemble members.
Figure 3: Overview of data downloaded and processed for September 4th, 2021. Includes Sentinel-1 (a) HH, (b) HV, and (c) HH and HV cross-polarization, RCM (f) HH, (g) HV, and (h) HH and HV cross-polarization, AMSR2 (d) 89.0 GHz H and (e) 89.0 GHz V data, as well as (i) NIC Ice Chart and (j) NASA Team SIC products.
Date
Dataset
Sentinel-1
RCM
AMSR2
2021/09/01
Training
5858
4393
10,194
2021/09/04
Validation
5270
4694
10,195
2021/09/07
Training
6333
3923
10,195
2021/09/10
Training
6058
4613
10,195
2021/09/14
Training
4612
2819
10,196
2021/09/18
Validation
6568
8014
10,195
Table 2: Temporal Distribution of 256 x 256 Samples for Training, Validation, and Testing.
Figure 4: Comparison of predicted SIC (left), the ice/water binary ground truth mask derived using adaptive thresholding of the NIC ice chart (middle), and corresponding SAR HV (right) for one 256 x 256 validation sample.
Figure 5: Pan-Arctic SIC from the Bayesian High-Resolution Transformer for validation data on September 4th, 18th, and 27th, 2021. SIC derived from (a) Sentinel-1, (b) RCM, and (c) AMSR2, and (d) is the final fused map of all data sources with Sentinel-1 layered on top, followed by RCM and AMSR2. SIC estimates are visually compared to (e) NASA Team SIC and (f) NIC Ice Chart.
Figure 7: Pan-Arctic SIC from the Bayesian High-Resolution Transformer for test data on September 3rd, 10th, and 30th, 2025. SIC derived from (a) Sentinel-1, (b) RCM, and (c) AMSR2, and (d) is the final fused map of all data sources with Sentinel-1 layered on top, followed by RCM and AMSR2. SIC estimates are visually compared to (e) NASA Team SIC and (f) NIC Ice Chart.
Dataset
Method
MIZ
Ice Pack
Sentinel-1
MC Dropout
0.0171
0.0226
Epoch Ensemble
0.0169
0.0207
Bayesian Transformer
0.0018
0.0022
RCM
MC Dropout
0.0145
0.0238
Epoch Ensemble
0.0144
0.0194
Bayesian Transformer
0.0019
0.0017
Table 3: Weighted Expected Calibration Error (ECE) for MIZ and Ice Pack. Best performance is bolded .
Figure 9: Comparison of uncertainty quantification approaches for Sentinel-1, RCM, and AMSR2. Uncertainty estimates are shown using (a) MC Dropout, (b) Epoch Ensemble, and (c) BBB implemented in the Bayesian Transformer for September 4th, 2021. The distribution of uncertainty for all validation samples are shown using (d) Gaussian Kernel Density curves for each dataset with corresponding mean and standard deviation.
Ice Concentration Classes (%)
Uncertainty Estimation
Input Data
0-10
10-20
20-30
30-40
40-50
50-60
60-70
70-80
80-90
90-100
Sentinel-1
0.17
5.38
5.59
5.35
4.85
4.37
3.68
2.65
1.44
0.82
RCM
0.18
7.53
7.94
8.94
8.86
7.80
6.15
4.13
2.07
1.05
MC Dropout
AMSR2
7.95
18.29
22.57
25.24
28.04
29.66
27.15
21.55
13.81
4.49
Sentinel-1
0.72
11.69
12.65
12.96
12.52
11.37
9.49
6.85
3.80
1.94
RCM
0.64
10.64
12.31
14.03
14.29
13.32
11.45
8.54
4.45
2.45
Table 4: Mean Uncertainty (%) per Ice Concentration Class Achieved by Different Methods for Validation Data. Lowest overall uncertainty estimates are bolded .
Method
Sentinel-1
RCM
NASA Team
0.55/0.81/0.28
0.55/0.83/0.26
Deterministic Transformer
0.70/0.83/ 0.57
0.69 /0.90/ 0.48
MC Dropout
0.70/0.84/0.55
0.69/0.91/0.47
Epoch Ensemble
0.70/0.86/0.53
0.59/0.85/0.33
Bayesian Transformer
0.70 / 0.90 /0.52
0.66/ 0.92 /0.40
Table 5: Uncertainty Estimation Comparison Overall Accuracy/Ice Accuracy/Water Accuracy for Feature Detection. Best performance is bolded .
Figure 10: Local visual comparison of SIC derived from Sentinel-1 on September 4th, 2021, where (a) NASA Team SIC, (b) Sentinel-1 HV Imagery, (c) Sentinel-1 HH Imagery, (d) Deterministic Transformer SIC, (e) Mean Monte Carlo Dropout SIC, (f) Mean Epoch Ensemble SIC, and (g) Mean Bayesian Transformer SIC (our approach) .
Model
Sentinel-1
RCM
U-Net
0.69/0.80/0.58
0.58/0.71/0.44
HRNet
0.64/0.72/0.56
0.61/0.89/0.32
ConvNeXt
0.71/0.77/0.65
0.62/0.77/ 0.46
Swin Transformer
0.63/0.75/0.52
0.60/0.85/0.36
SegFormer
0.64/0.77/0.51
0.61/0.79/0.42
HR Transformer
0.76/0.86/0.65
0.65/0.91/ 0.40
Table 6: Model Architecture Comparison Overall Accuracy/Ice Accuracy/Water Accuracy for Feature Detection. Best performance is bolded .
Figure 12: Local visual comparison of SIC in MIZ region derived from Sentinel-1 on September 4th, 2021 using (a) U-Net, (b) HRNet, (c) ConvNeXt, (d) Swin Transformer, (e) SegFormer, and (f) High-Resolution (HR) Transformer (our approach) . SIC predictions are compared to (g) SAR HV and (h) NASA Team. Our approach detects small sea ice floes and ice boundaries in the MIZ while maintaining large-scale SIC patterns.
Figure 15: Local visual comparison of SIC in MIZ region derived from RCM on September 18th, 2021 using (a) U-Net, (b) HRNet, (c) ConvNeXt, (d) Swin Transformer, (e) SegFormer, and (f) High-Resolution (HR) Transformer (our approach) . SIC predictions are compared to (g) SAR HV and (h) NASA Team. Our approach detects small sea ice floes and ice boundaries in the MIZ while maintaining large-scale SIC patterns.
Dataset
Model
Metrics
Sentinel-1
HR Transformer
0.70/0.83/0.57
LoFormer Block Removed
0.67/0.80/0.53
LL1−GW Removed
0.69/0.83/0.54
RCM
HR Transformer
0.69/0.90/0.48
LoFormer Block Removed
0.63/0.87/0.39
LL1−GW Removed
0.62/0.90/0.34
Table 7: Ablation Study Overall Accuracy/Ice Accuracy/Water Accuracy for Feature Detection. Best performance is bolded .
Figure 18: Predicted Bayesian Transformer SIC accuracy vs. ASI SIC scatter plot for all 2021 validation samples. SIC estimates are aggregated into 11 bins ranging from 0% to 100% for visualization, while the R2, RMSE, and MAE are calculated using all individual SIC estimates.
Figure 19: Predicted Bayesian Transformer SIC accuracy vs. ASI SIC scatter plot for all 2025 test samples. SIC estimates are aggregated into 11 bins ranging from 0% to 100% for visualization, while the R2, RMSE, and MAE are calculated using all individual SIC estimates.
Figure 20: Training data distribution of NASA Team SIC in MIZ and Ice Pack according to the NIC Ice Chart.
High-resolution sea ice mapping using Synthetic Aperture Radar (SAR) is crucial for Arctic navigation and climate monitoring. However, operational ice charts provide only coarse, region-level polygons (weak labels), forcing automated segmentation models to struggle with pixel-level accuracy and often yielding under-confident, blurred concentration maps. In this paper, we propose a weakly supervised deep learning pipeline that fuses Sentinel-1 SAR and AMSR-2 radiometry data using a U-Net architecture trained with a region-based loss. To overcome the severe under-confidence caused by weak labels, we introduce an Analytical Logit Scaling method applied post-inference. By dynamically calculating the temperature and bias based on the latent space percentiles (2% and 98%) of each scene, we force a physical binarization of the predictions. This adaptive scaling acts as a topological extractor, successfully revealing fine-grained sea ice fractures (leads) at a 40-meter resolution without requiring any manual pixel-level annotations. Our approach not only resolves local topology but also perfectly preserves regional macroscopic concentrations, achieving a 78% accuracy on highly fragmented summer scenes, thereby bridging the gap between weakly supervised learning and high-resolution physical segmentation.
Reda Elwaradi, Julien Gimenez, Stéphane Hordoir +3
Mastère Spécialisé IA Télécom Paris, Institut Polytechnique de Paris Palaiseau, France · DTIS ONERA, Université Paris-Saclay 91120, Palaiseau, France
Accurate and automated sea ice classification is important for climate monitoring and maritime safety in the Arctic. While Synthetic Aperture Radar (SAR) is the operational standard because of its all-weather capability, it remains challenging to distinguish morphologically similar ice classes under severe class imbalance. Rather than claiming a fully validated multimodal system, this paper establishes a trustworthy SAR only baseline that future fusion work can build upon. Using the AI4Arctic/ASIP Sea Ice Dataset (v2), which contains 461 Sentinel-1 scenes matched with expert ice charts, we combine full-resolution Sentinel-1 Extra Wide inputs, leakage-aware stratified patch splitting, SIGRID-3 stage-of-development labels, and training-set normalization to evaluate Vision Transformer baselines. We compare ViT-Base models trained with cross entropy and weighted cross-entropy against a ViT-Large model trained with focal loss. Among the tested configurations, ViT-Large with focal loss achieves 69.6% held-out accuracy, 68.8% weighted F1, and 83.9% precision on the minority Multi-Year Ice class. These results show that focal-loss training offers a more useful precision-recall trade-off than weighted cross-entropy for rare ice classes and establishes a cleaner baseline for future multimodal fusion with optical, thermal, or meteorological data.
David Mike-Ewewie, Panhapiseth Lim, Priyanka Kumar
Department of Computer Science, The University of Texas Permian Basin, Odessa, Texas, USA
Antarctic sea ice concentration (SIC) forecasting is an important yet challenging task due to the coexistence of complex spatial structure, long-range temporal dependencies, and strong seasonal variability. Conventional convolution-based models are effective at capturing local spatial patterns, but often have limited ability to model long-term temporal evolution. To address these challenges, we build on a hybrid convolutional transformer forecasting framework for monthly Antarctic SIC forecasting. This framework combines convolutional encoding for spatial feature extraction with space-time factorised self-attention for SIC modelling. We further introduce two season-based mechanisms: a month-aware positional encoding that injects calendar-month information into the token representation, and a seasonal temporal bias that encourages attention to periodically related historical states. Experimental results show that the proposed framework achieves better performance than convolutional and recurrent baselines as well as ECMWF's physics-based dynamical model SEAS5 under both classification and regression metrics. Ablation studies further indicate that the seasonality-aware components provide consistent additional gains in both short- and long-horizon prediction. These results demonstrate the value of combining convolutional structures, attention mechanisms, and periodic prior information for Antarctic SIC forecasting.
Danyang Li, John Taylor, Thang Bui +1
School of Computer Science, Australian National University, Canberra, ACT 2601, Australia · Yau Mathematical Sciences Center, Tsinghua University, Beijing, China