Weakly Supervised Polar Low Segmentation in Sentinel-1 SAR Imagery

Andrea Federici; Jakob Grahn; Giacomo Boracchi; Filippo Maria Bianchi

cs.CVAug 14, 2026

Weakly Supervised Polar Low Segmentation in Sentinel-1 SAR Imagery

Authors: Andrea Federici, Jakob Grahn, Giacomo Boracchi, Filippo Maria Bianchi

Organizations: Dept. of Mathematics and Statistics, UiT the Arctic University of Norway · NORCE, The Norwegian Research Centre AS · DEIB, Politecnico di Milano

Abstract

Polar lows are intense maritime cyclones that form rapidly at high latitudes. Deep learning can detect them in Synthetic Aperture Radar (SAR) imagery, but pixel-level segmentation remains an open challenge. No pixel-level masks are available for training, and a polar low's extent is inherently subjective, with diffuse boundaries that even experts delineate inconsistently. We propose Constrained Region Erasing with Soft Targets (CREST), a Weakly Supervised Semantic Segmentation (WSSS) framework that generates masks solely from image-level labels. Our approach builds on Adversarial Erasing (AER), which iteratively mines discriminative regions, erases them, and retrains a classifier to reveal complementary cues that become pseudo-labels for segmentation. However, standard AER also collects irrelevant background features, degrading pseudo-label quality. CREST addresses this with (i) a Constrained Ordinal Region Expansion (CORE) module that encodes the spatial-connectedness prior of polar lows, constraining region expansion from a high-confidence seed, and (ii) a Dynamic Bootstrapping (DB) loss that treats the mining order as a proxy for label reliability, attenuating supervision from noisier, later-mined regions. On Sentinel-1 SAR data, CREST follows the cyclone structure more closely than standard AER, and returns a multi-class rather than binary mask whose classes indicate the reliability assigned to each region. We further evaluate on BUS-UCLM breast ultrasound and PASCAL VOC person data, whose targets satisfy the same connectedness prior but come with the dense masks the SAR data lacks. On both datasets, CREST performs better than the equivalent AER pipeline under identical settings.

Figures & tables

Panel of Figure 1 — Fig. 1 : Examples of Sentinel-1 SAR imagery from our dataset. (a) A positive sample showing a polar low. (b) A negative sample showing a typical maritime background without cyclonic activity. The black regions are “no-data” areas in the SAR images.

Panel of Figure 2 — Fig. 2 : Spurious background activations during late AER iterations. Each row shows a different positive polar low SAR image. Left: original input. Right: Grad-CAM on the image after erasing mined regions. With most of the target regions removed, the classifier starts producing activations on the background, injecting noise into the aggregated pseudo-label.

Panel of Figure 3 — Fig. 3 : Impact of CORE on region expansion. The panels display the Grad-CAM saliency maps at each step of the mining process. Once activated, CORE introduces the yellow local envelope that limits subsequent mask growth. In Iteration 3, the classifier produces a strong activation on land outside this envelope; the activation is therefore excluded from the next mask update, as shown in Iteration 4.

Fig. 4 : AER+CORE pipeline and pseudo-label construction. Top: the proposed iterative train $\rightarrow$ extract $\rightarrow$ erase loop. We retrain the classifier after each erasing step and recompute Grad-CAM on the updated image. Once activated, CORE restricts the next erasure to a neighborhood (yellow envelope) around the previously mined region. Bottom: Grad-CAM heatmaps from successive iterations and the resulting multi-class pseudo-label, where the class records when each region was mined.

Panel of Figure 5 — Fig. 5 : Grad-CAM and Integrated Gradients on a Sentinel-1 SAR image. In this example, both methods highlight the most discriminative cues but not the full spatial extent of the cyclone.

Panel of Figure 6 — Fig. 6 : Zero-shot SAM outputs for one Sentinel-1 polar low SAR image. With grid- and box-based prompts, the model highlights imaging artifacts or misses the diffuse cyclone structure.

Fig. 7 : Overview of the proposed two-stage pipeline. Stage 1 generates mining-order pseudo-labels from image-level labels via AER with the CORE module. Stage 2 trains a segmentation network using these pseudo-labels and a Dynamic Bootstrapping strategy to handle label uncertainty.

Fig. 8 : Effect of dynamic bootstrapping under the same background-biased model prediction

\mathbf{p}=[0.9,0.08,0.02]

\gamma(e)=0.5

. The bootstrapped target mixes the hard pseudo-label

\mathbf{y}

with the detached prediction

\operatorname{sg}(\mathbf{p})

. For the high-reliability case,

\rho_{s}=0.9

and

\lambda_{s}(e)=0.05

, corresponding to CE gradient

\times\,0.95

(

5\%

attenuation); for the low-reliability case,

\rho_{s}=0.2

and

\lambda_{s}(e)=0.40

, corresponding to CE gradient

\times\,0.60

(

40\%

attenuation).

Panel of Figure 9 — Fig. 9 : Visual comparison of segmentation masks for maritime mesocyclones in SAR images. First column: input images. Second: segmentation network trained with single-pass Grad-CAM masks, which concentrate mainly around the cyclone eye in these examples. Third: standard AER extends the prediction to more peripheral regions, but can also incorporate background. Fourth: CREST predictions, with colors encoding mining order.

Fig. 10 : Illustrative case: a single SAR acquisition overlaid with auxiliary meteorological fields.

Method	BUS						VOC
	Macro		Foreground				Macro		Foreground
	IoU $\uparrow$	Dice $\uparrow$	IoU $\uparrow$	Dice $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	IoU $\uparrow$	Dice $\uparrow$	IoU $\uparrow$	Dice $\uparrow$	Precision $\uparrow$	Recall $\uparrow$
Grad-CAM	57.2	64.2	17.7	30.1	62.4	19.8	51.8	54.8	5.7	10.7	39.8	6.2
Standard AER	60.2	68.6	24.3	39.1	45.0	34.6	65.4	74.4	33.5	50.2	38.8	71.3
CREST	64.8	73.9	32.7	49.3	58.7	42.5	67.4	76.6	37.5	54.6	41.1	81.0

TABLE I : Best test performance on BUS and VOC . For the iterative methods, we report the mining iteration that achieved the best Macro IoU; Tables IV and V report all iterations. Best results per metric are highlighted in bold.

Panel of Figure 12 — Fig. 11 : BUS qualitative examples. The predicted multi-class masks preserve the mining order from AER : warmer colors (yellow) indicate earlier-mined, more discriminative regions, while cooler colors (blue) indicate regions discovered later. In the displayed examples, the earliest tiers fall on the lesion core and later tiers extend toward its margin, and the ground-truth masks are shown for visual comparison.

Panel of Figure 13 — Fig. 12 : VOC qualitative examples for the person class. Colors in the multi-class output indicate mining order: warmer colors are earlier discoveries and cooler colors are later discoveries. In these examples, early discoveries tend to emphasize discriminative features such as the face and upper body, while later discoveries extend to other body regions. The binarized output is used for evaluation against the ground-truth mask.

Method	BUS				VOC
	Macro IoU $\uparrow$	Macro Dice $\uparrow$	IoU c=1 $\uparrow$	Dice c=1 $\uparrow$	Macro IoU $\uparrow$	Macro Dice $\uparrow$	IoU c=1 $\uparrow$	Dice c=1 $\uparrow$
Standard AER	60.2	68.6	24.3	39.1	65.4	74.4	33.5	50.2
AER+CORE	61.4	70.0	26.4	41.8	66.5	75.6	35.6	52.5
AER+CORE+DB	64.8	73.9	32.7	49.3	67.4	76.6	37.5	54.6

TABLE II : Progressive ablation results on the BUS and VOC test sets, where each row shows the effect of adding components to the baseline AER pipeline. For each method, we report the mining iteration that achieved the best Macro IoU.

Fig. 13 : Macro IoU over retained adversarial-erasing iterations on the BUS test set. Standard AER degrades after iteration 2, whereas the AER+CORE and full CREST (AER+CORE+DB) trajectories vary less over the later iterations.

Fig. 14 : Discrete precision–recall trajectories on the VOC test set as additional mining iterations are retained. Numbers indicate the iteration index. In the legend, AER+CORE+DB is the full CREST configuration.

Method	BUS		VOC
	Macro IoU $\uparrow$	IoU c=1 $\uparrow$	Macro IoU $\uparrow$	IoU c=1 $\uparrow$
Standard AER	57.5	18.6	65.7	34.4
CREST	59.6	22.7	68.4	38.8

TABLE III : Ablation results using Grad-CAM++ attribution on the BUS and VOC test sets. For each method, we report the mining iteration that achieved the best Macro IoU (standard AER at

t=2

and CREST at

t=3

on BUS ; both at

t=6

on VOC ).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 18 — Fig. 15 : Mosaic of inputs and predicted masks for test images. Warmer colors (red/yellow) denote earlier-mined regions; cooler colors (blue/cyan) denote later-mined regions.

Panel of Figure 19 — Fig. 16 : Qualitative consistency under geometric transformations. Left: original image. Right: predictions for the original and three transformed versions.

Method	Macro IoU $\uparrow$	Macro Dice $\uparrow$	IoU c=1 $\uparrow$	Dice c=1 $\uparrow$	Precision c=1 $\uparrow$	Recall c=1 $\uparrow$
Grad-CAM	57.2	64.2	17.7	30.1	62.4	19.8
Standard AER
Iteration 0	49.2	51.0	2.0	3.9	70.6	2.0
Iteration 1	57.6	64.7	18.3	31.0	67.0	20.1
Iteration 2	60.2	68.6	24.3	39.1	45.0	34.6
Iteration 3	58.0	66.5	22.0	36.0	29.6	46.1

Appendix

TABLE IV : Per-iteration results on the BUS test set. For the iterative methods, bold marks the best value per column across mining iterations.

Method	Macro IoU $\uparrow$	Macro Dice $\uparrow$	IoU c=1 $\uparrow$	Dice c=1 $\uparrow$	Precision c=1 $\uparrow$	Recall c=1 $\uparrow$
Grad-CAM	51.8	54.8	5.7	10.7	39.8	6.2
Standard AER
Iteration 0	52.2	55.5	6.3	11.9	60.8	6.6
Iteration 1	60.3	68.0	22.8	37.1	44.3	31.9
Iteration 2	63.7	72.4	29.8	45.9	42.8	49.5
Iteration 3	65.1	74.0	32.8	49.4	40.5	63.3

Appendix

TABLE V : Per-iteration results on the VOC test set (person class). For the iterative methods, bold marks the best value per column across mining iterations.

Explore similar work

Mar 13, 2026

Analytical Logit Scaling for High-Resolution Sea Ice Topology Retrieval from Weakly Labeled SAR Imagery

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

Remote Sensing Image Segmentation Weakly Supervised Image Segmentation Topology-Preserving Image Segmentation Model Calibration

Ask PDF

May 27, 2026

A Deep Learning Iterative Framework for Sentinel-1 Stripmap Enhancement Based on Azimuth Doppler Decomposition

Synthetic Aperture Radar (SAR) imagery enables all-weather, day-and-night Earth observation; however, it remains difficult to interpret due to speckle noise and other intrinsic imaging artifacts. Sentinel-1 (S1) constitutes one of the most widely used spaceborne SAR missions, offering systematic global coverage, high temporal resolution, dual-polarization imaging, and free data availability. Among S1 modes, Stripmap (SM) provides the highest resolution, yet speckle noise and spatial constraints often hinder applications requiring finer spatial detail. This motivates the need for effective image enhancement strategies. In this work, we propose a self-supervised enhancement framework for S1 SM imagery based on azimuth subaperture decomposition. The method exploits the physical consistency between subaperture reconstructions and the corresponding full-aperture image to generate paired training data without external sensors, simulated ground truth, or multi-temporal stacks. The proposed framework integrates single- and multi-frame learning and incorporates an iterative inference scheme that progressively refines image quality. Experiments on real S1 SM data show that the proposed approach consistently outperforms the widely adopted self-supervised deep learning baseline MERLIN, in terms of PSNR and SSIM, while MERLIN attains higher ENL, highlighting a trade-off between structural fidelity and speckle smoothing. Overall, the results demonstrate that subaperture-based supervision provides a physically grounded, reproducible, and operationally viable approach for SAR image enhancement using S1 data. It is worth noting that the proposed approach can be extended to other SAR platforms, polarizations, and acquisition modes.

Juan Francisco Amieva, Christian Ayala, Roberto Del Prete +1

Tracasa Instrumental S.L. · European Space Agency · Public University of Navarre

Image Enhancement Self-Supervised Learning Remote Sensing Synthetic Aperture Radar Image Denoising

Ask PDF

Sep 14, 2026

Multi-Label Proportion Learning for Sea-Ice Type Prediction

Sea-ice type prediction is important for climate monitoring, maritime navigation, and decision-making in polar regions. The main source of label data for this task is the ice chart, produced manually by ice analysts who interpret satellite imagery to delineate ice zones into polygons. Although ice charts are valuable, their production is labor-intensive and expensive, motivating recent efforts to automate the process using deep learning. However, deep learning models require patch-level (or pixel-level) label data for training, while ice charts provide only polygon-level annotations. As a workaround, supervised approaches often create approximate patch-level labels from polygon-level ice chart labels by assigning each sample the dominant ice type of its parent polygon. This approach enables supervised training but creates an ill-posed learning problem with intrinsically approximate solution. In this paper, we redefine sea-ice type prediction as a weakly supervised multi-label proportion learning problem to be able to directly use the polygon-level ice chart labels and avoid unnecessary label approximation for improved prediction accuracy. To address this problem, we propose a two-module framework where first Multiple Instance Learning (MIL) is used for water--ice classification, and then a multi-label proportion learning (MLPL) is introduced for ice-type composition prediction. We further extend this framework with a multimodal model that integrates SAR imagery with AMSR2 brightness temperatures and ERA5 reanalysis data through modality-guided auxiliary regularization. Evaluated on the AI4Arctic dataset, the SAR-only model reduces MAE by 14.5% and more than doubles mean ice-class F1 over the best supervised baseline. The multimodal model further reduces MAE by 21.5% and raises mean F1 by 41.2% over the SAR-only model, and by 52.7% over the supervised multimodal baseline.

Samira Alkaee Taleghan, Younghyun Koo, Andrew P. Barrett +1

University of Colorado Denver Denver, Colorado, USA · National Snow and Ice Data Center (NSIDC), CIRES, University of Colorado Boulder Boulder, Colorado, USA

Label Distribution Learning Remote Sensing Weakly Supervised Learning

Ask PDF