SAR-based Ocean Observation

• SAR-based ocean observation is a technique that utilizes high-resolution, all-weather SAR data from satellites like Sentinel-1 to monitor ocean dynamics.
• The methodology involves radiometric calibration, tiling, geocoding, and self-supervised models such as WV-Net and OceanSAR-2 to automate geophysical retrieval.
• This approach supports quantitative tasks like wave height and wind speed estimation and land deformation monitoring, despite challenges such as sparse swath coverage and high computational demands.

Synthetic Aperture Radar (SAR)-based ocean observation encompasses the acquisition, processing, and analysis of microwave backscatter data collected by satellite-borne SAR sensors, principally via C-band platforms such as Copernicus Sentinel-1. The Sentinel-1 mission’s Wave Mode (WV) is engineered for global marine monitoring, capturing high-resolution, cloud- and daylight-independent imagery through systematic tiling of 20 × 20 km vignettes at 5 m ground-range resolution. Recent advancements in machine learning, particularly self-supervised foundation models like WV-Net (Glaser et al., 2024) and OceanSAR-2 (Tuel et al., 12 Jan 2026), have accelerated the transition from manual SAR image analysis to automated, scalable geophysical retrieval across vast ocean and land archives.

1. Sentinel-1 WV Mode: Acquisition and Data Characteristics

Sentinel-1’s Wave Mode (WV) is designed for high-fidelity snapshots of ocean surface dynamics. Each acquisition event, packaged in SAFE granules, yields 15–160 vignettes, typically 20 × 20 km in spatial footprint, with individual pixels representing either single-look complex (SLC: I + j Q) or gridded backscatter amplitudes. Vignettes are indexed according to their geographic footprint for rapid spatial queries and selective download (Agram et al., 2024). Data geometry features include:

• Spatial resolution: 5 m for standard products; 2.5 m for geocoded SLC stacks.
• Polarization: Primarily VV, with some HH coverage.
• Incidence angles: WV1 (near-range, 21.0°–25.0°, nominal 23.8°), WV2 (far-range, 31.0°–37.9°, nominal 36.8°).
• Temporal sampling: Repeat cycles of 12 days per satellite, phasing between S-1A and S-1B allowing 6-day revisits.
• Archive extent: ~9.9 million images (October 2015–December 2021), ~60 000 WV patches per satellite per month (Glaser et al., 2024).

C-band SAR provides uninterrupted all-weather, day/night coverage, and the mission’s open-data policy ensures accessibility for large-scale oceanographic and land deformation analysis.

2. Data Preprocessing, Calibration, and Access Mechanisms

Preprocessing approaches prioritize radiometric fidelity and efficient data handling:

• Radiometric calibration: Conversion of Level-1 GRD σ° to linear or dB scales; extreme values (e.g., < –30 dB, > 0 dB) are clipped (Tuel et al., 12 Jan 2026).
• Tiling and indexing: Vignettes are randomly cropped to 256 × 256 pixels (for machine learning pipelines) and further divided into smaller patches (e.g., ViT 16 × 16) (Tuel et al., 12 Jan 2026). In infrastructure monitoring, vignette footprints are indexed for targeted AOI retrieval, bypassing bulky SAFE granule downloads (Agram et al., 2024).
• Geocoding and coregistration: Each vignette is geocoded to a map grid (e.g., UTM) at high spatial posting using precise satellite orbits and digital elevation models. Large-window amplitude cross-correlation is used for geometric alignment across repeated acquisitions, enabling the assembly of tightly coregistered stacks suitable for InSAR analysis. Computationally, stack generation (20 km × 20 km AOI, 15–20 vignettes) completes in 3–5 minutes on a multicore node; individual vignette SLCs require 200–500 MB, with stack aggregates < 10 GB (Agram et al., 2024).
• Dynamic curation: OceanSAR-2 employs feature-based sampling to counteract overabundance of simple ocean scenes, up-weighting rarer phenomena (rain cells, icebergs) without explicit masks or manual stratification (Tuel et al., 12 Jan 2026).

No speckle filtering or semantic segmentation is typically applied in current foundation-model pipelines; data-driven feature learning handles statistical noise.

3. Foundation Models for WV SAR Imagery: Architectures and Training

Recent self-supervised learning frameworks have yielded robust foundation models for SAR-based ocean analysis:

WV-Net (Glaser et al., 2024):

• Architecture: ResNet-50 backbone with a 2-layer MLP projection head mapping to 128-dim unit-norm embeddings.
• Augmentations: SimCLR-style random crop/resize, horizontal flip, Gaussian blur, color jitter, mixup ($C=(1-m)\,A+m\,B$), random rotation, color inversion, sharpness, cutout, “no-zoom” crop (≥ 90% of image).
• Contrastive loss: InfoNCE objective over augmented view pairs, with cosine similarity:

$$\ell_{i,j} = -\log\frac{\exp\bigl(\mathrm{sim}(\mathbf{z}_i,\mathbf{z}_j)/\tau\bigr)}{\sum_{k=1}^{2N} \mathbf{1}_{k\neq i} \exp\bigl(\mathrm{sim}(\mathbf{z}_i,\mathbf{z}_k)/\tau\bigr)}$$

• Training: 8×V100 GPUs, batch size 1 024, LARS optimizer, 200 epochs, sampling a fresh 30% random subsample per epoch.

OceanSAR-2 (Tuel et al., 12 Jan 2026):

• Architecture: Vision Transformer (ViT) backbone dividing 256 × 256 images into 16 × 16 pixel patches, CLS token dimension n = 384; student and teacher networks of identical structure.
• Losses:
  • Global view invariance (DINO): $$L_\text{global} = -\sum_p p_\text{teach} \cdot \log p_\text{student}$$
  • Local patch alignment (iBOT): $$L_\text{local} = \sum_{\text{patches}} \text{KL}(p_\text{teach}^\text{patch} \parallel p_\text{student}^\text{patch})$$
  • KoLeo regularizer: $$L_\text{reg} = \lambda \cdot \sum_k [-\log(\|c_k\|)]$$
  • $$L_\text{total} = L_\text{global} + L_\text{local} + L_\text{reg}$$
• Dynamic sampling: Selection of feature-space-diverse training samples, periodic pruning to maximize informative content; σ°-calibrated input for physics-grounded normalization.

No explicit negative pairs are sampled; invariance is enforced via teacher-student distributions. The backbone contains 21 M parameters.

4. Downstream Tasks and Quantitative Performance

WV SAR foundation models support regression, classification, retrieval, and deformation analysis tasks:

Task	Metric	OceanSAR-2	WV-Net	Reference
Geophysical classification (TenGeoP)	Zero-shot kNN accuracy	94.0%	91.5%	(Tuel et al., 12 Jan 2026)
Wave height estimation (SWH)	RMSE (m)	0.52 (zero-shot), 0.40 (finetune)	0.64 (zero-shot), 0.427 (finetune)	(Tuel et al., 12 Jan 2026Glaser et al., 2024)
Wind speed estimation	RMSE (m/s)	1.32 (zero-shot), 1.01 (finetune)	1.71 (zero-shot)	(Tuel et al., 12 Jan 2026)
Air temperature estimation	RMSE (°C)	not specified	0.90 (linear probe)	(Glaser et al., 2024)
Iceberg detection (YOLOIB)	F1@IoU=0.1	0.865	not specified	(Tuel et al., 12 Jan 2026)
Phenomena classification (GOALI)	Micro-averaged AUROC	not specified	0.958 (linear probe)	(Glaser et al., 2024)
One-shot retrieval (rare classes)	Mean average precision (mAP)	not specified	AW: 0.127 vs. 0.013	(Glaser et al., 2024)

Models demonstrate strong transfer to key oceanographic tasks; OceanSAR-2’s σ° calibration and dynamic pruning further boost generalization, rare-class accuracy, and training convergence (~30% faster).

5. Deformation Monitoring and Land Applications

Sentinel-1 WV mode, originally intended for marine applications, also supports land deformation monitoring:

• Coverage: ~183,500 land-mass vignettes (2016–2024) for Africa, Australia, North and South America (Agram et al., 2024).
• Stack generation: Indexing and retrieval of repeated vignettes allow rapid assembly of coregistered, geocoded SLC stacks for AOIs ≥ 10 vignettes per track/beam.
• Workflow: Preprocessing, radiometric calibration, geocoding, and cross-correlation coregistration precede interferogram formation, phase difference extraction, coherence estimation (γ>0.3 for reliable scatterers), phase unwrapping, and time series inversion for LOS deformation retrieval.
• Performance: Case studies (Dalby mine) show annual deformation rates of ±10 mm, improved coherence relative to IW mode, and time-series error bars of ±2 mm (Agram et al., 2024).

A plausible implication is expanded use of WV-mode SAR as a proxy for commercial small-sat SAR (e.g., Iceye, Capella) and for long-baseline land infrastructure monitoring.

6. Benefits, Limitations, and Future Research Directions

Benefits:

• Spatial resolution superiority: 4–5 m vs. 20 m in IW/EW modes.
• Temporal sampling enhancement: 6–12 day revisit interval enables fine-scale tracking of dynamic phenomena.
• Open global coverage: Ensures applicability for both ocean and land.
• Annotation independence: Self-supervised models reduce reliance on costly manual labels.

Limitations:

• Sparse coverage: Narrow swath (20 km × 20 km) and ~100 km along-track separation yield incomplete spatial sampling.
• Signal-to-noise constraints: Lower SNR/NESZ (~–21 dB after improvements) leads to noisier interferometric phase.
• Processing overhead: Finer resolution increases computational burden by ~4× compared to coarser modes.
• Timing errors: Uncompensated along-track timing errors (up to 5 ms ≈ 35 m) require additional geometric correction.

Future Directions:

Models such as WV-Net and OceanSAR-2 highlight extensibility to other SAR acquisition modes, radiometric calibration strategies, and multimodal sensor fusion. Research trajectories include:

• Use of larger backbone architectures (e.g., ResNet-152, ViT variants).
• Expansion of pretraining archives.
• Integration of masked image modeling and multi-task SSL to embed richer geophysical priors.
• Operational real-time deployment in ocean stability, wave height nowcasting, and event detection workflows.
• Cross-modal adaptation for microwave radiometers and scatterometers.

Together, SAR-based ocean observation leverages advanced self-supervised learning on high-resolution, physics-calibrated imagery, supporting robust geophysical retrieval, automated monitoring, and enhanced deformation mapping over both marine and terrestrial domains (Glaser et al., 2024, Tuel et al., 12 Jan 2026, Agram et al., 2024).