OceanSAR-2: Micro-Sat Mission & Foundation Model

• OceanSAR-2 is an integrated system combining a micro-satellite SAR mission and a self-supervised foundation model, designed to deliver precise oceanographic measurements and robust feature extraction.
• The mission achieves ±10 cm sea-surface topography and ±5 cm/s surface current precision using advanced along-track interferometry, dual-polarization C-band SAR, and on-board FPGA compression.
• Its foundation model employs a compact Vision Transformer with self-supervised learning, yielding universal, physics-informed representations for diverse downstream oceanographic applications.

OceanSAR-2 denotes two technologically and methodologically distinct but thematically linked developments in ocean-focused Synthetic Aperture Radar (SAR): (1) a micro-satellite SAR mission architecture designed to deliver high-value oceanographic observations from < 100 kg platforms (Guerra et al., 2015), and (2) a second-generation foundation model for self-supervised representation learning on spaceborne ocean-SAR imagery, providing universal feature extraction capabilities for downstream oceanographic applications (Tuel et al., 12 Jan 2026). These efforts converge on the challenge of extracting robust, physically relevant ocean surface information from spaceborne SAR, under constraints of cost, data volume, and computational resources.

1. OceanSAR-2 Micro-Satellite SAR Mission: Concept and Objectives

The OceanSAR-2 mission concept targets oceanographic measurements of mesoscale phenomena and dynamic surface features from a micro-satellite platform. The primary objectives are:

• Sea-surface topography anomalies: Measurement with ±10 cm precision, addressing mesoscale circulation and eddy tracking.
• 2D wave spectra estimation: Frequency range 0.05–0.5 Hz, supporting offshore operations and climate modeling.
• Surface current field retrieval: Along-track interferometry (ATI) enables ±5 cm/s accuracy in current measurements.
• Monitoring coastal upwelling, fronts, and small eddies: Spatial resolution < 20 km.

Key geophysical products include radar-derived significant wave height, wave period and direction, wind-retrieval via σ₀ maps (backscatter coefficient), and sea surface height anomalies in specialized ATI modes. The platform complements large missions by leveraging cost-effective technologies without sacrificing core scientific deliverables (Guerra et al., 2015).

2. System Design: Platform, Payload, and Performance

Orbit and Coverage

• Orbit: Sun-synchronous, dusk–dawn, h ≃ 500 km, inclination i ≃ 97.4°, descending node at ≃ 10:30 LT; 8-day repeat cycle with ∼1-day revisit at mid-latitudes via ±20° off-nadir pointing.
• Coverage strategy: Ensures constant illumination (for any simultaneous optical payloads), thermal stability, and high temporal resolution for mesoscale process monitoring.

SAR Instrument Specifications

• Band: C-band, f₀ = 5.405 GHz (λ ≈ 5.55 cm), inheriting characteristics from Sentinel-1.
• Resolution: Slant-range δ_r ≈ 1.5 m, azimuth δ_az ≈ 3 m.
• Polarization: Dual-pol (HH + HV); VV + VH optional.
• Power: Peak transmit power P_tx ≃ 50 W, with a 20% duty cycle for average RF power of ~10 W.
• Antenna: Deployable planar array (1.2 m azimuth × 0.6 m range), corporate-fed patch design, gain ≈ 30 dBi, side-lobe ≤ –20 dB; stowed volume 0.016 m³, mass ~10 kg.

Power, Mass, and Platform Budgets

• Total platform mass: ~ 90 kg (< 100 kg including margins).
• Power: ~200 W end-of-life solar arrays (~1.5 m²), 100 Wh battery storage, supporting SAR and bus operations under eclipse and peak loads.
• Onboard processing: FPGA-based compression enables ≥ 4× downlink reduction, targeting raw data rates of ~600 Mb/s (~20 GB/pass).

Performance Equations

Key radar equations include the monostatic SAR SNR:

$$SNR = \frac{P_{tx}\,G^2\,\lambda^3\,\sigma_0\,A_p}{(4\pi)^3\,R^3\,k\,T_0\,B\,F_n\,L_s}$$

with typical SAR and oceanographic definitions for each parameter. Azimuth and slant-range resolutions follow:

$$\delta_{az} = \frac{L_a}{2} \qquad \delta_{r} = \frac{c}{2B}$$

Radiometric resolution and swath width are calculated as described in the system trade-off discussion (Guerra et al., 2015).

Design Trade-Offs

• Resolution–bandwidth–power: High range/azimuth resolution mandates wide bandwidth and thus increased data rate and power.
• Antenna–stowage: Large deployable antennas optimize performance but challenge mechanical complexity and mass constraints.
• Duty cycle: 200 W peak power strictly limits continuous acquisition (≈20%).
• Downlink/Processing: Data management requires aggressive on-board compression or fast X-band downlinks (~200 Mb/s).
• Thermal: Tight control (±2 °C) needed for C-band modules, constrained by bus radiator area.

Comparison with Contemporary Systems

OceanSAR-2’s architecture is positioned as a micro-sat alternative to Sentinel-1 (~230 kg, 250 W RF), with similar C-band instrumentation, and is commensurate in ambition and feasibility with ICEYE-X (~80 kg, 3 m resolution) and Capella-Space (~150 kg, 1 m resolution). The justification for feasibility is grounded in demonstrated small-sat SAR precedents for both payload and deployable antenna design (Guerra et al., 2015).

3. OceanSAR-2 Foundation Model: Architecture and Learning Paradigm

OceanSAR-2, as a universal SAR feature extractor, adopts a compact Vision Transformer (ViT) pretrained via self-supervised learning (SSL), and is specifically tailored for SAR oceanographic inputs (Tuel et al., 12 Jan 2026). The architecture comprises:

• Patch-wise encoding: SAR σ° images are partitioned into non-overlapping 16×16 pixel patches, each projected into a d-dimensional embedding space.
• Token-based representation: Positional encodings and a learnable global class token facilitate global image description.
• Transformer encoder: L stacked layers, each with multi-head self-attention (H heads) and 4d-dimensional FFN modules, using residual and layer normalization. Parameter count is ≈21 M, placing the model among efficient ViT variants.
• Physics-informed inputs: All data undergo radiometric calibration to σ°, preserving physical consistency and meaning.

As demonstrated in the original work, learned embeddings group semantically related phenomena effectively, illustrated by cosine-similarity maps of ocean scenes (Tuel et al., 12 Jan 2026).

4. Self-Supervised Pretraining and Dynamic Data Curation

Learning Objective

• DINOv2-style SSL: Training proceeds with a student–teacher dual-network configuration. Multiple “views” (random crops/augmentations) of each SAR image are processed; the student network is trained to replicate the teacher’s global (class-token) and local (patch) embeddings via cross-entropy prototype assignment.
• Mathematical objective: For input image $x$, augmented views $v$, $w$, and normalized embeddings $z_s$, $z_t$, a set of $K$ prototypes $\{c_k\}$ is learned. The global DINO loss:

$$\ell_{global}(v,w) = -\sum_{k=1}^K p_t(k|w)\,\log p_s(k|v)$$

with $p_s(k|v) \propto \exp(c_k^\top z_s(v)/\tau_s)$ (analogous for teacher).

• Local-patch loss and KoLeo regularizer: These promote fine-grained, spatially consistent representations and avoid prototype collapse.

Augmentation and SAR-specific Adaptation

All augmentations preserve the physical structure of σ° (e.g., avoiding intensity scaling). The pipeline does not inject speckle beyond inherent SAR noise, nor does it perform channel masking or additional preprocessing.

Dynamic Data Curation

• Dynamic pruning: The dataset is periodically reprioritized to maximize feature diversity, counteracting class imbalance in the SAR archive. During each pruning interval, embeddings are ranked, and the most diverse fraction is retained for ongoing SSL.
• Operational effect: Hastens convergence and increases the model’s exposure to rare but scientifically important scene types (e.g., icebergs, rain cells).

5. Benchmarking, Transfer Performance, and Evaluation Suite

The universality of OceanSAR-2’s features is validated through four systematically constructed benchmarks using Sentinel-1 Wave Mode data. Below is a summary of zero-shot (kNN) and fine-tuned (small MLP, DETR-style head) performance in comparison to other foundation models:

Benchmark	Model	Zero-shot	Fine-tuned
TenGeoP (%)	OceanSAR-2	94.0	98.5
SWH (m)	OceanSAR-2	0.52	0.40
Wspd (m/s)	OceanSAR-2	1.32	1.01
Wdir (°)	OceanSAR-2	–	16.9
Icebergs (F1)	OceanSAR-2	–	0.865

OceanSAR-2 consistently outperforms DINOv3 (RGB pretraining, 300 M parameters), TerraMind, and WV-Net on most tasks. This demonstrates strong transfer for geophysical classification, significant wave height regression (SWH), surface wind vector estimation, and object detection (icebergs), even using low-parameter fine-tuning heads.

The released benchmark suite (TenGeoP, WV-SWH, WV-wind, YOLOIB) emphasizes accessibility, reproducibility, and coverage of broad ocean-SAR phenomena, and it is structured as a “living” resource meant to evolve with future missions and data standards (Tuel et al., 12 Jan 2026).

6. Scientific, Technological, and Operational Context

The OceanSAR-2 mission and model developments fit within a broader context of transitioning ocean remote sensing from large, costly satellites to more nimble, distributed, and modular small-sat constellations and learning-based analytic pipelines:

• Feasibility: Budgets for mass, power, and antenna have been validated against ICEYE and Capella benchmarks. Deployable C-band antennas with similar characteristics have flight heritage; on-board FPGA-based SAR processing is available and effective at reducing downlink loads (Guerra et al., 2015).
• Model-application synergy: The availability of compact, SAR-native foundation models enables practically real-time, standardized downstream product derivation on edge or ground, unlocking new operational paradigms.

7. Prospects and Future Developments

Future research and development are likely to focus on:

• Incorporation of additional SAR modes and polarizations: Extension to Interferometric Wide and Extra-Wide swaths, as well as phase-exploiting objectives, is anticipated to broaden cross-mode generalization and physical inference.
• Benchmark suite expansion: Integration of operational products such as oil-spill detection, internal wave analysis, and in situ validation is planned.
• Multi-sensor synergy: Combining SAR-derived features with altimetric, SWOT, and microwave radiometry observations to enhance model robustness and product accuracy.

This suggests that OceanSAR-2 acts as a prototype for a new class of ocean remote sensing infrastructure, blending agile SAR observation with universal, physically-informed machine learning models for marine geoscience and operational oceanography (Guerra et al., 2015, Tuel et al., 12 Jan 2026).