WV-Net: Self-Supervised SAR Foundation Model

• WV-Net introduces contrastive self-supervised pretraining on ~10M unannotated 5m resolution SAR patches, bypassing manual annotation for ocean wave mode imagery.
• The model employs a truncated ResNet-50 backbone with a two-layer MLP projection head and enhanced data augmentation to preserve geophysical signal integrity.
• Quantitative evaluations reveal reduced RMSE in wave height and air temperature regression and superior retrieval performance compared to ImageNet-pretrained models.

WV-Net is a self-supervised foundation model developed for ocean wave mode (WV-mode) synthetic aperture radar (SAR) imagery acquired by the Copernicus Sentinel-1 mission. Pretrained on approximately 10 million unannotated 20 × 20 km image patches at 5-meter resolution, WV-Net produces compact semantic embeddings capable of supporting various geophysical applications, including wave height estimation, atmospheric and oceanic variable regression, geophysical phenomenon classification, and unsupervised image retrieval. Through contrastive self-supervised learning—a method that does not require labeled data—WV-Net addresses persistent bottlenecks imposed by the need for manual annotation in machine learning for large-scale satellite SAR datasets, establishing improved performance and data efficiency over conventional supervised computer vision pretraining (Glaser et al., 2024).

1. SAR WV-Mode Data and Preprocessing

WV-Net utilizes the Sentinel-1 (S-1) mission’s WV-mode SAR data, which consists of about 60,000 scenes per satellite per month, totaling approximately 9.9 million patches over 165 months. Each scene is a 20 km × 20 km patch, natively sampled at 5 m (roughly 4000 × 4000 pixels). The preprocessing pipeline comprises three compulsory steps:

• Incidence-angle normalization: The C-band backscatter ($\sigma_0$) is mapped to a sea surface roughness (SSR) proxy using the CMOD5N geophysical model at a 10 m/s reference wind speed and at a relative look angle, thereby removing first-order wind and incidence effects.
• Downscaling: Application of a 10 × 10 boxcar (50 m) smoothing followed by 10× subsampling yields a final ~400 × 400 pixel image.
• Intensity normalization: SSR values are clipped to the 1st and 99th percentiles, linearly mapped to $[0, 255]$, and saved as 8-bit grayscale.

To leverage standard computer vision pipelines, each grayscale image is stacked into three identical channels. In the contrastive learning framework, augmentation policies extend beyond those used in SimCLR (random crop/zoom, horizontal flip, color jitter, Gaussian blur) to include: Mixup (mixing weights 0.1–0.4), random rotation (±170°), random color inversion, and random sharpness adjustment. These policies were specifically selected to encourage embedding invariance to intensity, orientation, and local textural variability while preserving physically meaningful patterns (Glaser et al., 2024).

2. Model Architecture and Training Protocol

WV-Net is based on a standard ResNet-50 backbone, truncated at the global-average pooling layer (outputting a 2048-dimensional feature vector $f(x)$), followed by a two-layer MLP projection head—identical to the SimCLR design—mapping 2048$\rightarrow$2048 (ReLU)$\rightarrow$128. During self-supervised pretraining, the contrastive loss is applied to the projection output ($z = g(f(x))$, 128 dimensions); during downstream transfer, only the backbone’s 2048-dimensional $f(x)$, after L$_2$ normalization, is used as the semantic embedding.

Contrastive pretraining follows the SimCLR methodology: a batch of $N$ original images yields $2N$ augmented “views.” Positive pairs originate from augmentations of the same image, and the remaining pairs are negatives. A normalized cosine similarity is computed as

$$\mathrm{sim}(u,v) = \frac{u^\mathsf{T}v}{\|u\|\,\|v\|}$$

with a temperature parameter $\tau = 0.1$ in the contrastive loss. Training was executed for 200 epochs with a global batch size of 1024 (8×V100 GPUs), an initial learning rate of ~1.2 (warmup + decay), $10^{-6}$ weight decay, and the LARS optimizer. Training duration was approximately 12 days (Glaser et al., 2024).

3. Semantic Representation and Embedding

Upon completion of pretraining, each image is mapped by the backbone to a L$_2$-normalized 2048-dimensional embedding. These embeddings effectively capture multi-scale textural and pattern information, encompassing phenomena such as ocean wave fields, mesoscale and submesoscale features, convective cells, oil slicks, and atmospheric/oceanic fronts. This representation is suitable for a diverse set of downstream tasks requiring semantic discrimination of physical oceanic and atmospheric processes, without the need for extensive labeled datasets.

4. Quantitative Evaluation and Downstream Applications

WV-Net was benchmarked against (a) an ImageNet-pretrained ResNet-50 backbone and (b) a WV-mode SimCLR baseline trained without the four additional augmentations, using several geophysical and computational tasks:

Task	Metric	WV-Net Result	ImageNet Result
Significant Wave Height Regression	RMSE (m), linear probe	0.50	0.60
Near-Surface Air Temperature Regression	RMSE (°C), linear probe	0.90	0.97
Multilabel Geophysical Classification	Micro-averaged AUROC, linear probe	0.96	0.95
Multilabel Geophysical Classification	Micro-averaged AUROC, full finetune	0.939	0.931
Unsupervised Image Retrieval (rare classes)	mAP (e.g., atmospheric gravity waves, ship wakes)	up to 0.90	0.11

• For wave height regression (200,000 SAR-altimeter colocations), WV-Net embeddings yielded a 17% reduction in RMSE over ImageNet. Similar improvements were observed in SST/air-temperature regression (76,000 samples, RMSE decrease from 0.97°C to 0.90°C).
• On the GOALI dataset (16,400 expertly labeled samples, 15 multilabel phenomena), WV-Net achieved micro-AUROC values of 0.96 (linear probe) and 0.939 (finetune), each exceeding the corresponding ImageNet baselines.
• In one-shot k-NN unsupervised retrieval of rare geophysical classes (e.g., gravity waves, ship wakes), mean average precision of up to 0.90 far surpassed the ImageNet baseline mAP of 0.11, representing 30–80% relative improvement on rare class retrieval (Glaser et al., 2024).

5. Data Efficiency and Scaling Analysis

Performance as a function of available labeled data demonstrates that WV-Net provides substantially greater data efficiency in label-scarce regimes. On the multilabel GOALI dataset, with as few as 1,000 labeled examples, an MLP classifier on WV-Net embeddings achieves micro-AUROC exceeding 0.92, while the ImageNet model remains below 0.90. This suggests that domain-specific self-supervised pretraining lowers the annotation requirements for competitive model performance and accelerates methodological development within the SAR geoscience community (Glaser et al., 2024).

6. Practical Implications and Prospects

Due to its strong semantic discrimination across multiple scales and geophysical processes, WV-Net supports a broad spectrum of scientific applications, including studies of ocean surface dynamics, marine atmospheric boundary-layer structure, swell monitoring, and rapid assessment of extreme events (e.g., small-scale eddies, atmospheric gravity waves, submesoscale fronts). The foundational architecture enables fast adaptation with only a linear output head, requiring modest computational resources.

Planned future directions include scaling to larger model architectures (ResNet-152, vision transformers), extended pretraining durations, refinement for the detection of kilometer- and sub-kilometer-scale phenomena, and adaptation to alternate SAR and microwave remote-sensing modalities (e.g., Interferometric Wide Swath, SWIM, scatterometers). A plausible implication is broader adoption of SAR foundation models for unsupervised and semi-supervised satellite remote-sensing applications (Glaser et al., 2024).

7. Position Within Geoscientific Machine Learning

WV-Net exemplifies the use of contrastive self-supervised learning for building domain-specific foundation models. Compared to supervised pretraining on natural images, its approach permits the exploitation of vast unannotated satellite repositories and yields transferable representations for various downstream geoscientific inference tasks, including in contexts with limited labels. This foundation model paradigm is increasingly prominent across scientific domains seeking scalable and annotation-efficient workflows with robust, generalizable symbolic representations (Glaser et al., 2024).