WV-Net: A Foundation Model for SAR Ocean Satellite Imagery

Abstract The European Space Agency’s Sentinel-1 (S-1) satellite mission has captured more than 10 million images of the ocean surface using C-band synthetic aperture radar (SAR WV mode). While machine learning is a promising approach for detecting and quantifying various geophysical signatures in these images, scientists are limited by the cost of manual data annotation for any particular task. We propose to use contrastive self-supervised learning on the full archive of unannotated WV-mode images to train a semantic embedding model named WV-Net. In experiments, we show that WV-Net embeddings outperform those from models that were pretrained with natural images (ImageNet) on four downstream tasks: multilabel classification [0.96 average area under the receiver operating characteristic (AUROC) vs 0.95], wave height regression [0.50 root-mean-square error (RMSE) vs 0.60], near-surface air temperature regression (0.90 RMSE vs 0.97), and unsupervised image retrieval [0.41 class-averaged mean average precision (mAP) vs 0.37]. WV-Net embeddings also scale better in data-sparse settings, and fine-tuned WV-Net models are more robust to hyperparameter choices. The WV-Net foundation model is publicly available and can be adapted to a variety of data analysis and exploration tasks in geophysical research. Significance Statement Satellite imagery of the ocean from the Sentinel-1 (S-1) mission contains global and temporal information about ocean waves, atmospheric patterns, biological signatures, sea ice, and ships. However, the dataset is currently not fully utilized due to the size and complexity of the data. Our WV-Net foundation model makes this large dataset more accessible to researchers by providing a pretrained model that converts images into semantic vector embeddings, making the dataset amenable to analysis using advanced data-driven methods.