New research from a collaboration of institutions, led by researchers in the Departments of Information and Computer Sciences (ICS) and Ocean and Resources Engineering (ORE) at the University of Hawaiʻi at Manoa, introduces WV-Net, the first foundation model for satellite-based synthetic aperture radar (SAR) ocean imagery. The work, led by Yannik “Nick” Glaser of UH Manoa’s ICS department, was a joint effort with scientists from the University of Washington, the University of New Hampshire, and Ifremer, the French Research Institute for Exploitation of the Sea. This groundbreaking study was recently published in the journal Artificial Intelligence for the Earth Systems.
Work related to WV-Net was previously presented at both the NeurIPS AI for Physical Sciences workshop and the American Geophysical Union’s annual meeting.
The project addresses a major challenge in oceanographic research: the cost of manually annotating the millions of images captured by missions. The team used self-supervised contrastive learning on nearly 10 million unannotated images to train WV-Net, which converts images into semantic vector embeddings for easier analysis.
WV-Net has demonstrated impressive performance on a variety of downstream tasks. In experiments, its embeddings outperformed those from models pre-trained on natural images (like ImageNet). The model’s advantages also include better scalability in data-sparse settings and increased robustness to hyperparameter choices, which significantly reduces the need for extensive computational resources and time. The WV-Net foundation model is publicly available, a valuable tool for the remote sensing community, and is expected to accelerate research into phenomena such as air-sea interactions, weather prediction, and sea ice monitoring.
