A Deep Learning-Based Paradigm for Retrieving Ocean Wave Spectra from Spaceborne SAR

The ocean wave spectrum is an essential parameter to fully describe the dynamic ocean waves. Although the spaceborne Synthetic Aperture Radar (SAR) is a potential instrument to measure the ocean wave spectrum, its retrieval is intricated by the complicated non-linear SAR-wave imaging mechanisms. This study introduced a deep learning (DL) method to address the ocean wave spectrum retrieval tasks from SAR images. The DL-based framework modifies the classical CNN architecture according to the multiple outputs of the target parameters and imposes a soft constraint based on the characteristics of the ocean wave spectrum. Trained and evaluated by the measurements of in-situ buoys deployed in the global ocean, the proposed model achieved root-mean-square errors of 0.53 m for significant wave heights and 1.24 s mean wave periods, demonstrating a generalized and robust performance. Our study represents a significant breakthrough in the longstanding challenge of SAR retrievals in ocean remote sensing.