An Algorithm for One-Dimensional Wave Spectrum Retrieval from Gaofen-3 by Deep Learning

Abstract The goal of the present work was to confirm the applicability of one-dimensional wave spectrum retrieval from Gaofen-3 ( GF-3 ) synthetic aperture radar (SAR) using deep learning techniques. A total of 11 300 images were acquired in wave (WV) mode, and 1000 images have been employed in quad-polarized stripmap (QPS) mode during the period from 2017 to 2022. To simulate wave spectra that were collocated with GF-3 images, a third-generation numerical model, WAVEWATCH-III (WW3), was employed. Validation of significant wave heights (SWHs) hindcasted by WW3 against the operational products from Haiyang-2 (HY-2) altimeters during March–June 2019 in the China Seas yielded a 0.44-m root-mean-square error (RMSE), a correlation ( r ) of 0.91, and a 0.16 scatter index (SI). The basic deep learning method employed for SAR wave spectrum retrieval was based on three deep learning methods [multilayer perceptron (MLP), residual networks (ResNet), and convolutional neural networks (CNNs)], which were trained on the 11 300 WV images. SAR intensity spectrum in copolarization [vertical–vertical (VV) and horizontal–horizontal (HH)] and modulation transfer functions (MTFs) corresponding to three modulations (i.e., hydrodynamic, tilt, and velocity bunching) were treated as inputs in the training procedure. An MLP-based algorithm has the best inversion results. The retrievals by the MLP were compared with the collocated wave spectra from the Surface Wave Investigation and Monitoring (SWIM). Statistical analysis yielded a correlation coefficient (Cor) and a squared error (Err) of the wave spectrum of 0.79 and 1.69, respectively, and the RMSE of SWH was 0.35 m, with a 0.98 r value and 0.13 SI. Similarly, the RMSE of SAR-derived SWHs from 700 QPS images was 0.41 m, with a 0.95 r and a 0.28 SI validated against the altimeters on board HY-2. As validated against National Data Buoy Center (NDBC) buoys, the RMSE was 1.14 s, with an r of 0.23 and an SI of 0.16. Collectively, the proposed MLP-based algorithm demonstrated good performance in retrieving the one-dimensional wave spectrum from GF-3 SAR images.