Comparison of BPNN and Dual-Branch CNN for Significant Wave Height Estimation From Polarimetric Gaofen-3 SAR Wave Mode Data

The present study utilizes the backward propagation neural network (BPNN) and the dual-branch convolutional neural network (DB-CNN) algorithms to construct models for estimating significant wave height (SWH) from polarimetric Gaofen-3 SAR wave mode data, using a dataset of 11 164 images that are collocated with SWH from the European Centre for Medium-Range Weather Forecasts fifth generation reanalysis (ERA5). The models are assessed and compared across nine polarizations [vertical–vertical (VV), horizontal–horizontal (HH), RL, vertical–horizontal (VH), horizontal-vertical (HV), RR, 45° linear, RV, and RH] and various sea states using the SAR-ERA5 test samples as well as buoy and altimeter SWH observations. The results demonstrate the robust performance of BPNN models, with RMSEs around 0.30–0.32 m on SAR-ERA5 test data, 0.32–0.48 m on buoy data, 0.40–0.48 m on Jason-3 data, and 0.36–0.42 m on SARAL data. By comparison, the DB-CNN models, which additionally include two-dimensional (2-D) image spectra as input, only exhibit improved performance at VV, 45° linear, and RL polarizations, while showing negligible improvement at HH, RV, and RH polarizations and a notable degradation at VH, HV, and RR polarizations. Furthermore, the DB-CNN models generally fail to improve the overestimation (underestimation) in low (high) seas, and they even aggravate the overestimation (underestimation) under most polarizations. Additionally considering the heightened complexity, increased vulnerability to overfitting, and training times that are more than 200 times longer, the use of complex deep learning network structures to incorporate 2-D spectral information appears to be operationally limited for SAR SWH retrieval.