HybridWaveNet: a deep learning approach for significant wave height retrieval

Significant Wave Height (SWH) is a critical meteorological parameter that significantly impacts oceanic navigation, coastal engineering, and disaster prediction. Accurate and timely prediction of SWH relies on precise and rapid retrieval techniques. The vast expanse of the ocean and long satellite revisit times have historically posed significant challenges for SWH observation. The advent of GNSS Reflectometry (GNSS-R) technology has facilitated the rapid acquisition of sea surface information by using reflected GNSS signals. This technology, combined with advanced deep learning methods, presents a promising approach for retrieving SWH. Currently, mainstream methodologies primarily utilize GNSS-R data from NASA's Cyclone Global Navigation Satellite System (CYGNSS) satellites. In contrast, GNSS-R data from China's Fengyun satellite series have been underutilized. In 2021, China launched the first Fengyun satellites capable of receiving and processing GNSS-R signals, with official GNSS-R data products released in 2022. Given the recent availability of this data, research leveraging these domestic datasets for SWH retrieval remains limited, and their potential has not been fully realized. This study addresses this gap by focusing on the retrieval of SWH using GNSS-R data from the Fengyun satellites, employing deep learning methodologies. We designed and evaluated multiple models tailored to different characteristics of GNSS-R data. WaveANN model retrieves SWH based on one-dimensional feature elements; WaveCNN model utilizes the Delay-Doppler Map (DDM) power delay spectrum for two-dimensional feature retrieval; and HybridWaveNet model is a hybrid model that integrates both one-dimensional and two-dimensional features for enhanced SWH retrieval. Among these, HybridWaveNet demonstrated superior performance, achieving a root mean square error (RMSE) of 0.7 meters, thereby meeting international standards for SWH retrieval accuracy.