Space-Frequency Fusion Dual-Branch Convolutional Neural Networks for Significant Wave Height Retrieval From GF-3 SAR Data

Abstract

Deep learning in synthetic aperture radar (SAR) sea state retrieval is becoming increasingly prevalent. In current studies, convolutional neural networks (CNNs) are widely employed to extract either deep space features from normalized radar cross section (NRCS) of SAR images or deep frequency features from SAR spectra, with some studies combining artificially designed scalar features to retrieve significant wave height (SWH). When the quality of ocean wave imaging is poor, it becomes challenging for CNN models to extract useful features from a single space or frequency domain, and the scalar features are insufficient to describe the complex relationships within the data, thereby limiting the retrieval accuracy of the models. To harness the space and frequency domain information in SAR data effectively and acquire more expressive fusion features, we propose a space-frequency fusion dual-branch CNN (DB-CNN) model. The model separately extracts deep space features from NRCS of SAR images and deep frequency features from SAR image spectra. By employing the space-frequency feature cross layer (SFFCL) and the gated feature fusion layer (GFFL), it enhances and fuses space-frequency features, thereby achieving more accurate SAR SWH retrieval. Most retrieval models based on GF-3 data primarily focus on wave mode data, with limited utilization of data from other imaging modes. To fully leverage the diverse imaging modes of GF-3 data, this study collects GF-3 data across various imaging modes and establishes matched datasets with the fifth-generation European Centre for Medium-Range Weather Forecasts reanalysis (ERA5) and buoy for model training and evaluation. Consequently, our model exhibits applicability across diverse imaging modes and superior performance under different sea states. In addition, ablation experiments are conducted to evaluate the importance of the SFFCL and GFFL modules.