High-resolution sea surface height reconstruction method based on deep learning gradient constraint embedding method
Abstract
Sea surface height (SSH) data are widely used to study ocean phenomena. High-resolution SSH data helps to explore finer scale ocean phenomena in response to many applications in the field of ocean engineering, such as ship routing, search and rescue, oil spills, or offshore platform site selection. Accurately reconstructing highresolution SSH data from low-resolution SSH data products and focusing on the reconstruction of gradient information (information on ocean phenomena) has been a key challenge in ocean engineering. In this study, we propose a model based on generative adversarial networks that combines a multi-level residual networks with dense connections and gradient blocks (MRDG-GAN), which improves the spatial resolution of SSH data from 1/ 4 degrees to 1/12 degrees. The gradient blocks and gradient loss function proposed in the model are used to ensure the generation of high accuracy and high-resolution sea surface with more gradient information. Comparative experiments have shown that MRDG-GAN reconstructs more realistic and accurate high-resolution SSH data than the most advanced GAN models currently available. The results of comparative experiments and hyperparameter experiments indicate that the hyper-parameters of MRDG-GAN have reached their optimal level, which can accurately improve the resolution of SSH data.