We present an application of generative adversarial networks (GANs) to reconstruct the sea level of the North Sea using a limited amount of data from tidal gauges (TGs). The application of this technique, which learns how to generate datasets with the same statistics as the training set, is explained in detail to ensure that interested scientists can implement it in similar or different oceanographic cases. Training is performed for all of 2016, and the model is validated on data from 3 months in 2017 and compared against reconstructions using the Kalman filter approach. Tests with datasets generated by an operational model (“true data”) demonstrated that using data from only 19 locations where TGs permanently operate is sufficient to generate an adequate reconstruction of the sea surface height (SSH) in the entire North Sea. The machine learning approach appeared successful when learning from different sources, which enabled us to feed the network with real observations from TGs and produce high‐quality reconstructions of the basin‐wide SSH. Individual reconstruction experiments using different combinations of training and target data during the training and validation process demonstrated similarities with data assimilation when errors in the data and model were not handled appropriately. The proposed method demonstrated good skill when analyzing both the full signal and the low‐frequency variability only. It was demonstrated that GANs are also skillful at learning and replicating processes with multiple time scales. The different skills in different areas of the North Sea are explained by the different signal‐to‐noise ratios associated with differences in regional dynamics.

中文翻译：

---

利用海岸数据和生成对抗网络重建北海流域至整个海平面的变化

我们提出了使用对抗性网络（GAN）来利用潮汐仪（TGs）的有限数据重建北海海平面的应用。详细说明了该技术的应用，该技术学习了如何使用与训练集相同的统计数据生成数据集，以确保感兴趣的科学家可以在相似或不同的海洋学案例中实施该数据集。在2016年全年进行了培训，并使用2017年三个月的数据对模型进行了验证，并将其与使用卡尔曼滤波方法的重建进行比较。使用运行模型生成的数据集进行的测试（“真实数据”）表明，仅使用TG永久运行的19个位置的数据就足以对整个北海的海面高度（SSH）进行适当的重建。从不同来源学习时，机器学习（ML）方法似乎很成功，这使我们能够向网络提供来自TG的真实观测结果，并生成整个流域SSH的高质量重构。在训练和验证过程中使用训练和目标数据的不同组合进行的单独重建实验表明，如果数据和模型中的错误未得到适当处理，则与数据同化相似。当仅分析全信号和低频可变性时，所提出的方法表现出良好的技巧。事实证明，GAN还具有多种时间尺度的学习和复制过程的技能。