Rapid and precise forecasting of storm surge in coastal regions is crucial for ensuring safety of coastal communities’ life and property. Yet, learning a data-driven forecasting model from observation data such as gauges and post-event reconnaissance remains challenging, due to the observation data scarcity and the real-world complexity. Recently, deep learning has received increasing attention, but existing deep learning approaches solely focus on individual site scenarios, ignoring the value of information contained in neighboring sites’ observations. In this study, we propose to integrate graph neural networks (GNN) and gated recurrent unit (GRU) to capture the spatial and temporal storm surge dependencies across multiple observation stations. GNN provides the unique capability to model non-Euclidean complex spatial relationship across observation stations, while GRU enhances the data efficiency of temporal dependency modeling. To account for the effect of complex coastline topography, the Liang–Kleeman information flow theory is employed to establish a causal-inference based graph edge scheme connecting multiple observation stations. The Causal-inference based Spatio-Temporal Graph Neural Network (CSTGNN) were trained and evaluated on 13-year observation data from 4 observation stations along Florida coastline. Experiments affirm the competence of CSTGNN, which outperformed six commonly used competitive baselines across different metrics and observation stations, under lead times up to six hours. Furthermore, benefits of capturing the spatial dependency and leveraging causal inference are also comprehensively examined. To conclude, we believe that this novel spatio-temporal forecasting framework can result in enhanced coastal resilience by its improved storm surge forecasting capability.

中文翻译：

---

利用稀缺观测数据推进风暴潮预测：基于因果推理的时空图神经网络方法

快速、准确地预报沿海地区风暴潮对于保障沿海人民生命财产安全至关重要。然而，由于观测数据的稀缺性和现实世界的复杂性，从观测数据（例如仪表和事件后侦察）学习数据驱动的预测模型仍然具有挑战性。近年来，深度学习受到越来越多的关注，但现有的深度学习方法仅关注单个站点场景，忽略了邻近站点观测中包含的信息价值。在本研究中，我们建议集成图神经网络（GNN）和门控循环单元（GRU）来捕获跨多个观测站的空间和时间风暴潮依赖性。 GNN 提供了对观测站之间的非欧几里得复杂空间关系进行建模的独特功能，而 GRU 则增强了时间依赖性建模的数据效率。为了考虑复杂海岸线地形的影响，采用梁-克利曼信息流理论建立连接多个观测站的基于因果推理的图边方案。基于因果推理的时空图神经网络 (CSTGNN) 根据佛罗里达州海岸线 4 个观测站的 13 年观测数据进行训练和评估。实验证实了 CSTGNN 的能力，它在不同的指标和观测站上优于六个常用的竞争基线，交付时间长达 6 小时。此外，还全面研究了捕获空间依赖性和利用因果推理的好处。总之，我们相信这种新颖的时空预报框架可以通过提高风暴潮预报能力来增强沿海地区的抵御能力。