In the Coastal Hazards System's (CHS) Probabilistic Coastal Hazard Analysis (PCHA) framework developed by the United States Army Corps of Engineers (USACE), historical records of tropical cyclone parameters have been used as data sources for statistical analysis, including fitting marginal distributions and measuring correlations between storm parameters. One limitation of the available historical databases is that observations of central pressure and radius of maximum winds are not available for a large number of storms. This may adversely affect the results of statistical analyses used to develop hazard curves. This study uses machine learning techniques to develop a data imputation method to “fill in” missing storm parameter records in historical datasets used for Joint Probability Method (JPM)-based coastal hazard analysis such as the USACE's CHS-PCHA. Specifically, Gaussian process regression (GPR) and artificial neural network (ANN) models are investigated as candidate machine learning-derived data imputation models, and the performance of different model parameterizations is assessed. Candidate imputation models are compared against existing statistical relationships. The effect of the data imputation process on statistical analyses (marginal distributions and correlation measures) is also evaluated for a series of example coastal reference locations.

中文翻译：

---

机器学习驱动的沿海灾害评估中使用的风暴数据的数据插补

在美国陆军工程兵团 (USACE) 开发的海岸灾害系统 (CHS) 概率海岸灾害分析 (PCHA) 框架中，热带气旋参数的历史记录已被用作统计分析的数据源，包括拟合边际分布和测量风暴参数之间的相关性。可用历史数据库的一个限制是，无法对大量风暴进行中心气压和最大风半径的观测。这可能会对用于制定危险曲线的统计分析结果产生不利影响。本研究使用机器学习技术开发一种数据插补方法，以“填充”用于基于联合概率法 (JPM) 的沿海灾害分析（例如 USACE 的 CHS-PCHA）的历史数据集中缺失的风暴参数记录。具体来说，研究了高斯过程回归（GPR）和人工神经网络（ANN）模型作为候选机器学习衍生数据插补模型，并评估了不同模型参数化的性能。将候选插补模型与现有统计关系进行比较。还针对一系列沿海参考位置示例评估了数据插补过程对统计分析（边际分布和相关性测量）的影响。