Deep learning motivated data imputation of tropical cyclone radius of maximum winds

Probabilistic coastal hazard assessments support risk-informed design approaches for critical infrastructure.These assessments typically are based on the Joint Probability Method, which requires accurate characterization of the joint distribution of tropical cyclone (TC) parameters.However, historical TC datasets used to derive the joint distribution often contain incomplete records for the Radius of Maximum Winds (R max ), a key variable in these analyses.While prior research has employed machine learning (ML) methods like Gaussian Process Regression and shallow Artificial Neural Networks to estimate missing R max values, the potential of deep learning techniques for this task remains largely unexplored.Building on previous work, this study investigates the use of historical storm data to evaluate and compare various deep-learning strategies for R max imputation.Specifically, we examine the performance of one-dimensional Convolutional Neural Networks (1DCNNs) and Long Short-Term Memory networks (LSTMs).This study also introduces physically informed model refinement and assessment approaches.We also compare these deep learning methods with previously unexplored classical ML methods, such as decision trees, particularly XGBoost.The study focuses on simplifying input features and leveraging the time series characteristics of storms to improve the imputation of missing R max values in historical datasets.By exploring deep learning methodologies, this research aims to enhance the accuracy and reliability of imputing missing data in historical storm records.