Extreme Events and Probability Analysis Along the United States East Coast Based on High Spatial-Coverage Reconstructed Storm Surges

Analyzing the features of extreme events and estimating their probabilities robustly require high spatial coverage, high temporal resolution, and sufficiently long storm surge (SS) records. However, in situ observations cannot always meet these demands due to spatiotemporal sparseness. Here, we proposed a novel regional all-site modeling framework based on a machine learning method, the extreme gradient boosting tree. This framework can reconstruct long SS records simply and quickly and can estimate storm surges simultaneously at both gauged and ungauged locations. Compared to in situ observations, the distribution patterns of SS variations during extreme events can be recognized easily from the reconstructed hourly SS data set. Since its available record is longer than 60 years (1959-2020), the estimation uncertainties of extreme event probabilities are significantly decreased. Noticeably high extreme SS return levels were found along the coast of the northern Gulf of Mexico, which should be given great attention.

Automated summary

Analyzing extreme event features and estimating their probabilities require comprehensive storm surge records, which are often lacking. We propose a novel regional modeling framework based on machine learning to reconstruct long storm surge records and estimate surges at both gauged and ungauged locations simultaneously.