Reconstructing historical storm surges levels with models and machine learning

Reliable sea-level data is needed for accurate coastal risk assessments, but historical records often have biases, missing entries, and inaccuracies. This study presents a comprehensive framework for augmenting and rectifying storm surge records through the integration of machine learning, hydrodynamic modelling, and statistical analysis. Using a block-median approach and standard statistical methods, station-specific biases are found and fixed. To add more historical data, a combination of methods is used: a machine learning model like Random Forest is trained to fill in the gaps in storms' time series when only model data is available, and hydrodynamic simulations are used to find extreme events that aren't in the observational record. The framework is used at more than 50 tide gauge stations along the Danish coast to create a high-quality, validated dataset of reconstructed extremes and bias-corrected observations for the years 1961 to 2024. This dataset is very useful for climate adaptation and accurate coastal risk assessments because it focusses on critical windows about 24 hours before and after surge peaks.