Predictions of extreme coastal water levels are important to coastal engineering design and hazard mitigations in Florida. Annual maximum water levels are often used in frequency analysis of 1% annual chance flood in coastal flood hazard mapping. However, because of the damage to measurement instruments during hurricanes, some annual maximum water levels may be missed, which makes coastal flood hazard analysis difficult. In this study, a technique was developed to use artificial neural network and harmonic analysis for recovering extreme coastal water levels during hurricanes. The total water levels are decomposed into tidal components and storm surge. Tidal components can be derived by harmonic analysis, whereas storm surge can be predicted by neural network modeling on the basis of the observations of local wind speeds and atmospheric pressure. The neural network model uses three-layer feed-forward back-propagation structure with advanced scaled conjugate training algorithm. The method presented in this study has been successfully tested in Panama City Beach, located on the Florida coast, for Hurricane Dennis (2005), Hurricane Ivan (2004), and Hurricane Opal (1975). Model-predicted peak elevations reasonably match with observations for the three hurricane events. The decomposed storm surge hydrograph also make it possible for the analysis of potential extreme water levels if the storm surge occurs during spring high tide.