This study introduces two artificial neural network (ANN)-based methodologies to predict hourly water levels (WLs) in wetlands characterized by water tables at or near the surface that respond rapidly to precipitation. The first method makes use of hourly precipitation data and WL data from nearby sites. The second method is a combination of ANN, recursive digital filter, and recession curve method and does not require any nearby site. The proposed methods were tested at two headwater wetlands in coastal Alabama. Site 17 had two nearby sites whose WLs were highly correlated with Site 17’s. The root-mean-square error and Nash–Sutcliffe efficiencies were 2.9 cm and 0.98, respectively, when the first method was applied to Site 17. The second method was tested at Site 32. For this, the WL time series was separated into quick- and slow-response components. A combination of ANN and base-flow separation methods proved to be very efficient for WL prediction at this site, especially when the duration of quick-response components of individual events was less than 6 h. The proposed methodologies, therefore, proved useful in predicting WLs in wetlands dominated by both surface water and groundwater.