To automate extraction of bathymetric soundings from lidar point clouds, two machine learning (ML[1]) techniques were combined with a more conventional density-based algorithm.  The study area was four data “tiles” near the Florida Keys.  The density-based algorithm determined the most likely depth (MLD) for a grid of “estimation nodes” (ENs).  Unsupervised k-means clustering determined which EN’s MLD depth and associated soundings represented ocean depth rather than ocean surface or noise to produce a preliminary classification.  An extreme gradient boosting (XGB) model was fitted to pulse return metadata – e.g., return intensity, incidence angle -- to produce a final Bathy/NotBathy classification.  Compared to an operationally produced reference classification, the XGB model increased global accuracy and decreased the false negative rate (FNR) – i.e., undetected bathymetry – that are most important for nautical navigation for all but one tile.  Agreement between the final XGB and operational reference classifications ranged from 0.84 to 0.999.  Imbalance between Bathy and NotBathy was addressed using a probability decision threshold that equalizes the FNR and the true positive rate (TPR).  Two methods are presented for visually evaluating differences between the two classifications spatially and in feature-space.