Three ML techniques – artificial neural networks (ANNs), extreme gradient boosting (XGB), and regularized logistic regression (RLR) – have been applied to lidar pulse return meta-data to estimate the probability that each return is bathymetry -- p(Bath) -- for four tiles from a NOAA Remote Sensing Division shallow-water data in the Florida Keys. To facilitate operationalization, the meta-data are extracted from the point cloud alone – i.e., no ancillary data are employed. Three types of meta-data are being explored: return-specific, ocean floor topography, and flight path crenularity.

Major conclusions to date are:

• ANNs provide better p(Bath) estimates than XGB and RLR, but provide the least amount of information about what meta-data are most important for prediction.
• The p(Bath) signal strength as measured by the true positive rate for bathymetry varies considerably across the four tiled data sets – i.e., from 11% to 92% for ANN models.

Ultimately, the p(Bath) estimate will be integrated into an already-operational lidar-extraction algorithm.