Operational performance of a combined Density- and Clustering-based approach to extract bathymetry returns from LiDAR point clouds

Abstract

Conventional techniques for extracting bathymetric soundings from LiDAR point clouds are at best semi-automated and require considerable manual effort. An algorithm that couples a widely used sonar data processing method with a newly developed machine-learning(ML)-based algorithm was evaluated for accuracy and potential operationalisation. Data representing an operationally realistic range of environmental and data conditions comprised 103 500 m-by-500 m data tiles for method development/calibration and 20 tiles for validation located in the Florida Keys. Tiles are processed individually to classify each LiDAR pulse return (“sounding” in hydrographic terminology) as bathymetry or not. Compared to a reference classification an average agreement of about 90% was produced for the calibration and validation data sets, and accuracy varied depending on ocean bottom and data conditions. The average false negative rate – the most important metric in hydrographic mapping – was about 5%. Processing time for tiles containing the average number of soundings (seven million) on a desktop computer was approximately 100 min. The algorithm does not require in situ ground-“truth” data for training or calibration, although its adaptation to other geographic and data conditions might require data-guided adjustment of ML tuning parameters.