Applying Machine Learning to LIDAR Pulse Return Meta-Data to Improve Bathymetric Mapping in Shallow Water

TitleApplying Machine Learning to LIDAR Pulse Return Meta-Data to Improve Bathymetric Mapping in Shallow Water
Publication TypeJournal Article
YearSubmitted
AuthorsLowell, K, Calder, BR
JournalComputers and Geosciences
Date Published2019
PublisherTaylor and Francis
Place PublishedLondon, UK
KeywordsExtreme Gradient Boosting, Florida Keys, Neural Networks, Regularized Logistic Regression, True Positive/Negative Rate

To improve an existing operational methodology for extracting bathymetry, the strength of bathymetric signal in meta-data was evaluated for lidar point clouds in shallow waters (less than 20 m depth) for four areas near the Florida Keys using three machine learning (ML) algorithms — extreme gradient boosting, neural networks, and regularized logistic regression. Meta-data considered included “pulse specific” information such as the number of a pulse return and the pulse incidence angle as well meta-data that addressed the consistency and stability of the flight path.  (Pseudo) r2 values for ML classification models for Bathy/NotBathy average about 0.50 for point clouds having between 1 and 7 million observations (pulse returns). Global accuracy for all models is above 80% but this reflects in part a Bathy/NotBathy imbalance in lidar point clouds. A method for assigning pulses to the Bathy/NotBathy classes using a probability decision threshold other than 0.50 is presented as a pathway to usage in operational workflows.

Check When Not CCOM Publication