Abstract
To streamline data processing, ML is being used to detect hidden relationships in lidar meta-data that relate to the probability that a given lidar pulse return is bathymetry. Three types of meta-data have been extracted from lidar point clouds: 1) return-specific, 2) macro flight-path, and 3) topographic. Four 500m-by-500m tiles from NOAA Remote Sensing Division shallow-water lidar data in the Florida Keys are being used as a testbed. Major conclusions to date are:
- Artificial Neural Networks (ANNs) outperform Extreme Gradient Boosting and Regularized Logistic Regression, but ANNs provide the least amount of information about what meta-data are most valuable.
- The meta-data that are most useful appear related to ocean depth.
- Signal strength as measured by the true positive rate for bathymetry classification varies considerably across the four tiled data sets – i.e., from 11% to 92% for ANN models.
The integration of this work into operational procedures will also be discussed
Presenter Bio
Kim Lowell has a B.Sc. (Forestry), two M.Sc.s (Forest Biometric and Data Analytics), and a Ph.D. (Forest Biometrics). He has worked for universities, government agencies, and national research organisations. He has extensive experience using geospatial analysis, image processing, and applied statistics for subjects as diverse as carbon accounting, creating markets to buy ecosystem services, and socio-environmental landscape monitoring.
Publication Date
4-19-2019
Document Type
Presentation
Recommended Citation
Lowell, Kim, "Better Shallow-water Bathymetric Maps from Airborne Lidar Data Using Machine Learning" (2019). Seminars. 278.
https://scholars.unh.edu/ccom_seminars/278