Testing for Agreement between Human Digitizers and a Machine Learning Algorithm when Digitizing Rip Currents in Time-Averaged Imagery
Abstract
Rip currents result from alongshore variations in wave dissipation; they transport sediment and nutrients within the surfzone but also pose a surf safety hazard. Studying these ephemeral currents is logistically difficult, but remote sensing provides avenues for identifying and tracking rip currents, as well as the bathymetric rip channels which foster rip current development. Machine learning applications can be trained to identify visual signatures in time-averaged optical imagery - where those visual signatures were assumed to signal the presence of a rip channel and/or current. Multiple studies have developed machine learning algorithms to automate the digitization of these visual signatures, which performed well when trained and tested with visual information (only). In this study we investigate the spatial coincidence of rip channels present in lidar bathymetry and the visual signatures they produce in optical time averaged images, as digitized by both a trained machine learner and a group of human digitizers. We found that the majority of target features in the bathymetric data were not signaled by visual signatures, regardless of whether those visual signatures were manually digitized by a group of subject matter experts or automatically by the trained machine learner. These findings suggest that bathymetric depressions may not reliably produce the visual signature in the time averaged imagery, as previously assumed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 30, 2023
- Accession Number
- AD1209537
Entities
People
- A. Penko
- Sarah Trimble
Organizations
- United States Naval Research Laboratory