2D CNN-based Nearshore Bathymetry Inversion from Synthetic Imagery
summary
A 2D FCN learns to map time-averaged "timex" images from the surf zone to depth using photorealistic synthetic imagery generated by Celeris, trained with a 2D fully convolutional U2-Net (Bayesian encoder–decoder) on ~1,700 training / 150 test images produced by Latin-hypercube wave sampling. Results show RMS errors around 0.32–0.39 m (comparable to existing timelapse methods) over an 860 m offshore domain, with higher uncertainty where few breaking waves provide signal.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:16Significance and Motivation
- 3:19Breaking Waves to Water Depth How Surfing Works Breaking Waves
- 4:12Remote Sensing for Observation
- 4:41Observation to Timex
- 5:33Parametric Beach Tool
- 6:23Beach Wizard
- 7:15Deep Neural Networks for Bathymetry Inversion
- 7:58Transfer Learning for Domain Adaption Smal Real Large Synthetic Dataset
- 8:34Fully Convolutional Neural Network
- 9:05Celeris Wave Model
- 9:54Cross-shore Bathymetric Profile • First input to the Celeris wave
- 10:51Create Synthetic Bathymetries
- 12:14Wave Conditions - Observations
- 15:00Create Synthetic Timex Images
- 19:37Preliminary Summary Statistics Wave Conditions Test Error Values
- 21:32Dataset Setup & Feature Inputs
- 22:20Exploring Uncertainty
- 23:03Celeris Video Imagery Video frame data used to
- 23:43Synthetic - Real Timex Images
related talks
10mLightning Talk: Seismic Data to Subsurface Models with OpenFWI
01h 34mEroding Coastlines: A Geospatial & Computer Vision Analysis
25mDeep learning in deep nets: Helping fish farmers feed the world
48m(Deep) Learning to Fly
21mLand Boundary Delineation using U-Nets
39mInsights About Places with Deep Learning Computer Vision
33mWhat and where? - Machine learning for geospatial image analysis
26mComputer Vision in the Era of Deep Learning