Deep Sensing of Ocean Wave Heights with Synthetic Aperture Radar
summary
Curates 2015–2018 co-locations between Sentinel-1 SAR and altimeters to train a convolutional neural network on 2D image modulation spectra (real+imag after 2D FFT, 72×60) plus auxiliary features; trains with Sherpa tuning, LR annealing and early stopping on 303k train / 185k test samples and predicts significant wave height. Results halve RMSE from 0.6 m to 0.3 m on 2018 test data (50% RMSE reduction), generalize across platforms and incidence angles, and still struggle above ~8 m due to scarce examples.
AI-written summary · about
chapters (10)
- 0:00Introduction to SAR wave sensing
- 1:34Problem statement and wave height prediction
- 4:52Building large datasets with collocations
- 8:44Data preprocessing for deep learning
- 11:09Model architecture and training
- 14:21Results and performance comparison
- 17:58Feature importance analysis
- 20:50Future work and oceanography phenomena
- 22:38Summary and applications
- 23:47Q&A session
related talks
36mClassification of Satellite Imagery
27mThe View from Above Getting Started with Satellite Images
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
45mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet an Open Source Workflow
17mFree satellite imagery and DEM analyses enabling NRM in the developing world
13mAnomaly Detection on Remote Sensing with Ray + Horovod
06mDetecting marine heatwaves (MHWs) by using R-programming
17mSLNet: A Spectrogram Learning Neural Network for Deep Wireless Sensing