Prototype for automated damage detection in Sentinel 1 imaging of iceselves using unsupervised learning.
The prototype makes use of a VAE architecture, and employs a bespoke input genertation framework to create cutouts for network training from larger imaging tiles.
Notebooks have been used for intial prototyping. These have been converted to scripts for a first production prototype.
Within each directory another README file is included provide more details
- files : contains list of data tiles used for training/testing
- notebooks : contains python notebooks for training, pre- and post-processing
- scripts : contains python scripts for traininig, pre- and post-processing
- training: contains one of of the trained VAE models in a .zip, and an overview of the trained models that are available at /projects/0/einf512/trained_models/
Last analysis (nov2023) concluded that the predcited damage from the trained VAE, applied to Antarctic wide Sentinel-2 data, was very similar to simple threshold on spectral bands (and so the VAE was not able to deduce interesting patterns after all). View results at: https://code.earthengine.google.com/9af710497fac6b76c710873839e797b6