Physics-informed Variational Autoencoders for Improved Robustness to Environmental Factors of Variation

This repository contains code for the following paper, under review in Springer Machine Learning:
Physics-informed Variational Autoencoders for Improved Robustness to Environmental Factors of Variation

A short version of this work has been accepted as a workshop paper at Machine Learning for Remote Sensing, ICLR 2024.

Please cite this paper if you use the code in this repository as part of a published research project (see bibtex citation below).

Extrapolation of damped pendulum time series

In order to reproduce the results, you can train the networks with the scripts exps/extrapolation/train.sh (baselines) and exps/extrapolation/p3vae_train.sh (our method).

Hyperspectral image classification

The airborne hyperspectral images acquired during the CAMCATT-AI4GEO experiment in Toulouse, France are publicly available here: https://camcatt.sedoo.fr/

To load and save image patches, use an instance of the GeoDataset class in the exps/classification/data.py file.

In order to reproduce the results, you can train the networks by running exps/classification/train.py with default arguments.

Methane plume inversion

Scripts to run the optimal estimation algorithm and apply p$^3$VAE to the inversion of methane plume from hyperspectral satellite data are in the exps/inversion folder.

Setup

The code was run using python 3.8:

1. create a python virtual environment
2. clone this repo: git clone https://github.com/Romain3Ch216/p3VAE.git
3. navigate to the repository: cd p3VAE
4. install python requirements: pip install -r requirements.txt

Feedback

Please send any feedback to romain.thoreau@cnes.fr, or open an issue.

Bibtex citation

@misc{thoreau2025p3vae,
      title={Physics-informed Variational Autoencoders for Improved Robustness to Environmental Factors of Variation}, 
      author={Romain Thoreau and Laurent Risser and Véronique Achard and Béatrice Berthelot and Xavier Briottet},
      year={2025},
      eprint={2210.10418},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2210.10418}, 
}