Modelling forest fire dynamics using conditional variational auto-encoders

Reviewed Article | Dataset

TL;DR

This research applies deep networks to simulate the spatial and temporal evolution of a forest fire's impact on a region. Using a Conditional Variational Autoencoder (CVAE) model, we generate intermediate representations of the burnt area's evolution. we also employ a CVAE model to predict future fire propagation behavior, estimating the burnt area at various time horizons and propagation stages. Real-world data is used to assess the model's performance, with results showing high similarity and temporal consistency metrics, indicating that the application of CVAE in this context is capable.

Prerequisites

• conda for managing Python environments.
• git for version control.
• zenodo-get downloader.
• Python 3.10.

Repo Structure

spatiotemporal-vae-reconstruction/ 📁 
├── checkpoints/      💾: Checkpoints for model training 
├── data/             📂: Data
├── logs/             📃: Log files 
├── notebooks/        📓: Jupyter notebooks 
│   ├── 1_data_preparation.ipynb  📊: Data preparation
│   ├── 2_modelling.ipynb         📚: Model training
│   ├── 3_evaluation.ipynb        📈: Evaluation
│   └── ...                        
├── outputs/          📺: Output files generated during execution 
├── src/              📜: Source code files
│   ├── __init__.py   
│   ├── cvae_model.py 📑: CVAE model implementation
│   ├── eval.py       📐: Evaluation script
│   ├── utils.py      🛠️: Utility functions
│   └── ...                
├── .gitignore        🚫: Files to ignore in Git
├── config.yml        🎛️: Configuration settings 
├── README.md         📖: Project README file 
└── requirements.txt  📄: Dependencies

Operating System

This project has been tested on the following operating systems:

• Pop! OS 22.04 LTS (Nvidia)
• Windows 11

Main Settings

• Tensorflow 2.15.0
• CUDA 11.5

Setting up the environment

# Create a new Conda environment named 'cvae' with Python version 3.10
conda create --name cvae python=3.10  

# Clone the 'spatiotemporal-vae-reconstruction' project from GitHub
git clone https://github.com/Tiago1Ribeiro/spatiotemporal-vae-reconstruction.git

# Change the current directory to the cloned project's directory
cd spatiotemporal-vae-reconstruction

# Install the Python dependencies listed in the 'requirements.txt' file
pip install -r requirements.txt  

# Downloads *BurnedAreaUAU* dataset and save it in the 'dataset' directory
zenodo_get --output dataset 10.5281/zenodo.7944963 # or download it from the link below

To-do list (🚧):

• Include the U-Net samples (Base, RED, 3D) and sampled_masks.txt in the dataset (new Zenodo version)
• Test repo in other machines

Article Citation

@article{ribeiro2024modelling,
  title={Modelling forest fire dynamics using conditional variational autoencoders},
  author={Ribeiro, Tiago Filipe Rodrigues and Silva, Fernando Jos{\'e} Mateus da and Costa, Rog{\'e}rio Lu{\'\i}s de Carvalho},
  journal={Information Systems Frontiers},
  pages={1--20},
  year={2024},
  publisher={Springer},
  doi = {https://doi.org/10.1007/s10796-024-10507-9}
}

Dataset Citation

@misc{ba_uav_dataset,
  author       = {Ribeiro, Tiago F. R. and Silva, Fernando and Moreira, Jos\'e and Costa, Ro\'erio Lu\'is de C.},
  title        = {BurnedAreaUAV Dataset (v1.1)},
  month        = may,
  year         = 2023,
  publisher    = {Zenodo},
  version      = {1.1},
  doi          = {10.5281/zenodo.7944963},
}