Rapid Wildfire Hotspot Detection Using Self-Supervised Learning On Temporal Remote Sensing Data

Dataset and code for the paper Rapid Wildfire Hotspot Detection Using Self-Supervised Learning On Temporal Remote Sensing Data (IGARSS 2024).

https://arxiv.org/abs/2405.20093v1

Note

Dataset available at hf.co/datasets/links-ads/multi-temporal-hotspot-dataset.

---

Installation

First, create a python environment. Here we used python 3.9 and torch 1.9, with CUDA 11.1. We suggest creating a python environment, using venv or conda first.

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

Training

You can launch a training with the following commands:

$ CUDA_VISIBLE_DEVICES=... python src/train.py  --catalog_file_train=... --catalog_file_val=....  --catalog_file_test=... <..args>

You can specify the following args:

• batch_size
• max_epochs
• lr
• gpus
• log_dir
• seed
• optimizer
• scheduler
• compute_loss_lc (False if not specified)
• positive_weight_loss_class (default 1)
• lc_loss_weight (default 2)
• mask_strategies (use "random_timesteps")
• mask_ratio (default 0.75)

Inference

To produce inference maps, run something like the following:

$ CUDA_VISIBLE_DEVICES=... python src/test.py --model_checkpoint <args>

Citation

@misc{barco2024rapid,
      title={Rapid Wildfire Hotspot Detection Using Self-Supervised Learning on Temporal Remote Sensing Data}, 
      author={Luca Barco and Angelica Urbanelli and Claudio Rossi},
      year={2024},
      eprint={2405.20093},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}