WSTSPlus

Improved Wildfire Spread Prediction with Time-Series Data and the WSTS+ Benchmark

Title: Improved Wildfire Spread Prediction with Time-Series Data and the WSTS+ Benchmark
Conference: IEEE Winter Conference on Applications of Computer Vision (WACV) 2026
Paper: arXiv
Dataset: Zenodo
Model Weights: HuggingFace
Repo:: GitHub Authors: Saad Lahrichi, Jake Bova, Jesse Johnson, Jordan Malof

---

This repository extends the original WildfireSpreadTS benchmark with new models, improved training, and an expanded benchmark dataset, WSTS+.

---

Benchmark Results (AP ± Standard Deviation)

Mean Test AP for T = 1 and T = 5 Across Feature Sets

---

Datasets

WSTS+ (Extended Benchmark)

• Name: WSTS+
• Years: 2016–2018; 2021-2023
• Link: https://doi.org/10.48550/arXiv.2502.12003

Original WSTS Dataset

• Name: WildfireSpreadTS (WSTS)
• Years: 2018–2021
• Link: https://doi.org/10.5281/zenodo.8006177

Both datasets are compatible with the same preprocessing and training code in this repository.

Model Weights

We release our best T=1 and T=5 models (Res18-Unet and Res18-UTAE) as PyTorch .pth files containing the raw state_dict. They follow a consistent naming convention: fold_<foldID>_testAP<value>.pth and they are organized in folders by architecture (Res18UNet, Res18UTAE), temporal dimension (T=1 or T=5), and feature set used (Veg, Multi, or All).

Each model directory contains 12 files: one per cross-validation fold (fold_0 … fold_11). The filename includes the Test AP, allowing for easy identification of best- and worst-performing folds. Link: HuggingFace

Loading pretraind Models

We provide a utility script load_trained_model.py to allow for quickly loading the pretrained models. Example calls:

python load_trained_model.py \
    --weights_path /path/to/unet/model/fold_X_testAP0.X.pth \
    --model unet

Or for UTAE:

python load_trained_model.py \
    --weights_path /path/to/utae/model/fold_Y_testAP0.Y.pth \
    --model utae

Model Comparison Table

---

WSTS vs. WSTS+ Dataset Comparison

---

Citation

If you use this fork or the WSTS+ benchmark, please consider citing:

@inproceedings{
    lahrichi2026improved,
    title={Improved Wildfire Spread Prediction with Time-Series Data and the WSTS+ Benchmark},
    author={Saad Lahrichi, Jake Bova, Jesse Johnson, Jordan Malof},
    booktitle={IEEE Winter Conference on Applications of Computer Vision (WACV) 2026},
    year={2026},
    url={https://arxiv.org/abs/2502.12003}
}