WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction

Sebastian Gerard, Yu Zhao, Josephine SullivanNeurIPS 2023

Wildfires are becoming more frequent and severe due to climate change, yet our ability to predict their movement in real-time remains surprisingly limited. While satellite products are excellent at telling us where a fire is, they struggle to tell us where it will be tomorrow. A new paper from researchers at KTH Royal Institute of Technology, presented at NeurIPS 2023, introduces WildfireSpreadTS, a comprehensive dataset designed to move the field from simple detection to sophisticated time-series prediction.

TL;DR

WildfireSpreadTS is a new multi-modal, multi-temporal dataset featuring over 13,000 images of 607 wildfire events in the US. By providing daily snapshots of fuel, weather, and topography at a 375m resolution, it enables researchers to build deep learning models that predict fire spread 24 hours into the future.

What Problem Does This Paper Address?

Predicting wildfire spread is a "wicked" problem. It involves complex physical processes—combustion, heat transfer, and atmospheric fluid dynamics—all occurring across varying scales. Historically, researchers have relied on heavy physical simulators (like FARSITE) which are computationally expensive and hard to sync with real-time satellite data.

On the machine learning side, previous datasets were often "mono-temporal," meaning they provided a single snapshot of a fire and asked the model to predict the next day. This ignores the "momentum" of the fire—how it has been growing or shrinking over the previous days. Furthermore, many existing datasets are not public, creating a high barrier to entry for researchers who aren't experts in remote sensing.

Methods

The authors established several baselines to test the dataset’s utility, ranging from simple linear models to state-of-the-art spatio-temporal architectures:

  • Logistic Regression: A simple 3x3 convolution baseline using only the current day's data.
  • U-Net (ResNet18): A standard computer vision architecture used for image segmentation, tested in both mono-temporal and multi-temporal (stacking days as channels) configurations.
  • ConvLSTM: A recurrent neural network that combines convolutional layers with Long Short-Term Memory units to capture temporal dependencies.
  • UTAE (U-Net with Temporal Attention): A sophisticated model that applies self-attention across the time dimension at the "bottleneck" of a U-Net, allowing the model to focus on the most relevant historical days.

To handle the extreme class imbalance (fire pixels are very rare compared to unburned land), they used Average Precision (AP) as the primary evaluation metric.

Datasets

The core contribution is the WildfireSpreadTS dataset:

Overview of the multi-modal inputs in WildfireSpreadTS used to predict the next day's active fire.
  • Scope: 607 fire events in the United States (2018–2021).
  • Modality: 23 input channels including VIIRS active fire masks, surface reflectance (to detect smoke/burned area), GRIDMET weather data, GFS weather forecasts, and SRTM topography (slope, aspect, elevation).
  • Resolution: 375m spatial resolution with a 24-hour temporal resolution.
  • Comparison: Unlike the previous benchmark NextDayWildfireSpread, this dataset is multi-temporal (time-series) and offers higher spatial resolution (375m vs 1km).

Key Insights and Ideas

Topography features (Aspect, Elevation, Slope) proved to be strong predictors of fire spread.
  • History Matters: Multi-temporal models (like UTAE) significantly outperformed mono-temporal models. Knowing how the fire behaved over the last five days provides critical context that a single snapshot lacks.
  • The Power of Topography: In ablation studies, topography (elevation and slope) emerged as one of the strongest predictors. Fire naturally moves faster uphill, and the dataset captures this physical reality.
  • Freshness of Data: The authors found that using daily-updated vegetation indices from VIIRS performed better than the standard NDVI (which only updates every 8 days). In a fast-moving fire, week-old vegetation data is "stale."
  • The Optimization Paradox: Interestingly, adding all available features sometimes led to worse performance than using a curated subset. This suggests that high-dimensional input spaces make the optimization landscape very difficult for current deep learning models to navigate.

Key Results

The models were evaluated on their ability to predict the binarized fire mask 24 hours ahead:

  • Persistence Baseline: (Predicting the fire stays exactly where it is) achieved an AP of 0.193.
  • U-Net (Mono-temporal): Achieved an AP of 0.341.
  • UTAE (Multi-temporal, 5 days): The top performer, reaching an AP of 0.372.
  • Efficiency: The UTAE model achieved these results with only 1.1M parameters, making it significantly more efficient than the 14M+ parameter U-Net.

Limitations and Future Work

The authors are transparent about several hurdles:

  1. Label Noise: Satellite detections are often obscured by the very thing they are trying to track: smoke and clouds. This introduces "noisy" ground truth.
  2. Geographic Bias: The current dataset is limited to the US. Expanding this to different climates (e.g., Australia or the Mediterranean) is a necessary next step for generalization.
  3. Static vs. Dynamic: Currently, models treat static features (like elevation) the same as dynamic ones (like wind). Future work could involve specialized architectures that process these differently.
  4. Generative Modeling: Since fire spread is inherently stochastic (probabilistic), the authors suggest that generative models or uncertainty estimation could provide more useful "risk maps" than simple binary predictions.

Paper Citation: Sebastian Gerard, Yu Zhao, Josephine Sullivan. "WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction." NeurIPS 2023 Track on Datasets and Benchmarks. Link to Code/Dataset

Summary updated

BibTeX
@inproceedings{gerard2023wildfirespreadts,
  author = {Sebastian Gerard and Yu Zhao and Josephine Sullivan},
  title = {{WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
}
Figure