Zhao et al. (2025) TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction

Identification

• Journal: Scientific Data
• Year: 2025
• Date: 2025-11-19
• Authors: Yu Zhao, Sebastian Gerard, Yifang Ban
• DOI: 10.1038/s41597-025-06271-3

Research Groups

• KTH Royal Institute of Technology, Stockholm, Sweden

Short Summary

This paper introduces TS-SatFire, a comprehensive multi-temporal remote sensing dataset designed for integrated wildfire monitoring and prediction using deep learning models. It provides benchmarks for active fire detection, daily burned area mapping, and next-day wildfire progression prediction across the contiguous U.S.

Objective

• To present a comprehensive multi-temporal remote sensing dataset (TS-SatFire) to facilitate the development of robust deep learning models for active fire detection, daily burned area mapping, and next-day wildfire progression prediction.

Study Configuration

• Spatial Scale: Primary VIIRS imagery at 375 meters. Auxiliary data resolutions vary from 10 meters (ESRI 10 m Annual Land Cover) to 27.83 kilometers (Global Forecast System).
• Temporal Scale: Dataset covers wildfire events from January 2017 to October 2021. Data frequency ranges from twice-daily (VIIRS imagery) to 8-day (Vegetation Indices) and yearly (Land Cover).

Methodology and Data

• Models used: GRU, LSTM, T4Fire (Transformer-based), U-Net, Attention U-Net, UNETR, SwinUNETR (2D and 3D variants) were used as baseline models for benchmarking the dataset.
• Data sources:
  • Satellite Imagery: VIIRS (Suomi-NPP, NOAA-20, NOAA-21) Imagery Bands (I1-I5) and Moderate Band (M11) from NASA LAADS.
  • Vegetation Indices: Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) from VNP13A1 VIIRS Vegetation Indices product.
  • Weather Data: Gridded Surface Meteorological Dataset (GRIDMET) for wind speed/direction, precipitation, specific humidity, Palmer Drought Severity Index (PDSI), minimum/maximum temperature, and Energy Release Component.
  • Weather Forecast Data: Global Forecast System (GFS) for wind, temperature, and humidity forecasts.
  • Topography: NASA SRTM Digital Elevation dataset (derived slope and aspect).
  • Land Cover: MODIS Land Cover Type Yearly Global product (MCD12Q1.061) and ESRI 10 m Annual Land Cover dataset.
  • Labels: NASA VIIRS Active Fire (AF) product (training AF), National Interagency Fire Center (NIFC) perimeters, and accumulated VIIRS AF detections (training Burned Area), with manual quality assurance for test labels.
  • Wildfire Events: GlobFire dataset (for training set fire IDs) and MODIS monthly burned area product (for BA/Prediction test set fire IDs).
  • All auxiliary data processed via Google Earth Engine.

Main Results

• The TS-SatFire dataset, totaling 71 GB, comprises 3552 surface reflectance images and extensive auxiliary data for 179 distinct wildfire events in the contiguous U.S.
• Benchmarks for three tasks were established using various deep learning models:
  • For active fire detection, UNETR-3D achieved the highest F1 score of 0.811 and IoU of 0.706.
  • For daily burned area mapping, SwinUNETR-3D achieved the highest F1 score of 0.855 and IoU of 0.768.
  • For next-day wildfire prediction, SwinUNETR-3D achieved the highest F1 score of 0.374 and IoU of 0.331, indicating the significantly higher complexity of this task compared to detection.
• Leveraging temporal information in models (e.g., 3D models over 2D) generally improved performance for detection tasks.
• Feature importance analysis revealed that Near Infrared (Band I2) and Short-wave Infrared (Band M11) are most crucial for burned area mapping and prediction, while Medium Infrared (Band I4) is key for active fire detection.

Contributions

• Introduction of TS-SatFire, the first multi-task, multi-temporal remote sensing dataset specifically designed for integrated wildfire detection, monitoring, and prediction.
• Provision of a comprehensive dataset covering a wide range of wildfire events in the contiguous U.S. (January 2017 – October 2021), integrating diverse satellite imagery and auxiliary environmental data.
• Establishment of robust deep learning benchmarks across three distinct wildfire-related tasks, offering a foundation for future research and development of Earth observation foundation models.
• Demonstration of the critical role of temporal information and specific spectral bands in improving wildfire detection and prediction accuracy.

Funding

• Formas (Swedish research council for sustainable development) for the 'Sentinel4Wildfire' project.
• Digital Futures for the 'EO-AI4Global Change' project.

Citation

@article{Zhao2025TSSatFire,
  author = {Zhao, Yu and Gerard, Sebastian and Ban, Yifang},
  title = {TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction},
  journal = {Scientific Data},
  year = {2025},
  doi = {10.1038/s41597-025-06271-3},
  url = {https://doi.org/10.1038/s41597-025-06271-3}
}