Dataset Card for TS-SatFire

This dataset card aims to be a base template for new datasets. It has been generated using this raw template.

Dataset Details

Dataset Description

The TS-SatFire dataset is a comprehensive multi-temporal remote sensing dataset designed to cover the entire life cycle of wildfires. It provides a unified framework to support three critical and interconnected wildfire monitoring tasks: active fire detection, daily burned area mapping, and next-day wildfire progression prediction. The dataset encompasses 179 wildfire events recorded primarily in the contiguous United States between January 2017 and October 2021. It includes 3,552 surface reflectance time-series images derived from VIIRS satellites, coupled with extensive multi-modal auxiliary data such as topography, weather conditions, and fuel types, spanning roughly 71 GB in total. Ground-truth labels for current active fires (AF) and burned areas (BA) are provided per image, with manual quality assurance conducted to ensure reliable evaluation benchmarks.

• Curated by: Yu Zhao, Sebastian Gerard, and Yifang Ban (KTH Royal Institute of Technology)
• Language(s) (NLP): English (en)
• License: CC-BY-4.0

Dataset Sources

Repository (Kaggle): https://www.kaggle.com/datasets/z789456sx/ts-satfire Repository (GitHub): https://github.com/zhaoyutim/TS-SatFire Paper: TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction (arXiv:2412.11555)

Uses

Direct Use

TS-SatFire is intended to stimulate research in deep learning-based wildfire monitoring and remote sensing. Suitable direct uses include:
Training 2D spatial, 1D temporal, or 3D spatial-temporal deep learning models (like U-Net, Attention U-Net, or Transformers) for multi-spectral image segmentation.
Pixel-wise classification for active fire (AF) detection and burned area (BA) mapping.
Forecasting applications predicting how existing wildfires will progress on a day-to-day basis using time-series inputs combined with multi-modal weather/topography features.

Out-of-Scope Use

Real-time emergency deployment: Models trained solely on this dataset should not be immediately used in operational, life-critical emergency response systems without thorough validation and data integration.

Zero-shot global generalization: Because the dataset specifically captures North American vegetation and climates (contiguous US), models may suffer from severe domain shifts if applied to ecosystems in other continents out-of-the-box.

Dataset Structure

The dataset contains 71 GB of multi-modal files structured around 179 individual wildfire events. Satellite Images: Visible Infrared Imaging Radiometer Suite (VIIRS) surface reflectance data (Imagery Bands I1-I5 at 375m resolution and Moderate Band M11 at 750m resolution).
Auxiliary Data: Environmental context features including Topography (elevation, slope, aspect from SRTM), Weather (historical conditions and GFS forecasts), and Land Cover (MODIS fuel types).
Labels: Binary masks for active fires and burned areas.

Dataset Creation

Curation Rationale

Previous wildfire datasets were highly specialized and focused on individual, isolated tasks (e.g., detecting a fire or predicting a spread). TS-SatFire was created to bridge this gap, offering an integrated, multi-task approach that allows researchers to use the exact same interconnected time-series framework to observe a fire's full life cycle—from initial detection to mapping the burned region, to forecasting the next day's spread.

Source Data

Data Collection and Processing

The raw observations were aggregated from satellite imagery and weather tracking APIs across 179 US fire events between January 2017 and October 2021. Spectral bands particularly sensitive to heat (Mid-Infrared and Long-wave Infrared) were leveraged for detection alongside Near Infrared for visualizing burn scars.

Who are the source data producers?

NASA / NOAA (Suomi-NPP, NOAA-20, NOAA-21 satellites for VIIRS imagery)
NASA SRTM (Topography)
GRIDMET / GFS (Meteorological and drought index data)
MODIS (Fuel and land cover)

Annotations

Annotation process

Active Fire Labels: Formulated from NASA's baseline VIIRS active fire products.

Burned Area Labels: Derived using a union approach that combines the accumulated history of active fire detections with official fire perimeter maps.

Progression Labels: Extracted as the pixel-wise difference between the burned area masks of consecutive days, encouraging models to strictly forecast newly burned zones.

Who are the annotators?

The dataset's creators, researchers at the KTH Royal Institute of Technology.

Personal and Sensitive Information

The dataset is composed strictly of satellite and sensor data. It does not contain personal or sensitive human information.

Bias, Risks, and Limitations

Geographic Bias: As the dataset focuses primarily on the contiguous United States, deep learning models will heavily adapt to the topography, atmospheric conditions, and forest types native to North America. Sensor Limitations: VIIRS relies on optical/infrared sensing which can be physically blocked by heavy cloud cover, thick smoke, or severe weather conditions. Furthermore, the dataset operates at a resolution of 375m and 750m per pixel, meaning very small localized fires might not be easily detected.

Recommendations

Users should be aware that wildfire dynamics are highly complex. To mitigate geographic biases, researchers looking to build global active fire detection systems should augment TS-SatFire with localized data from their specific target regions.

Citation

BibTeX:

@article{Zhao2024TSSatFireAM, title={TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction}, author={Yu Zhao and Sebastian Gerard and Yifang Ban}, journal={Scientific Data}, year={2024}, volume={12}, url={https://arxiv.org/abs/2412.11555} }

APA:

Zhao, Y., Gerard, S., & Ban, Y. (2024). TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction. Scientific Data, 12.

More Information

For full benchmark performance and baseline models including U-Net, Attention U-Net, and UNETR-2D/3D architectures, please refer to the main scientific paper.

Dataset Card Authors

Initial template automatically generated and completed by AI.

Dataset Card Contact

Yu Zhao, KTH Royal Institute of Technology (GitHub: @zhaoyutim).