TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction.

Area of Science:

• Earth and Environmental Sciences
• Computer Science

Background:

• Wildfire monitoring and prediction are crucial for understanding fire behavior and mitigating risks.
• Earth observation data offers vast potential for enhancing wildfire analysis.
• Multi-task deep learning models can integrate diverse data sources for improved wildfire insights.

Purpose of the Study:

• To introduce a comprehensive multi-temporal remote sensing dataset for wildfire research.
• To support three key tasks: active fire detection, daily burned area mapping, and wildfire progression prediction.
• To provide a foundation for advancing wildfire research using deep learning.

Main Methods:

• Development of a multi-temporal remote sensing dataset covering U.S. wildfires (2017-2021).
• Inclusion of surface reflectance images and auxiliary data (weather, topography, land cover, fuel).
• Utilizing multi-task deep learning for pixel-wise classification (detection) and integrated data modeling (prediction).

Main Results:

• A 71 GB dataset with 3552 images and detailed wildfire lifecycle documentation.
• Manual quality assurance for active fire (AF) and burned area (BA) labels.
• Established benchmarks for the three supported wildfire research tasks.

Conclusions:

• The presented dataset and benchmarks are foundational for deep learning-based wildfire research.
• This resource facilitates advancements in active fire detection, burned area mapping, and wildfire prediction.
• Enhanced wildfire monitoring and prediction capabilities are achievable through integrated data and advanced modeling.