FireCastNet: earth-as-a-graph for seasonal fire prediction.

Area of Science:

• Earth System Science
• Artificial Intelligence
• Climate Science

Background:

• Climate change exacerbates fire weather conditions globally.
• Accurate seasonal wildfire forecasting is crucial for disaster preparedness and ecosystem management.

Purpose of the Study:

• To introduce FireCastNet, a novel deep learning architecture for global wildfire prediction.
• To leverage the SeasFire dataset for forecasting burned area patterns up to six months in advance.

Main Methods:

• Developed FireCastNet, combining 3D convolutional encoding with Graph Neural Networks (GNNs).
• Utilized the SeasFire dataset, a multivariate Earth system datacube.
• Modeled Earth as an interconnected graph to capture spatio-temporal dependencies and teleconnections.

Main Results:

• FireCastNet demonstrated superior performance in global burned area forecasting compared to state-of-the-art models.
• Achieved strong results in fire-prone regions like Africa, South America, and Southeast Asia.
• Longer input time-series and spatial context integration improved prediction robustness and accuracy.

Conclusions:

• FireCastNet offers a significant advancement in long-term wildfire prediction.
• Modeling Earth system interactions is vital for accurate wildfire forecasting.
• The approach enhances region-specific predictions through local area modeling techniques.