Layering coarse-but-frequent over fine-but-rare satellite views to track wildfire — from detection through spread to suppression planning

1Two kinds of satellite with opposite weaknesses

Viewing a fire from orbit involves two approaches with inverted weaknesses. A geostationary weather satellite watches the same place continuously, so it updates fast, but its resolution is coarse. A polar-orbiting imager sees fine detail but passes overhead only at intervals. Take speed and lose detail; take detail and arrive late.

This project layers the two through diffusion-based super-resolution to obtain fineness and speed together.

2Three objectives in sequence

1. 1First objectiveFuse geostationary observations with polar-orbiting imagery to detect fire rapidly from time-series imagery
2. 2Second objectiveCombine physical constraints with data-driven learning to improve spread prediction
3. 3Third objectiveIntegrate fire management knowledge with weather, fuels, terrain, transportation networks and population distributions to support suppression strategy and resource allocation
4. 4MaterialObservations and case studies from more than 100 large wildfire events across the United States

Find it, read where it is going, decide how to fight it. That the three follow in sequence is the design of this project. Improving detection alone changes nothing on the ground if it does not carry into the decisions that follow. Bringing transportation networks and population distributions into the third objective reflects that suppression is a problem of allocating limited crews and equipment as much as a physical one.

3Choosing not to discard the physics

The phrase physics-guided in the second objective marks a design stance in this field. Spread can be learned from data alone, but much of the physics of combustion and airflow is already known, and supplying it as a constraint makes the model less likely to break outside its training data. That difference matters for conditions observation does not reach and for fires at scales without precedent.

4The geosciences and AI program

Of the NSF awards this site holds as of 2026-09-02, six fall under AI and Geosciences. The same program includes an award on marine food webs and one building a digital twin for earthquakes and landslides, all sharing a direction of reworking observation that already exists. There are 70 distinct programs in total, and this one sits among those weighted toward reuse of observational data. Amounts are the obligated amount as of the check date and may change.