Satellites like VIIRS are crucial for monitoring fires, providing data on location and intensity. However, they often misidentify human-made heat sources like landfills, gas flares, and power plants as actual vegetation fires. This misclassification complicates wildfire monitoring and related analyses.

Our research addresses this by first compiling a global database of potential static heat sources (volcanoes, refineries, etc.). We then analyzed VIIRS detections, examining their temporal persistence and spatial proximity to this infrastructure. Our findings confirm that numerous high-persistence detections near human infrastructure are indeed incorrectly labeled as vegetation fires, often exhibiting distinct heat signatures on non-burnable land.

To resolve this, we've developed a machine learning model. Trained on verified fire events versus persistent static emitters, it leverages features like detection persistence, Fire Radiative Power (FRP), land cover type, and spatial context. Our goal is a more accurate classification system that minimizes misidentifying industrial heat as wildfires. This will significantly enhance the reliability of satellite fire monitoring for policy, emissions tracking, and wildfire management.