The data set harmonizes satellite observations of fire, atmospheric reanalysis, vegetation, fuel and topography, with a progression of the front observed every twelve hours.
Predicting the spread of a fire involves crossing things that are not measured at the same scale or pace: weather conditions, available fuels, vegetation, relief. Existing datasets often lack the resolution or coverage needed to capture how these factors interact, which hinders both the development and the comparative evaluation of models. PyroStack, presented by Arya Kondur, Giosue Migliorini, Cameron Schmitt, Francesco Immorlano and a team of seventeen researchers, aims to fill this gap.
The collection is organized by event and covers the contiguous United States as well as Alaska. It brings together satellite-derived fire observations, atmospheric reanalysis, vegetation, fuel characteristics, and topographic information on 6,994 fires that occurred between 2012 and 2024 in a wide variety of ecosystems and climate regimes. Spatial resolutions range from 30 meters to 9 kilometers, temporal resolution is hourly, and fire progression is provided in twelve-hour increments to allow for both model initialization and evaluation.
This combination of geographical extent and spatio-temporal finesse explicitly targets the two families of approaches — physical simulation and machine learning — by offering them a common basis for comparison. The authors announce future extensions to other regions and to the integration of data on firefighting operations, a component that is currently absent.
Read the full article on the original source
Go to the source →