Our tracking system is based on data from firms, a publicly available system set up by America’s National Aeronautics and Space Administration (nasa), which was originally designed to detect forest fires. We have built a machine-learning algorithm that evaluates the location of each fire detected by firms, and assesses whether or not it is related to the war. It consists of 100 separate models built to predict fire activity in non-war years. If at least 95 of these models agree that the number of fires identified at a given place and time is abnormally high by pre-war standards, the algorithm marks these “extra” fires as war-related.

Also explained / showcased in this YouTube video: The Economist - War in Ukraine: Tracking the fighting with satellites