A third kind of spread model neither writes an energy balance nor fits a small closed-form curve. It trains a statistical model — often on satellite-derived fire progression, gridded weather and fuel maps — to predict directly where the fire front will be at the next observation, learning whatever relationship the data contains. Its attractions are real: it can be evaluated in milliseconds once trained, it can absorb inputs a physical model has no term for, and it is fitted to actual wildfires rather than laboratory fuel beds. Its weaknesses are equally real and follow from the same fact. It can only be trusted in conditions resembling its training fires, so the extreme fires people most want predicted are the ones least represented; it will happily produce a confident answer outside that range with no signal that it has left it; and its output rarely comes with a mechanism you can argue with, so a wrong forecast is hard to diagnose. Position it on the model ladder as an empirical model with a very large number of fitted parameters. After this Concept you can state what a machine-learned spread model is fitted to and where its confidence is unearned.
Linear regression can be applied to spatial, pixel-level data by treating multiple environmental factors as stacked input features per spatial unit (pixel) and learning a relationship between those f…