Twelve years of damage
The UC Irvine team processed 100,000 inspection records kept by Cal Fire between 2013 and 2024. They cover nearly 55,000 buildings damaged or destroyed, 86 per cent of them in the belt where country and city interlock. Satellite data, fire-weather analyses and building-level structural attributes were combined by machine learning into a continuous probability surface covering all of California at 100-metre resolution, with accuracy reaching 88 per cent across three successive model configurations.[1]
The leading factors form a chain. A low dew-point temperature marks the dryness of the air; high wind carries both flame and ember; at mid elevations of around 500 metres, slopes and canyons channel that wind and multiply its effect. On the vegetation side, grassland, chaparral and oak woodland stand out. The building's own share is a separate entry: wooden fences, flammable siding, protruding eaves and unprotected vents.[1]
The lever in hand
This is a statistical surface fitted to past fires, not a forecast of any particular fire. That the factors form a reproducible combination means the loss is not fated: part of the combination is weather and terrain, part is alterable building detail. Even so, the same accuracy could come from the model learning where fire arrives rather than how a building ignites; 86 per cent of the damage falling inside a single belt is on its own a strong signal of exposure.[1]
The lever sits where lead author Somnath Bar points: local jurisdictions can prioritise defensible-space inspections and fuel-reduction investment against this surface, and households can work through fence, siding, eaves and vents in turn. The map itself saves no building; what saves one is the inspection carried out and the material replaced. The measurable counterpart is plain: whether inspections in a county that adopts the map shift towards the high-risk grid squares can be read from published inspection counts over the next two fire seasons.[1]