Growth-only precision & recall of fire-spread forecasts against Cornea perimeter snapshots. Each run is scored against perimeters in the 7 days after it, each at the perimeter's own time (no window or tolerance): every perimeter with a mapped time, the latest one published on a day without one (timed by its last edit), and the perimeters scored before November 2026. One row per (model, fire, run, weather percentile, perimeter), all kept: no minimum growth or off-grid filter. Predicted area = time of arrival before the perimeter's time; both sides subtract what had burned when the run began: the latest perimeter mapped before the run, and where the forecast itself started burning (Pyrecast seeds its start with satellite detections beyond the perimeter). How it's calculated, with a worked example.
Every scored row at its exact time after the run (no averaging buckets). Filled dots: perimeter with a mapped time; open dots: timed by its last edit.
Fires contributing rows under the current filters. Click a name to open it on the forecast map.
Pick any scored row (or click one in the table below) to see the exact areas behind its precision & recall — recomputed in your browser from the same archived time-of-arrival raster and Cornea perimeters the pipeline used.