Pick a point to see the inputs behind its risk class: the gridded 2 m minimum, dew point (the overnight floor — a low dew point removes the latent-heat brake and enables black frost), the estimated leaf minimum, and the night's wind & cloud (calm + clear = stronger radiative cooling).
Engine: for each of 10 representative points, hourly 2 m temperature, dew point, wind, cloud and soil moisture are taken at the coldest hour of each local night. The leaf/canopy minimum is estimated as the gridded 2 m minimum minus a radiative offset (0 °C under wind/cloud → ~4.5 °C under clear, calm, dry conditions), floored near the dew point (latent-heat floor). That estimated leaf minimum is classified into None / Watch / Moderate / High / Severe. A black-frost flag marks damaging+ classes reached under a low dew point. Data: Open-Meteo (recent actuals + 7-day forecast; ~7 days is the skill horizon for minimum temperature). Generated 2026-06-20.
Calibration: class thresholds are proxy-calibrated to the gridded ERA5 distribution of historical frost events, not raw leaf-kill physiology. ERA5 grid cells run warm versus real frost pockets, so the model is validated to light up the southern tier (Severe + black frost at the coldest cells) in the 1994 and 2021 frosts while staying quiet in mild winters (2023 control). 1975 — the century's worst frost — is excluded from validation: ERA5's pre-1979 back-extension renders it as a mild winter, an input-data limit rather than a model failure. Read the signal as a regional tendency; the absolute classes describe conduciveness relative to past events.
Scope: frost is a market-scale risk to coffee essentially only in Brazil (subtropical 18–23°S, austral-winter polar intrusions); equatorial origins do not frost. The Patrocínio and Araguari points are deliberate low-risk benchmarks — their lighting up signals an exceptionally deep cold outbreak.