Frost in cold hollows: why the forecast says 4 degrees and you wake up to ice

The forecast said 4 degrees

It is the same conversation every winter. The area forecast called for a low of 4 °C, and at seven in the morning the bottom of the orchard is burned while the upper part, three hundred meters away, has not lost a single leaf.

It is not that the forecast was wrong: it was not about your plot. The reanalysis and forecast models feeding almost every app work in cells 9 to 25 kilometers across. That cell averages hills, valleys and flats, and reports the air temperature two meters above that average surface. Your 300-meter-wide hollow does not exist on that map.

Cold air runs downhill

The physics is simple, and it is what makes the problem predictable. On a clear, windless night the ground loses heat by radiation to the sky; the air touching it cools, becomes denser and, like any heavy fluid, drains down the slopes to the lowest point, where it pools. That is a thermal inversion: colder below than above, the opposite of normal.

You do not need a canyon. A 10 to 15 meter depression with a flat bottom is enough to dam cold air, and the difference between the rise and the bottom, inside the same plot, reaches 3 to 5 °C. That is why radiation frosts are so local and so repetitive: the same low spot burns every year, and the map that predicts it is not the climate map, it is the terrain map.

It is also why the mitigation measures are what they are: move the air (fans, helicopters), add heat or water (sprinkler irrigation, which releases heat as it freezes), and above all do not plant the most sensitive material at the bottom. The calculator below makes the correction the area forecast does not.

What we measured, and what we found

Before building this we tried the obvious: using climate reanalysis to count sub-zero nights per plot. In a well-known frost area of Los Altos de Jalisco, the 9-kilometer reanalysis returned zero sub-zero nights in ten years. That is not a bug in the model, it is its resolution: the cell averages the valley, and the average never drops below zero even when the bottom does.

So risk is built from three pieces, and each one is honest about what it measures (it is part of satellite monitoring):

• Terrain (30 m elevation model): the plot's elevation minus the mean of a ring at 1.5 km —the topographic position index— plus local slope. Negative and flat means a hollow. It is static: computed once and it does not change.
• Night surface temperature from satellite (1 km, two passes a night, around 22:30 and 01:30): how many nights per season that specific plot dropped below 0 °C, over ten November-to-March seasons. It is an observed frequency, not a model.
• Forecast seven days out for the area, corrected to each plot by elevation and topographic position. It is the only one of the three you can act on.

Surface temperature is not air temperature

This nuance decides whether the number is useful or misleading, so it goes stated. The satellite measures surface temperature —soil, leaf, roof—, not the two-meter air temperature a weather station reports. On a clear night the surface cools further than the air, so a night with a surface at −1 °C can correspond to air at 2 or 3 °C.

That is why what we publish is a frequency comparable between plots —how many nights per season this plot crossed the threshold, against what the others crossed— and not a thermometer. To decide where to plant the sensitive material, that comparison is exactly what you need. To decide whether to irrigate tonight, the corrected forecast is what rules.

And there is one more limit worth knowing: clouds block the measurement. A cloudy night is not a frost-free night, it is a night with no data, so the frequency ships alongside the fraction of nights actually observed.

How the forecast is corrected to your plot

Two corrections, both with numbers you can check. By elevation: 6.5 °C per kilometer of difference from the area's reference point; sixty meters higher is 0.4 °C less, and that already moves the verdict when the forecast hovers around 2 °C. By terrain: an extra 1.5 °C discount if the plot sits in a hollow, which is the cold-air pooling the model cannot see.

And the alert threshold is not zero but 2 °C, or 4 °C in a hollow. The reason is honest: forecasts carry error, and the margin is what covers it. Alerting at exactly 0 °C would mean alerting late half the time.

The actionable part is this: with two or three days' notice you can irrigate, move air, delay a pruning that would push tender shoots, or bring a pick forward. The rest —terrain and climatology— will not change tonight, but it does decide what gets planted at the bottom of the plot and what gets planted on the slope.