How we know what we know

Data & methodology

Every date on when.garden is traceable to a source and carries a confidence score. Here is exactly how a planting date is produced — by whichever method fits the place — and where it can be wrong.

Live dataset: every figure is produced by the automated pipeline, town by town, from raw NOAA GHCN-Daily station records and NASA POWER over the 1991–2020 normal. “Station-based” means a real station is close enough to use directly; “modeled” means we fell back to the grid. We show the method, not just the number.

How a planting date is decided

A useful planting date depends on the climate a place actually has, so when.garden picks the method that fits — and tags every figure with which one produced it:

  1. Frost-anchored (temperate). From a 30-year daily record we compute the last-spring and first-fall frost at the 50% and 10% probability levels, then offset each crop by its frost tolerance and days-to-maturity. This is the method behind our temperate-climate figures.
  2. Wet/dry-season (tropical). Where nothing freezes, sowing is timed to the rainfall calendar (monsoon onset) from gridded precipitation normals.
  3. Observed calendars. Where an authoritative planting calendar exists (e.g. FAO GIEWS, national agriculture bulletins), we ingest it directly and label it observed.
  4. ECOCROP modelling. Everywhere else, we run FAO ECOCROP's per-crop temperature and rainfall requirements against a town's monthly climate normals (a limiting-factor suitability model) to derive the window — labelled modeled, with lower confidence.

All four methods are live: the frost method — detailed next — covers temperate towns, while the wet/dry-season, observed-calendar and ECOCROP methods cover the tropics and everywhere else.

Deriving a frost date

For the frost method, the underlying data comes from one of two places:

Station-based
A real weather station is nearby

We use its daily minimum-temperature record — ideally a 1991-2020 normal — to compute the dates when the probability of frost crosses 50%. We show the source and its distance.

Modeled
No close station — we interpolate

We fall back to gridded reanalysis (NASA POWER) when the nearest station is over 50 km away or differs by 300 m in elevation. Useful, but wider error bars — always labelled "modeled", never dressed up as measured.

The confidence score

A 0–100 score combining three factors (example: a dense-station location scores ~88).

Record length30 yrs available
95
Station distance8 km away
86
Year-to-year variancemoderate spread
82
The honest disclaimer

Frost dates are probabilities, not guarantees. We headline the safe date — the P90 last-spring / P10 first-fall date, about a 10% frost risk, the standard threshold for setting out tender crops — and show the P50 median as “typical”. Cold-hardy crops can be planted nearer the median. Cold-air pockets, slopes, cities and coastlines still shift your real risk by a week or two. Treat these as a well-informed starting point — then watch your own sky.

Sources

Dataset
License
Coverage
CC BY 4.0
2,500+ crops
Public domain (CC0)
~40,000 stations
CC BY 4.0
1 km global grid
US-gov open data
0.5° global grid
ODbL
global gazetteer

All datasets are open or openly licensed · computation code published open-source · derivation method station-derived, reference window 1991-2020, min 20 years.