GardenFrost Date Lookup
Enter your address and get frost odds computed for your spot — roughly an 800-foot square, corrected for its elevation — from 26 years of NASA Daymet daily data (1995–2020). Last spring and first fall frost as 10/50/90% odds at 36, 32, and 28°F, cross-checked against your nearest NOAA station. Free, no signup.
How to read the numbers
Every result here is a band of three dates, never one date. The primary answer is computed for your spot — an ~800 ft × 800 ft square, corrected for its elevation — from what the daily record actually did there in 1995–2020.
Take downtown Salt Lake City at 32°F. On the surface, the last spring freeze was done by March 17 in 1 year out of 10, by April 13 in half of years, and by May 1 in 9 years out of 10. Plant tomatoes on April 13 and you're flipping a coin. Wait until May 1 and you're down to 1-in-10 odds.
Fall runs the same way, flipped. The first freeze has arrived by the early date in 1 year in 10 — that's your "could come this soon" warning — and by the late date in 9 of 10.
One date can't carry that information. That's why the almanac date and your neighbor's date disagree — they're both real answers to different questions about the same odds.
A climatological average is not a forecast
These numbers are climatology: what frost usually does at your spot. Any single year can swing two weeks either way, and one clear calm night is all it takes.
So use this page to plan the season — when to lay out the bed, when to start hardening off. Then, the week you actually plant, check the live forecast. Every time. A freeze warning on Tuesday outranks decades of averages.
What the surface can and can't see
The primary answer comes from a continuous fine-grained surface: 26 years of Daymet daily minimum temperatures on a 1 km grid, downscaled to ~250 m cells and corrected to each cell's own elevation using USGS 3DEP terrain. On sloped ground that removes the biggest error a station lookup bakes in — the station sitting a few hundred feet above or below you.
Your 2–3 nearest NOAA stations still appear with every answer, as the cross-check: official 1991–2020 normals, with the distance shown and the agreement measured in days. Where the surface has no data — open water, spots too warm to compute freeze odds, outside the lower 48 — the tool falls back to stations entirely and says so.
What the surface cannot see yet: cold-air pooling. Two towns at the same elevation can differ by weeks because one sits on a bench that drains cold air and the other in a basin that collects it. That terrain correction is in the works as the next stage of this tool — until it ships and passes validation, we won't claim it. It also can't see man-made frost traps — walls, berms, dense canopy. Your own logged frosts beat any model at your exact spot.
Which threshold should you use?
| Threshold | What it means | Crops on the line |
|---|---|---|
| 36°F | Light frost can form on clear, calm nights even when the thermometer reads above freezing. | Basil, cucumbers, squash seedlings, tomatoes fresh out of the house |
| 32°F | Freezing. The default — what most people mean by "frost date." | Tomatoes, peppers, beans, most summer crops |
| 28°F | Hard freeze. Ends the season for almost everything unprotected. | Kale, spinach, peas shrug off 32°F but check out around here |
Planting after frost is half the plan. Who goes next to whom is the other half.
How this works — data, method, measured accuracy
Method, in one paragraph. We take Daymet's 1 km grid of daily minimum temperatures (1995–2020), compute for every cell the empirical 10/50/90% dates for the last spring and first fall crossing of 36, 32, and 28°F, then downscale each 1 km cell to ~250 m: every fine cell takes its parent's frost odds shifted by a locally-estimated temperature lapse from the parent's mean elevation to the fine cell's actual elevation, both from USGS 3DEP terrain. Everything is precomputed; the live lookup reads 36 bytes for your cell. No zones, no station-snapping, no made-up precision — where the record can't support a date, you get an honest "no data."
That's the short version. The full write-up — every step, every source and license, the validation design, and the accuracy bar we had to change — lives at how we compute frost dates.
Data sources
- Daymet V4 — daily minimum temperature, 1 km (NASA / ORNL DAAC) The base climate record: 26 years (1995–2020) of daily gridded minima. Public domain (CC0-equivalent); the open-access archive we bake from ends 2020-12-30, which caps the window.
- USGS 3DEP — 1 arc-second (~30 m) elevation The terrain behind the elevation correction. US federal public domain. At a small number of documented coastal gaps the bake falls back to Copernicus DEM: contains modified Copernicus DEM GLO-30 data © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA; all rights reserved.
- NOAA NCEI — U.S. Climate Normals 1991–2020, freeze/frost probabilities The ground truth. ~7,300 stations' first/last freeze dates at 10–90% probability, used as the on-page cross-check, the fallback where the surface has no data, and the held-out validation set for the accuracy numbers below. US federal public domain.
- US Census Bureau — Census Geocoder Turns addresses into coordinates and ZIPs into centroids. Public domain. Coordinates you type and the "use my location" button leave your browser only to run the lookup.
Measured accuracy — validation against held-out NOAA stations
Accuracy is measured by baking the surface, hiding real NOAA stations from it, and scoring the surface's median-date predictions against what those stations recorded. Errors are in days; the baseline is a "nearest station" lookup — what almanac-style tools do.
| Validation test | Surface error (median-date MAE) | Nearest-station error (same test) | Error cut | 10–90% band held |
|---|---|---|---|---|
| Utah pilot — complex terrain (69 stations) | 6.3 days | 10.3 days | 38.9% | 92% of cases |
| Central Nevada pilot — complex terrain (34 stations) | 6.1 days | 11.1 days | 44.7% | 94% of cases |
| Lower 48, all terrain (6,857 stations scored) | 6.0 days | 7.2 days | 16.8% | 94.3% of cases |
| Lower 48, complex terrain (1,297 stations) | 7.6 days | 11.0 days | 31.2% | — |
The short version, measured coast to coast: about a 6-day median-date error nationally, and the surface earns its keep exactly where frost gets weird — in complex terrain where the nearest station is also far away, it measured roughly 4–5 days more accurate than a nearest-station lookup (a 35% error cut where the nearest station sits more than 20 km out). Where the country is flat and stations are dense — the South — a good nearby station is already as accurate as the surface, so the tool shows you both and lets the agreement speak. Two honest weak spots we found and kept: coastal California and the marine Northwest, where frost doesn't come every year and median dates are shaky for every method we tested — there, trust the 10%/90% ends of the band over the middle.
The honest limits
- It can't see cold-air pooling yet. Elevation is corrected; drainage is not. A frost pocket in a basin can run colder than this surface says. We've run the cold-air correction as a pre-registered public experiment twice — neither attempt cleared its accuracy bar, so it doesn't ship. The full record, misses included, is in the experiment log; a redesigned attempt is in the works.
- Marginal-frost coasts are hard for everyone. Where frost doesn't arrive every year — coastal California, the marine Northwest — the national validation measured 10–20-day median-date errors for every method tested, station lookups included. The date odds there are honest but wide; lean on the 10% and 90% ends of the band, not the median.
- The window caps at 2020. The open-access Daymet archive ends 2020-12-30, so the surface is 1995–2020 (26 years). Station cross-checks run 1991–2020. An extension rebake is planned.
- An ~800 ft square is not your exact bed. Walls, berms, canopy, and other man-made cold traps are invisible to a bare-earth model. Log what actually happens in your garden — your ground truth beats any model at your spot.
- It's climatology, not a forecast. Any single year swings. Check the live forecast the week you plant.
FAQ
What do the 10%, 50%, and 90% frost dates mean?
They're odds from decades of daily records — 1995–2020 for the fine-grained surface, 1991–2020 for NOAA stations. For spring, the 50% date means half of years still get a freeze after it. The 90% date is the careful one: only 1 year in 10 sees a freeze later. Same 1-in-10 / median / 9-in-10 framing NOAA uses.
When is it safe to plant after the last frost?
For tender crops — tomatoes, peppers, basil — wait for the 9-in-10 date at 32°F, then check the 10-day forecast before you plant. Hardy crops like peas and kale can go in weeks earlier, closer to the 28°F dates. The live forecast always outranks a climatological average.
Why does the tool show three temperatures?
Because plants die at different numbers. 36°F can burn tender things like basil and cucumber seedlings, 32°F takes out most summer crops, and 28°F is the hard freeze that ends even the tough ones. Pick the threshold that matches what you grow.
How accurate are these dates for my yard?
Measured, not promised. Validated against 7,006 held-out NOAA stations across the lower 48: the surface's median-date error is about 6 days nationally, and roughly 4–5 days better than a nearest-station lookup where terrain is complex and the nearest station is far away. The 10–90% band held in 94% of cases. Where stations are dense and the land is flat, a good nearby station ties the surface — so the tool shows you both. It still can't see cold-air pooling in your valley; we're working on that correction in public and it hasn't cleared its accuracy bar yet. For what the usual nearest-station method costs, we measured that too: 13.1% of stations within ten miles of each other disagree about the last frost by two weeks or more.
What if my area never gets a freeze?
Then the tool says that instead of inventing a date. Where the record is too warm to compute reliable freeze odds — most years, no freeze — the surface reports no data honestly and the tool falls back to your nearest NOAA stations, which say the same thing in their own numbers.
Why does the surface data stop at 2020?
The open-access Daymet daily archive this surface is baked from ends December 30, 2020, so the window is 1995–2020 — 26 years. NOAA station cross-checks use 1991–2020. A rebake extending the window is planned; frost climatology moves slowly, so 26 recent years is a solid base — but we would rather tell you the cap than hide it.
This lookup's free and it stays free. No signup, no wall.
It's also the data behind the frost alert we're building into GardenTrack — the garden app in beta now. The plan: it knows your garden's frost odds from this same surface, watches your local forecast, and gives you a heads-up when frost's actually coming — so you're not the person carrying 40 seedlings inside at 11pm on a hunch.
Meanwhile the one-off math lives here in the free tools: how many plants fit in the bed, who plants next to whom, and now when frost actually ends.