How the fall foliage forecast works

Where the temperatures come from

Every date on this site starts as daily mean and daily minimum temperature from the ERA5 reanalysis, fetched through Open-Meteo (CC BY 4.0), which builds on ERA5 from the Copernicus Climate Change Service. A reanalysis is not a weather station. It is a physical model rerun over the historical observing record, which is what makes the series consistent from Vermont to Lapland, and it resolves to roughly 9 km. One value covers a whole valley.

Those temperatures are sampled on a 0.75 degree lattice of 4,401 anchor points, placed only over land inside the 123 regions this site covers across 51 countries. Nothing outside those regions is fetched or drawn. Where the map is blank, that is deliberate, because nothing broadleaf grows there worth forecasting.

What the model computes

Two signals drive senescence, and both are in the daily series. The first is photoperiod, which follows from latitude and date. The second is accumulated cooling, summed as the amount each day's mean temperature falls below 14 °C. When the blend of the two crosses its threshold, onset starts. Peak follows onset by about two weeks, a little sooner after a warm early autumn, and a hard frost pulls peak forward and shortens the tail behind it.

One more term matters more than it sounds like it should. The chilling threshold scales with the local seasonal temperature swing, so a mild maritime autumn has to bank more cooling before senescence counts. Without it the model cannot tell apart three places at nearly the same latitude that peak three weeks apart: Seattle (swing near 14 °C, peak around October 20), Spokane (26 °C, around October 12), and Duluth (28 °C, around September 30). Each anchor ends up with four day-of-year numbers, for onset, peak, past peak, and bare.

Elevation, and what happens between anchor points

The lattice is coarse and the map is not. Between anchors the tracker interpolates the four surrounding points, then shifts the result with 30 m terrain tiles, because higher ground turns earlier by about a day per 300 m of elevation. That is what puts an early ridge next to a later valley floor inside a single grid cell.

It is interpolation, though, not measurement. The underlying weather is still a 9 km average, so a fine cell is a good guess about a hillside and not an observation of it. Anyone reading a single slope off the map is asking it for precision it does not have.

What “updated daily” actually means

From August through November (and the first ten days of December, while the last leaves come down), a scheduled job rebuilds the whole season every morning at 06:30 UTC. It runs the model on the weather that has actually happened so far, then falls back to the 2023 to 2025 normals for the rest of autumn, validates the result, and publishes only if a date moved.

The ERA5 archive lags about five days behind real time, so each run uses observed weather through six days ago, snapped down to a three-day boundary. The 2026 season on the site right now runs on observed weather through August 26, 2026, with normals after that. So the map is rebuilt daily, but the newest real weather inside it is about a week old. Nothing here is a live satellite view of leaves.

Each region also carries a baseline peak alongside its forecast peak, which is the same model run on normal weather for that place. The difference between the two is the only honest way to say a season is running early: a region that always peaks in late September has not moved just because it is peaking in late September.

Correcting each region by hand

Ground truth is a file of 103 locations with a published peak-color window and a source URL on each one. These are climatological windows for a typical year, not single-season observations, so the model is scored against the window midpoint averaged over four weather years rather than against any one autumn. The residual per region becomes a whole-curve day shift, and 41 regions carry one today, from -9 to +8 days.

Those shifts are patches for what the inputs cannot see, and mostly that means species. A temperature series does not know whether it is sitting over sugar maple, oak, or birch, and those trees do not turn together. A region with no reference point gets no shift and ships raw model output, which is currently true of every Canadian region.

How close it lands

Scored on 2026-08-21 against all 103 reference locations over four weather years, with the per-region shifts switched off so the global model answers for itself, the weighted RMSE is 6.2 days and the mean bias is +0.6 days. Average miss is about five days. That average hides a split by continent, which is the more useful number. Across the 62 North American locations the average miss is 3.5 days and 87% land within a week. Across the 41 European ones it is 7.2 days, with the model running about three days early on average and considerably worse than that in the Alps.

Two caveats belong with those numbers. The model constants were fitted against this same reference set, so the figures are in sample and flatter the model; the honest out-of-sample number comes from a cross-validated fit, which is a separate exercise. And the references themselves are peak windows other people published, not field surveys, so “error” here measures distance from the published consensus rather than distance from the trees.

Compared against the most cited public map

In July 2026 this forecast was compared with the most widely cited public foliage map, the only one that publishes machine-readable data. Their forecast is one number per county in weekly steps, so any residual under about 3.5 days sits below their own resolution and means nothing. This measures agreement, not accuracy.

Sampled at 3,099 county centroids, the two agree within one week on 55% of counties and within two weeks on 92%, with almost no net bias. The disagreement is not scattered. Across the Plains and the upper Midwest this model runs 10 to 14 days earlier than theirs, and scored against independent reference windows their map is the one that is late: Kansas City by 20 days against our 5, Omaha by 17 against our 2, Minneapolis by 10 against our 2 the other way. In the deep South the verdict flips. Both maps run late there and this one runs later, by 8 days at Atlanta and 14 at Jackson.

The comparison is not a scorecard win. Many of the reference windows were themselves read off that same map, and the per-region shifts were fitted on those residuals, so any head-to-head total is circular by construction. What survives the caveats is narrow. In the continental interior, independent sources back this model over theirs.

What it still gets wrong

• Species composition is not an input anywhere. The seasonal-swing term gets close on climate, and the per-region shifts carry the rest by hand.
• Canada has no external validation at all. Nine regions, including Ontario and Quebec, ship raw model output with no correction, because no Canadian reference windows have been curated yet.
• The European reference set is tourism-grade. Authoritative phenology data for Europe is either registration-walled (PEP725) or published only inside a graphic (the German weather service), so those windows come from tourism boards and travel writing, converted on a fixed convention and weighted below every North American entry. The Alps are the worst region on the board, off by more than three weeks at some sites.
• The deep South runs late. Atlanta, Columbia, Jackson, and Lost Maples all peak later in the model than the references say, and a second independent forecast agrees the model is late there.
• Accumulated cooling has no floor on how early it can fire, so cold cells can start senescence in midsummer. In the 2026 grid on the site right now, 6% of populated cells peak before September 1 (277 of 4,393), and 100% of those sit above 50°N, where early September is not far off. The earliest cells are alpine, in the western Alps and the Canadian Rockies, and they land in late July, which is simply wrong. Region dates are medians and mostly absorb this. A zoomed-in mountain cell in August does not.
• Alaska rests on two reference points, and the far north generally is thin.

Frequently asked

How does the foliage color prediction work?

Every colored cell blends three signals. Daily temperatures from the ERA5 weather reanalysis (via Open-Meteo) drive a phenology model of when leaves turn: shortening day length plus accumulated autumn cooling. 30-meter terrain data shifts timing locally, since high slopes turn about a day earlier per 300 m of elevation. Finally, each region is calibrated against published peak dates, because temperature and latitude cannot see which trees are growing there. Forecast seasons use recent-year climate normals; archive seasons use that year's actual observations.

How accurate is the peak timing?

Weather truth is roughly 9 km resolution, so the fine cells are weather plus terrain interpolation rather than 1-mile weather stations. Tested against published peak dates the model lands within about five days on average, closer than that in the United States and further off in Europe, and a week or more of error happens in mountains and along coastlines. Treat the peak window as a planning estimate, and check a local report before you travel.

Why does the southern half of the map turn in April?

Autumn happens in March to June south of the equator. Scrub the date slider through the whole year to watch color sweep Patagonia, New Zealand, and Australia in April, then the northern hemisphere in September to November.