Methodology
How our fall foliage forecast works
Every peak date on this site is a prediction with a stated confidence range and a traceable source. This page explains where the numbers come from, what they can and cannot tell you, and where we do not have enough data to answer at all.
Model climatology_v0 · rebuilt September 27, 2026 · latest satellite observation September 25, 2026
What we predict, and what we do not
We predict when peak color is likely to arrive at a specific place, expressed as a date and a window either side of it. That is a timing forecast built from a multi-year climatology — how this location has behaved across the satellite record.
We do not currently publish a live report of today's color. The model takes no weather inputs, so it cannot know that this particular October has been warm and wet. A page that showed a climatological date under the words “current conditions” would be claiming an observation we did not make, so we label predictions as predictions.
The practical consequence: use these dates to choose a travel week months ahead, which is what they are good at. In the last few days before a trip, a local webcam or a ranger station will beat any climatology, including this one.
How a peak date is calculated
Follow the canopy through the satellite record
For each location we pull the seasonal vegetation-index curve from NASA MODIS MOD13Q1 vegetation indices across every available year, at the sample points that fall inside the site rather than at a single coarse grid cell where we can.
Find the day the canopy draws down
Peak is taken as the day the index falls through a calibrated fraction of its seasonal amplitude — currently 65% of the drawdown. Greenness alone does not mark peak color; the transition out of green does.
Calibrate that threshold against the ground
The threshold is fitted against PhenoCam canopy-greenness cameras — daily photographs of real forest canopies, with a fitted band of about ±12.6 days. This is what ties a satellite definition of “peak” to something a person standing under the trees would agree with.
Average across years, and keep the spread
The published date is the multi-year mean. The ± figure beside it is one standard deviation of the observed years, so roughly two autumns in three land inside the stated window. Across the whole dataset the mean confidence range is ±8.2 days.
How much evidence stands behind each page
Not every location has equal evidence, and the site says which is which rather than presenting 589 identically confident pages.
| Evidence level | Locations | Share | What it means |
|---|---|---|---|
| Observed at the site | 573 | 97% | The location has its own usable satellite history — either its own sample points or a satellite cell with labeled years. These are the strongest predictions on the site. |
| Interpolated from neighbors | 12 | 2% | The location sits in a cell without usable observations, so the estimate is borrowed from the nearest cells that do have them. A real prediction, weaker evidence, and labeled as such on the page. |
| No local data | 4 | 1% | Neither the site nor its neighbors resolved. These pages show no peak date at all. We suggest nearby locations that do have observations instead of printing a national average under a local place name. |
Data sources
Every dataset the pipeline draws on. The peak dates published today come from NASA MODIS MOD13Q1 vegetation indices; the others feed calibration, species and elevation context, or work that is not published yet. We list all of them rather than only the ones that flatter the model.
| Source | Product | Cadence | Latency |
|---|---|---|---|
| Copernicus DEM GLO-30 | Copernicus DSM COG 30 m | static release | — |
| Drawdown to canopy-percentage crosswalk | derived | refit as needed | — |
| gridMET daily climate | gridMET | daily | 2 days |
| LANDFIRE Existing Vegetation Type | LF2025 EVT | static release | — |
| MODIS MOD09GA/MYD09GA daily surface reflectance | MOD09GA | daily | 3 days |
| MODIS MOD13Q1 16-day vegetation indices | MOD13Q1 | 16-day composite | 18 days |
| NASA SRTM GL1 elevation | SRTMGL1 v003 | static | — |
| National Land Cover Database | NLCD 2025 land cover | static release | — |
| PhenoCam canopy greenness (GCC) | PhenoCam 90th-percentile GCC | daily | 1 days |
| USDA FIA forest species composition | FIA DataMart | annual | — |
Previously used and since replaced: MODIS MCD12Q2 land cover dynamics, USA-NPN phenophase observations. We list these because older write-ups elsewhere may still cite them.
Known limits and failure cases
- No weather in the model. A drought summer, an early hard frost or an unusually warm October can shift real peak by a week or more. The confidence range covers typical year-to-year variation, not an exceptional year.
- Satellite lag. The composite products behind the climatology arrive on a 16-day cycle with roughly two and a half weeks of processing latency — currently 2 days behind. Nothing here is same-day.
- One date per destination. A large park spans elevations that peak days apart. Where we have enough sample points we publish that internal spread on the location page; treat the headline date as the middle of the site, not the whole of it.
- Evergreen and mixed country. In areas with little deciduous canopy the seasonal signal is weak and the fitted date is correspondingly less meaningful. Location pages carry the deciduous fraction so you can judge that yourself.
- 4 locations have no local data. Those pages show no peak date. That is deliberate: a national average printed under a specific place name is a wrong answer that looks like a right one.
Corrections
If a date here contradicts what you saw on the ground, that is useful to us — ground observations are how the calibration improves. Every location page states its own evidence level and update date, so a report is checkable against the exact prediction that was live at the time.