Open data

The forecast this site draws is a handful of files, and this is where they live. Each one covers the 2026 season. For any place in it you get the day color starts, the day it peaks, the day it goes past peak, and the day the trees are bare. Those dates come out of a phenology model run on ERA5 daily temperatures and corrected for elevation, not from an average of what past Octobers did.

Take them. The license asks for a credit line and nothing else. There is no API and no key to request, because these are static files sitting on a CDN, so there is nothing here to rate limit.

The files

  • Region forecasts (CSV) /data/regions.csv
    One row per region: onset, peak, past peak and bare, as day-of-year and as dates.
  • Places (CSV) /data/spots.csv
    One row per mapped place, with coordinates and its own sampled peak date.
  • Season 2026 (JSON) /seasons/2026.json
    The season file the site itself renders from, including grid metadata.
  • Places (GeoJSON) /spots.geojson
    The same places as a FeatureCollection of points, ready for a map library.
  • Region shapes (GeoJSON) /regions.geojson
    Region boundaries, derived from Natural Earth.
  • Region registry (JSON) /registry.json
    Names, countries, hemispheres and centroids for every region we cover.
  • Anchor grid 2026 (binary) /seasons/2026.grid.bin
    The raw model output the season file describes, at 8 bytes per grid cell.

The two CSVs are generated from the same data the region and place pages render, so a row and its page never disagree. Both carry a url column if you want to link back to the page a number came from.

Reading the grid file

The .grid.bin is the model output before any region or place has been sampled out of it, and it is the only file here that is not self describing. It is a flat array of cells in row-major order, 60 rows by 269 columns, starting at latitude 24.75 and longitude -170.25 and stepping 0.75 degrees each way. Every cell is 8 bytes, little endian: onset as a uint16 day of year, where 65535 means the cell has no data, then three uint8 offsets carrying peak, past peak and bare as days from the milestone before, then confidence as a byte over 255, then the cell’s reference elevation as a uint16 with 1000 added so ground below sea level still fits.

The geometry above also sits in the grid object inside the season JSON, which is the copy to trust if the two ever part company. Sampling a point means bilinear interpolation across its four surrounding cells and then a shift of about a day per 300 m of elevation difference. Our reader for it is SeasonGrid in src/lib/seasons.ts.

How often it changes

Through the summer the season file is a pure forecast, built from the last few years of ERA5 normals. From mid-August it gets rebuilt against temperatures that have actually happened, so peak dates drift by a few days as real weather replaces the normals, and they tighten as the season closes in. Whichever kind you have is stamped in the season file. generatedAt is when it was built, and every region carries its own kind, either predicted or actual. The copy on this page was built September 2, 2026.

Past seasons stay put once archived. If you need a stable snapshot, take a dated copy rather than fetching the live file, because the current season’s numbers are meant to move.

License

Creative Commons Attribution 4.0 (CC BY 4.0). Use it commercially, change it, build on it. The only condition is that this line travels with whatever you make:

Autumn Guide (https://www.autumnguide.com), built on ERA5 via Open-Meteo (CC BY 4.0)

Open-Meteo publishes under CC BY 4.0 and the underlying reanalysis is ERA5 from the Copernicus Climate Change Service, which is why their names ride in that line alongside ours. Region boundaries in the GeoJSON come from Natural Earth and are public domain. Photographs on the site are not covered by any of this; each one carries its own credit and license on the page it appears on.

Cite as

Autumn Guide. (2026). Autumn Guide fall foliage forecast, 2026 season [Data set]. https://www.autumnguide.com/data

Add the date you retrieved it if the season was still live when you pulled it. For a specific region, cite its page instead, for example https://www.autumnguide.com/regions/vermont.

If the dates look wrong

A region that peaked a week off what we said is worth an email to hello@autumnguide.com. Local observations are how the model gets corrected, and mountains and coastlines are where it is weakest. More about how it works is on the about page.