The fall color model
Our own forecast of when the leaves turn, for every 5 km of the lower 48. How it works, what it is built from, and how far out it is on the autumns it never saw.
The popular foliage prediction maps are commercial products whose method is not published and whose terms do not allow republishing. This is not one of those. Everything below is built from open data, the method is written out here, and the error is measured and printed rather than implied.
Open the map, then switch on Fall color forecast under Earth and sky in the Layers drawer. Play the frames or drag the slider from three weeks back to six weeks ahead.
How it works
Leaf senescence is clocked by day length, which is the same every year. The weather then moves that clock by a few days. So the model is a normal date for each place plus this year's shift, and the shift is small: a fortnight either way at the extremes, a few days in most autumns.
The normal date is mid green-down: the day a place's greenness has fallen half way from its summer plateau to its leaf-off plateau. This year's date is that normal plus three terms. A run-up warmer than the rest of the country holds the leaves green and a cooler one brings the turn on. Severe drought brings it forward. And where the satellite has already seen the decline begin, the prediction is pulled toward the date that decline is heading for, which is why the recent frames are closer to observation than to forecast.
| Term | What moves it | Most it can move the date |
|---|---|---|
| Temperature | The mean nightly low over the 45 days before the normal date, against the 2001 to 2020 normal of the same days, relative to the rest of the country | 12 days |
| Drought | Severe drought or worse on the US Drought Monitor | 2.5 days earlier |
| Satellite | Where this autumn's observed decline in the vegetation index is heading, weighted by how far the decline has already gone | 21 days |
Where the colors come from
A map that says peak color over a pine plantation is wrong, so the model carries what the ground is actually under. The Forest Service's forest inventory gives the forest type group of every 30 m of the country, and the national land cover map gives how much of each cell is under trees that change at all. Under 8 per cent, nothing is drawn.
The forest type sets the hue at peak, because that is the part that is a fact about the species rather than about the date: maple, beech and birch go scarlet, oak and hickory russet, aspen and birch gold. Everything before and after peak is the same green, gold and brown everywhere.
What it is built from
| Data | Source | Terms |
|---|---|---|
| Normal turn date, shape | USGS mid green-down median 2001 to 2017, via the USA National Phenology Network | CC BY 4.0 |
| Normal turn date, level and spread | NASA GIBS, MODIS 16-day vegetation index, 24 autumns measured here | NASA open data |
| This autumn's nightly lows | NASA POWER daily minimum temperature | NASA open data |
| The fortnight ahead | Our own GFS ingest | NOAA, public domain |
| Drought | US Drought Monitor | Free with credit |
| This autumn's decline | NASA GIBS, MODIS 8-day vegetation index | NASA open data |
| Forest type group | USDA Forest Service, FIA BIGMAP 2018 | Work of the US Government |
| Land cover | MRLC National Land Cover Database 2021 | Work of the US Government |
How well it does
The test is the autumns the model was never built on. Its one fitted number comes from 2018 to 2021; it is scored on 2022 to 2025, against what the satellite went on to record, and given only what it would have had on the day it was run.
Six weeks out it is no better than the long-term average. Two to four weeks out it is, by a tenth to a quarter. That is the honest shape of this problem: until the leaves start to turn there is very little in the weather to go on.
| Run on | Long-term average is out by | The model is out by | Better by |
|---|---|---|---|
| 1 September | 8.5 days | 8.4 days | 1% |
| 15 September | 8.5 days | 8.3 days | 3% |
| 1 October | 8.5 days | 7.5 days | 11% |
| 15 October | 8.5 days | 6.5 days | 23% |
Mean absolute error over 3,760 place-years. Some of that error is in the measurement rather than the model: the same place turns 7.8 days either side of its own average from one autumn to the next, and the satellite reads the date off composites 16 days apart. An eight-day error is close to the floor of what these inputs allow.
Nearly all of the improvement comes from watching the decline rather than from the weather. Where the satellite has seen half the turn happen, the model cuts the error from 15 days to 9. Where the canopy is still green, it changes almost nothing. The temperature signal is real and measurable, 1.8 days later per degree Celsius the autumn runs warm, but it explains under four per cent of the variation, so it moves the answer by a tenth of a day. A frost term was built, tested, and removed for making the answer worse.
Every figure here comes from a script that can be run again, and the full method, the coefficients and the year-by-year numbers are kept with the code. Ask if you want them.
What it does not do
- The lower 48 only. Every field behind it stops at the border.
- It predicts a satellite's date, not a person's. Mid green-down is the average of a whole 5 km cell. In the Appalachians one cell holds a thousand feet of relief, and a single maple in a yard turns on its own schedule.
- The forecast is short. Real forecast temperatures reach 16 days. Past that the model holds this autumn's anomaly where it is, which is persistence rather than a forecast, and the drawer says which frames those are.
- It is a guide, not a booking. Use it to pick a week, not a weekend.