Methodology
How this forecast is built
No black boxes: every step between raw weather data and the number on the meter.
A baseline peak date per location
Every tracked place has a historical median peak plus observed peaks from recent seasons. Those are blended, weighted toward recent years, into a baseline date. This is the calendar anchor, set by day length, which weather cannot move. Early in September, before enough of the season has actually happened, the forecast stays on that baseline rather than over-reacting to a few days of data.
Live weather, elevation-corrected
We pull daily observed and forecast temperature, rainfall and evaporation demand for each location's actual coordinates, then apply a 3.5°F per 1,000 ft lapse correction so a ridge-top reading is not treated like a valley reading.
Shift the date
Cold nights below 45°F pull peak earlier. A warm anomaly against the seasonal normal pushes it later. Dryness since September 1 accelerates leaf drop: we use the water balance between rainfall and evaporation demand where we have it, the official US Drought Monitor category where we do not, and a high vapour-pressure deficit compounds both. In most seasons the combined shift lands within a few days either side of the baseline; the weather term is allowed to reach twelve days in genuinely unusual years.
Correct against what we can see
Satellite canopy greenness tells us whether the forest is already ahead of or behind schedule, and can move timing by up to two and a half days once there are enough clear looks at the canopy. A learned bias correction folds in how far off our own forecast was at that location in past scored seasons, capped at five days and deliberately under-applied so one odd year cannot swing it. After all of that, the total shift is capped at fourteen days in either direction. The model is never permitted to move peak more than two weeks off its daylight anchor.
Score the intensity
Vibrancy is separate from timing. Cool-but-not-freezing nights plus sunny days plus adequate soil moisture produce saturated reds. Drought, warmth and dry air produce yellows and browns. Vibrancy caps how high the color index can reach in a given year.
Build the daily curve
For each day we compute distance from the predicted peak and run it through a logistic rise, a four-day plateau at full color, and a faster Gaussian decay, because leaves come on slower than they fall off. That produces the 0-100 color index and its stage.
Pull today toward the real canopy
Today is not left to the model alone. Recent satellite passes and scored frames from park webcams both nudge today's reading toward what the canopy actually looks like, weighted by how fresh and how readable the image is. Any correction we make is then carried forward and faded out over the following two weeks, so the next few days stay continuous with what you can see right now instead of snapping back to the curve.
Warn about leaf drop separately
Color and loss are different questions. Active National Weather Service warnings (high wind, winter storm, hard freeze, flooding) plus raw forecast gusts, heavy rain and hard freezes are scored onto the same day grid as a separate leaf-drop risk, scaled by how much color there is to lose. A ridge can read near peak and still be flagged high risk for the same afternoon.
Attach honest confidence
Confidence starts high where we have observed weather and decays with forecast distance. Peak probability widens as confidence drops, so a date three weeks out shows a wider, humbler window than tomorrow does. Alongside it we publish a reliability horizon: the date real evidence runs out and climatology takes over, which stretches further when satellite and camera readings are fresh.
Score the trip
Your trip score blends the best available color each day, peak probability, how many strong alternatives you have, and how many distinct elevation bands you can reach. A short trip on exactly the right date and a long trip spanning several bands can both score well.
Where each number comes from
Every input is a public, checkable source. Here is what we use and how often it updates.
- Daily weatherRefreshed nightly
- Observed and forecast daily temperature, rainfall and evaporation demand for each location's own coordinates, from the Open-Meteo weather and reanalysis service.
- Long-term climate normalsStored once, rarely changes
- Thirty-year day-by-day averages for each location, used to judge whether this autumn is running warm or cold against its own normal rather than a park-wide average.
- Satellite canopy greennessChecked daily
- Sentinel-2 imagery over each tracked spot, turned into greenness measures. Cloudy passes are marked unusable rather than quietly averaged in.
- Drought statusChecked daily, published weekly
- The official US Drought Monitor category for the county each location sits in, used when we do not have a clean local water balance.
- Weather warningsChecked hourly
- Active National Weather Service alerts for the park counties. These drive leaf-drop risk and the caution flags on the road cards.
- Park webcamsCaptured three times a day
- Frames from public park and partner cameras, scored for how much color is in the canopy. Fogged, dark or offline frames are kept and labelled as unreadable, never dropped to make the record look better.
- Past seasonsRe-scored weekly
- Recorded peak dates and our own past forecasts for the same locations. This is what the learned bias correction is trained on, and what the accuracy page reports.
- Nationwide county mapRefreshed daily
- County-level weather across the lower 48, combined with latitude, elevation and regional timing to produce the coarser national estimate.
What this model cannot do
- The color index itself does not drop for a single wind event. We flag those separately as leaf-drop risk, but one night of 50 mph gusts during peak week can still strip an exposed ridge faster than any forecast can follow.
- It works at the scale of a location, not a single tree. A sheltered north-facing cove can run a week behind the overlook above it.
- Long-range forecasts past about three weeks lean on climatology, so treat them as planning ranges rather than promises.
- The nationwide map is a coarser, county-level estimate built from latitude, elevation and regional timing. Only the Smokies locations get the full weather, satellite and camera treatment described above.
- It does not know about live road closures. Road cards follow standard seasonal park schedules with caution flags from active weather alerts, so always confirm with the park the morning you travel.