How this map is made

In plain words

Most foliage maps are a picture of the country colored by latitude. This one is built from the trees. For every county in the lower 48 and every national park and forest with fall canopy, we read what grows there from the federal 30-meter vegetation map and how high it stands from USGS terrain data. Each species turns on its own clock and color climbs down a mountain at a measured rate, so every place gets a peak date of its own, and a second date where a second species turns a week or more after the first.

Then we make it this year. Every morning we read September’s temperature against the thirty-year normal and shift the dates by a rate we measured ourselves, and we read every report that comes in from the ground. A report moves a date only if it passes a gate, because we measured that letting every report in makes the forecast worse.

Then we check ourselves. Every forecast is scored against long-running forest cameras that never saw it, region by region, and the scorecard is on this page.

What moves a date

Six things decide when a place peaks on this map, and each one is measured, not assumed.

• The treesLANDFIRE, the federal 30-meter vegetation survey, tells us which species stand in every county and every forecast place. Aspen, red maple, oak, cottonwood and larch each turn on their own clock, and where two share a place the map prints two dates.
• ElevationColor climbs down a mountain at about 1.0 days per 100 meters, and faster in the southern and central Appalachians, about 1.5, read from USGS terrain where the deciduous canopy actually stands, not at the county average. It’s why the ridge above Gatlinburg turns about three weeks before the valley floor.
• LatitudeThe north turns first. Latitude and elevation together, fitted against the camera record, give every place its date for a normal year.
• Big waterA Great Lake or the ocean holds color back on the ground beside it. Water that soaked up a summer of heat gives it back all fall and keeps the first hard frosts off the shore. We measured it on ten years of satellite readings at 1,120 forested points around the five Great Lakes and the Atlantic coast: about 9 days at the Lake Superior shore, fading to half of that 6 kilometers inland. It’s why Marquette and Pictured Rocks, on the Lake Superior shore, turn a week or more after the inland woods of the western Upper Peninsula. West of the Cascade crest the Pacific does the same thing to the whole lowland, about 5 days, measured the same way.
• This September’s weatherEvery day we read this month’s temperature against the 1991 to 2020 normal at NOAA’s station grids and shift the dates by 1.64 days per degree Celsius, a rate we measured over 639 camera-years. A hot September moves the map by days, and the place page says by how many.
• Reports from the groundForty forest research cameras that photograph the same trees every day, state foresters’ weekly reports, the National Phenology Network’s observers, research-grade geotagged photographs, and a decade of our own dated frames. Each is admitted through a gate that we measured against a full season, and where one has moved a date the place page says so.

The number

Our peak color forecast lands within a median of 6 days of the peak recorded by 68 long-running deciduous forest monitoring stations, and 40 of those 68 land within a week. That’s an out-of-sample number: each station is scored by a version of the model that never saw it.

Those stations are almost all eastern, so we score the West a second way, against dated photographs of known ground from Montana to southern Utah. Across 11 of them this forecast lands within a mean of 3.64 days.

Checked against the people who look at the trees

Several states publish a weekly foliage report written by their own foresters. We read the ones that are structured enough to read, work backwards from a season of reports to the day peak happened in each place, and score what we had forecast there. A camera describes one hillside; a district forester grading a county is describing the county, which is the thing you asked about.

• PhenoCam Network (GCC): within 7.1 days252 places over 18 seasons. 72 within three days. On average we run 2.1 days late against them.
• Maine Forest Service: within 3.9 days93 places over 12 seasons. 47 within three days. On average we run 0.9 days late against them.
• Pennsylvania DCNR: within 4.5 days67 places over 1 season. 31 within three days. On average we run 0.8 days late against them.
• USA National Phenology Network (Nature's Notebook): within 12.3 days16 places over 1 season. 0 within three days. On average we run 12.2 days early against them.

These are eastern states. A new source earns the right to move a forecast only after we have watched it through a full autumn and measured what its words mean in days.

Region by region

A national average hides regional error, so here is the scorecard by region: how far our dates landed from the peak the cameras recorded last season.

• Southern AppalachiansScored against 16 monitoring sites: two dates in three landed within 7 days of the camera’s peak. Average bias -0.3 days.
• New England and New YorkScored against 27 monitoring sites: two dates in three landed within 10 days of the camera’s peak. Average bias -1.8 days.
• Midwest and Great LakesScored against 13 monitoring sites: two dates in three landed within 9 days of the camera’s peak. Average bias +1.1 days.
• The SouthScored against 7 monitoring sites: two dates in three landed within 13 days of the camera’s peak. Average bias +2.1 days.
• The WestNo monitoring sites in this region, so the model is scored by satellite and by dated photographs instead.

The West is scored by satellite and by our own frames. Only 2 of our 68 camera sites sit west of the 95th meridian, so out there we score against NASA’s MODIS land surface phenology record, sampled inside each unit’s real boundary across 495 western destinations and 136 million acres, and against the dated peak frames in our own archive.

Elevation, read where the trees stand

Peak day is fitted from latitude and elevation against the camera record, and elevation does most of the work, so where we read it matters as much as the rate. For every county the number is the weighted median elevation of its deciduous canopy on the LANDFIRE map, not the county mean. Flathead County, Montana averages 1,977 meters, carried by the Whitefish and Swan ranges and Glacier’s peaks; its cottonwood and larch stand at about 1,250 meters along the rivers and the highways, which is where every visitor is. That difference alone is a week.

For every forecast place the elevation is sampled inside its own boundary, and the number we publish is the median of its river corridors, because a canyon road is where you see the color.

Every layer we pulled, and where it works

Each dataset is named with the job it does and the surface where a reader meets it.

• LANDFIRE vegetation type, 30 mThe species mix of every county and every forecast place; the second peak where two species split; the canopy gate that decides whether a place is forecast at all. Seen on the county card, the place card (What turns) and the place page.
• LANDFIRE canopy cover, 30 mThe mask under the map: color is painted only where deciduous canopy stands. Seen on the map from zoom 4.5.
• USGS 3DEP elevation, 30 mThe elevation each date is fitted at, per county (canopy-weighted median) and per place (inside the boundary, and its river corridors). Seen on the place card.
• PhenoCam NetworkThe peak definition, the calibration set, the scorecard on this page, and the live camera on the county and place cards.
• September temperature (NOAA station grids)The daily read of this September against the same days of the 1991 to 2020 normal, shifting dates by the rate we fitted. Seen as the shift on a place page when it moves a date.
• State foresters (ME, NH, VT, PA, NC)Weekly ground reports, each with its own fitted offset, gated into the correction layer. Seen on the tape and the county card, and linked on the county panel.
• USA National Phenology NetworkSpecies-level observations, the ground source that reaches the West. Gated into the correction layer.
• iNaturalistResearch-grade geotagged photographs of the four fall-color species, stage read by machine and validated against 571 frames read by eye. Gated into the correction layer; the only in-season source that reaches Montana.
• More Than Just Parks field archiveA decade of dated, geotagged frames read for the stage of each species in the frame, entered as residuals like a forester’s report, and the check the western species layer was scored against.
• National Weather Service point forecastsFrost and wind advisories inside the seven-day forecast, per place. Seen on the place page. Never a model input.
• USGS PAD-USThe boundaries and the acreage of every unit; which places exist at all.
• Our own photographsThe frames on the front and the place pages, and the film on the front door. Agency frames never open a page.

What counts as ground truth

We calibrate against the PhenoCam Network, a research camera network that photographs the same patch of forest every day and publishes a fitted curve of canopy greenness. Peak color is taken as the 25 percent decline point on that curve, the day the reds and golds are at their best, not the day greenness first starts to drop, which comes a week or two earlier and would send you north too soon.

Of the sites available we use 68 and exclude 63, every exclusion recorded: cropland, grassland, wetland and street trees (43), experimentally heated research plots (10), cameras cataloged as broadleaf but pointed at conifers (6), and records whose year-to-year swing is too large to serve as a target (2).

Our own decade in the field

The one source no other map can use. More Than Just Parks has filmed America’s public lands in the fall for more than a decade, and every frame carries a date and a location. The library holds more than 107,000 frames; over six thousand are geotagged autumn frames from the West alone, across more than 700 trip-days from the Flathead to the Colorado Plateau, which is exactly where the research cameras are not.

We read those frames the way a forester reads a stand. A Flathead Valley frame from 2 October 2018 is two readings, cottonwood along the river at 80 percent and the larch above it at 18, because those are two trees on two clocks. Each reading enters the forecast through the same gate a state forester’s report passes, and the western figure above is the model scored against them.

Official state reports

Where a state publishes a real foliage report we link it on the county panel and name the publisher, because a state forester and a visitor bureau aren’t equally disinterested about whether it’s peak this weekend. We link 8 and read 5 of them into the forecast. Where a state publishes nothing we show nothing, rather than a tourism page.

• Maine Forest ServiceME · State forestry agency
• Pennsylvania DCNRPA · State forestry agency
• Tennessee State ParksTN · State forestry agency
• Fall Color Guy (Dr. Howard Neufeld)NC · University research report
• Vermont Dept. of Forests, Parks and RecreationVT · State forestry agency
• Visit New HampshireNH · State tourism board
• I Love NYNY · State tourism board
• Travel WisconsinWI · State tourism board

Peak is the forest, not the roadside

Our date is when the canopy across a hillside is at its fullest color, the way you’d see it from a ridge or an overlook looking out over the woods. It’s measured that way too: the research cameras we score against are aimed across a forest canopy, not at the trees along a road.

The roadside turns first. A maple on a ditch or a swamp edge is stressed and gets more light than the trees behind it, so a highway can look like peak while the woods a mile in are still mostly green. That’s why so many peak reports feel early once you’re on a trail. If you’re judging from the car, you’ll see color a few days before our date. If you’re walking in, our date is the one to aim at.

When we’ll say a place is at peak

Every place on the map names a peak date and the four days to aim for. We call a place at peak only when the days you can go land within four days of that date, and the strongest thing we will say otherwise is near peak, however close it looks.

The whole year

Every date starts as a long-run expectation for its place, and the two layers that make it this year’s date move it by days, not weeks: the temperature read, and the ground reports that pass the gate. Where either has moved a date, the place page says so and by how many days.

The same phenology predicts green-up, full canopy, peak and dormancy, so the map runs the whole year: the green wave moving north in spring, canopy through summer, the turn in fall, and bare ground through winter. Spring green-up will be scored against the same cameras, and it will carry its own line on this page.