Autumn in Cluj, seen from a satellite: when the trees turn in each forest, park and district

The forests’ half-way day, year by year#

From one year to the next all the forests move together, earlier or later. Each autumn’s day is read straight from the images: a straight line is drawn between the last image in which the forests are more than half green and the first in which they are less, and the half-way day is where it crosses half. The closer the two images, the surer the day.

How sure the dates are#

What is measured. The vegetation index NDVI (normalised difference vegetation index), computed from the red and the infrared light that leaves reflect. For every square, the summer level is the median between 15 August and 19 September and the leafless level the median between 11 and 30 November. The half-way day is the day on which a curve fitted to all clear-sky measurements passes half-way between the two levels (half of the distance between them, not half of the index’s summer value).

What colour a forest is on the half-way day. Colour was measured separately, from the three bands of visible light (blue, green, red), on crowns: squares of trees surrounded only by squares of trees. On the colour wheel green is at 120°, yellow at 60° and orange at 30°. In summer the forests are at 118°. On the half-way day they were, in the six autumns in which it can be read, between 46° and 63°: from yellow towards orange. Blue falls from 26% of visible light to 12–19%, and the forest reflects more visible light than in summer, so it is neither grey nor dark brown. Red and green are then in almost equal parts (104 to 100; in summer, 61 to 100). In the city, at the same point, the crowns in the districts are between 53° and 67°, also yellow but paler; those in the Central Cemetery, between 65° and 79°, still lean towards green. The share of red reaches its October maximum about eight days after the half-way day: anyone who wants the forest rust-coloured will find it about a week after that day. The photographs show the same: on 22 October 2024, three days before that year’s half-way day, the forests are yellow-green; on 29 October, four days after it, they are orange. The colour in a single image also depends on the air that day, which is why ranges are given. Red light also enters NDVI, and green and blue light do not, so the two measures are only partly independent. Towards the end of November the leafless forest, under a low sun, has an even larger share of red without being more colourful.

Two clocks of autumn. A second index, EVI, which also takes blue light into account, reaches its own half-way point before NDVI does: in the forests by one to seven days (three on average, in six autumns), in the crowns of the districts by up to eight days. At the EVI half-way point the forests are still yellow-green (61°–71°). Wang and colleagues (2024) compared the two indices, from MODIS images with squares of about five kilometres, with observations on the ground in the United States: the end of season read from NDVI (the mean of the half-way point and the day of fastest decline) came on average 16 days after the day on which half of the leaves were coloured, and the one read from EVI a day after it. In Cluj, with 10-metre squares, the two indices are a few days apart, and there are no ground observations to compare them with. The half-way day is therefore the day on which a forest, seen from above, is yellow shading to orange. The yellowing starts before it: the share of red is half-way through its rise three to four days earlier.

The second satellite. The American Landsat satellites photograph the same place less often, with 30-metre detail. For Landsat the curve was fitted separately, on squares fully covered by forest. In six autumns there is one Landsat image in which the forest is still more than half green and a later one in which it is less than half. The day found here falls between those two Landsat images in four of the six autumns; in the other two it lies outside them by at most 3.1 days. The two Landsat images are 7 to 24 days apart, so this is a broad check, not one to the day. Where the Landsat images are at most 16 days apart (four autumns), the day computed from Landsat alone is one to four days earlier; with another choice of the summer and leafless levels, up to seven. On days when both satellites passed, Landsat saw the forests, in mid-change, about 3 percentage points less green than Sentinel-2: what a forest loses in less than a day. The rest of the difference is unexplained; most likely it comes from the small number of Landsat images. So the days on this page are uncertain by a few days and, against Landsat, tend to be later.

Three readings of the day. When two groups of trees are compared year by year, each one’s day is read in three ways: between the last image above half and the first one after it below half; between the first two consecutive images that step from above half to below it; and from a curve fitted to all the autumn’s images, which a single bad image cannot move far. The difference in a year is the median of the three, and the page says “whichever way the day is read” only when all three put the same group behind, by at least a day. A comparison is made only if the images around the day are at most 16 days apart. For the forests taken together the three readings differ by at most 1.5 days; for small places, by far more.

Other methods. The curve fitted to each square gives, for each year, a day that differs by at most 2.7 days from the one read directly between images. Computed by other methods (a curve on the forest median, interpolation between images), the forests’ usual days come out one or two days earlier, and the district trees come out, on average, three to four days behind the forests (3.2 days from the fitted curve, 3.1 from the year-by-year reading).

How much of the drop can come from the lower sun. In the same images, the conifer stands mapped in OpenStreetMap lose less than a tenth of NDVI between September and the end of October (0.03–0.07, depending on how their edges, which include broadleaved trees, are drawn), while the broadleaved forest drops from 0.86 to 0.56. Conifers keep their leaves, so their drop shows roughly how far autumn light and air can move the index; the comparison is not an exact measure of that effect on broadleaved trees.

Limits of the measurement. A 10-metre square in a city district holds tree, grass and roof; its day belongs to the mixture. A place is listed if it has at least 120 measured squares of trees, so pocket parks and small public gardens are missing. Each forest’s dates for individual years are estimates: its usual day, moved by that year’s lead or lag across all the forests. A date on the map is the mean of eight autumns and is uncertain by a few days: a day or two between two places means nothing. Each place’s card shows how far ahead of or behind the forests it was in the autumns in which its day can be read directly; for small places and for districts the range is wide. Neighbouring squares resemble one another and are measured on the same days, so the nearly 910,000 squares are not as many independent measurements; that is why, for places, the page gives ranges between years, not statistical margins. Eight autumns are not enough to say whether autumn now comes later than it used to: Landsat has images from 1984, but in years with one image every 16 days it cannot pin down the day.

The images. All photographs are rendered with the same contrast curve and the same saturation, stronger than the eye would see from orbit, so that a difference in colour between two dates means a difference on the ground.

Sources#

Contains modified Copernicus Sentinel data (2017–2026).

• The images: Copernicus Sentinel-2, level 2A (surface reflectance), European Space Agency, at dataspace.copernicus.eu
• Read from the open Earth Search archive, from which the 263 images were chosen (2 August 2017 – 12 October 2026), at earth-search.aws.element84.com
• The second satellite: Landsat 4–9, Collection 2 Level-2, U.S. Geological Survey, 520 images from 1984–2026 in which at least half of the forest is visible, at usgs.gov
• Where the trees are: ESA WorldCover 2021, 10 m (CC BY 4.0), at esa-worldcover.org
• Outlines and names of forests, parks and districts: © OpenStreetMap contributors (ODbL), at openstreetmap.org
• Altitude: Copernicus DEM GLO-30, at dataspace.copernicus.eu
• Copernicus also publishes a European map of the end of the growing season at 10 m from the same images (“End-of-season Date”). There the end of season is the day on which another index (PPI) falls to 15% of the seasonal amplitude, later than the half-way day here. The days on this page have not been compared with it: land.copernicus.eu
• Satellite indices compared with ground observations of leaf colouring and leaf fall (MODIS images, deciduous forests in the United States): Wang et al., “Divergence in Autumn Phenology Extracted From Different Satellite Proxies Reveals the Timetable of Leaf Senescence Over Deciduous Forests”, Geophysical Research Letters, 2024, at doi.org
• Tree species against planting site, in Washington, DC: Alonzo et al., “Canopy composition drives variability in urban growing season length more than the heat island effect”, Science of the Total Environment, 2023, at doi.org
• Light at night and leaves that last longer, in 452 cities: Chen et al., “Increased artificial illumination delays urban autumnal foliar senescence”, Nature Communications, 2026, at doi.org

The same city in winter: Cluj under the lid, on the nights when cold air is trapped over the city.