How this imagery was made

Source

Every pixel is a Landsat observation from the USGS archive, Collection 2, read through Microsoft’s Planetary Computer. Nothing is painted in, interpolated between dates, or generated. Three instruments cover the record: MSS (1972–1984, 79 m), TM (1985–2011, 30 m) and OLI (2013–2026, 30 m).

One day per year — actually one day

Forty-eight of these fifty frames are a single afternoon. Not a blend of a season, not a median of six passes: one satellite overpass, every pixel from the same minute.

That is possible because of where this rectangle sits rather than how it was processed. A Landsat path images a strip about 185 km wide, and the area was chosen to fall wholly inside one strip — path 38 of the modern grid, path 41 of the one Landsat 1–3 flew. A rectangle that hangs over the edge of a strip can never be a single-day image, however it is assembled, because the satellite was not looking there that day.

The exceptions are 1996 (97.3% from its one pass) and 1997 (94.8%). The remainder in those two comes from other passes in the same season, and both frames say so on screen.

Which day

For each year the clearest single pass of the season is used, measured on the imagery itself rather than on the catalogue’s cloud estimate. In practice that lands between 12 June and 15 October: fourteen frames each in July, August and September, three in June and five in October.

That spread is worth knowing when you watch the sequence. A year anchored in June or October carries snow on the high ground that a July frame does not, and the mountains will appear to gain and lose snow from one frame to the next. That is the calendar moving, not the climate — each frame names its own date.

Cloud was not allowed to drive the choice past the point of usefulness. Where a summer had no clear day, the least-clouded one is shown with its weather intact: a real cloud on a real afternoon is more honest than a patch stitched in from a different month.

Years that are not here

Five summers are missing, and each for a reason the archive imposed rather than a choice of convenience:

• 1977 — every acquisition that summer is completely clouded over this valley. There is no usable day at all.
• 1975, 1976 and 1982 — the imagery exists but carries defects in the multispectral scanner that cannot be removed: detector striping across the frame in 1976, and heavy cloud in the other two. The least-striped alternative days in those summers measure no better.
• 2012 — Landsat 7’s scan-line corrector had failed, leaving wedges of the frame with no data at all, and 2012 is the one year with no second satellite in orbit to supply them.

They are left out rather than included badly. A missing year means the archive could not support one.

Color

Brightness is not adjusted per frame. Every year passes through one fixed curve per instrument, anchored to surface reflectance, so identical ground renders identically in 1978 and 2026 — which is what makes a change in the picture mean a change on the ground.

1972–1984 are false color. The MSS instrument carried no blue band; it simply never recorded the data, so true color is impossible for those years. They use near-infrared, red and green, and healthy vegetation appears red. We tested reconstructing a blue band from same-day MSS and TM pairs; it was not accurate enough to publish, so the honest rendering is the one you see.

Which satellite took which year

The controls show only the year, because the instrument was never a choice — one Landsat was flying at a time. This is the record behind them, built from the frames actually in this viewer:

How a satellite actually sees

Landsat is not a camera in the ordinary sense. It does not capture a picture; it measures how much light comes back from the ground in a few specific slices of the spectrum, one slice at a time, and the picture is assembled afterwards from those measurements.

Some of those slices are the ones your eye uses — red, green, blue. Others are invisible. The most useful of them is near-infrared, just past the red end of human vision. Healthy leaves reflect near-infrared very strongly, far more than they reflect green, because the internal structure of leaf tissue scatters it. That single fact is why satellites can tell a thriving forest from a dying one, irrigated hay from dry range, and water from land — water absorbs near-infrared almost completely and comes out black.

It is also why the 1970s frames here are red. Those years are drawn with near-infrared standing in for red, so vigorous vegetation — the thing reflecting most strongly in that slice — appears brightest and reddest.

Fifty years of getting better

Each generation of Landsat measures more slices, in finer detail, with more precision:

• MSS (1972–1984) — 4 slices, 79 m ground detail. No blue at all, and only 64 brightness steps per slice.
• TM (1984–2012) — 7 slices including blue and two shortwave-infrared, 30 m, 256 brightness steps. Shortwave infrared was the leap: it separates snow from cloud, and wet ground from dry, neither of which the earlier bands could do.
• ETM+ (1999–) — adds a sharper 15 m black-and-white band. It is not used here: its scan-line corrector failed in 2003, and the one year that would have needed it (2012) is missing for that reason.
• OLI (2013–) — 9 slices, 30 m, and around 4,000 brightness steps instead of 256. That precision is why modern frames hold detail in dark forest and bright snow at the same time, where the older sensors had to give up one or the other.

Detail stops at the sensor — about 30 m since 1985, about 79 m before. Zooming further magnifies pixels; it does not reveal more.

What comes next: hyperspectral

Every sensor above splits the spectrum into a handful of wide slices. A hyperspectral instrument splits it into hundreds of very narrow ones, so each pixel carries something close to a complete spectral curve rather than a few numbers.

That matters because materials have signatures. A wide band can say “this is vegetation”. A few hundred narrow bands can say which species, whether it is water stressed, what mineral a rock face is, and — because gases absorb at very specific wavelengths — whether methane is leaking from a wellhead.

Planet’s Tanager satellites, the first launched in 2024, do exactly this: roughly 400 contiguous bands from visible light through shortwave infrared, at about 30 m. The trade is the one it has always been — splitting light that finely means less of it per band, so hyperspectral instruments buy spectral richness with coverage and signal. Landsat still wins on the thing this viewer depends on: an uninterrupted, free, consistently calibrated record going back to 1972. Nothing else has that, and nothing launched today can create it retroactively.

Landsat imagery courtesy of the U.S. Geological Survey.