Methodology

How these numbers are made

Every figure in this app is an area measured from Landsat imagery, not a mass and not a published result. This page says how a year is chosen, composited and measured, and the four ways the resulting number can mislead you.

What is measured

Ice is separated from rock and water with the Normalised Difference Snow Index, NDSI = (Green − SWIR1) / (Green + SWIR1), thresholded per site. Snow and ice are bright in the visible and dark in the shortwave infrared; almost nothing else on a glacier landscape is.

That gives clean-ice area — and the adjective is load-bearing. Debris-covered ice is spectrally indistinguishable from bedrock, so NDSI does not see it. Spectral-mixing work in this project put the limit at roughly 40% debris fraction for reliable detection, with detection lost beyond about 50%: above that, real ice reports as zero area.

Ice sheets are different enough to need a different headline. Greenland reports bare-ice extent, never total area: the total changes by a fraction of a percent per decade, so a headline built on it would read as “Greenland is fine”, which is the opposite of the truth.

When each year is imaged

The target is the minimum-seasonal-snow date — the end of the melt season, when what remains is glacier ice rather than last winter's snow. That date is not the same everywhere, so it is set per site, not globally:

Sites	Anchor	Window
Northern hemisphere (18 of 20)	15 August	± 45 d
Southern Patagonia	1 March	± 45 d
Antarctica	20 January	− 60 / + 40 d

On the two southern sites, year Y means the austral summer that ends in Y — Patagonia “2026” is January–April 2026. Without that rule, comparing a northern and a southern site compares two different things.

How a year is composited

Every usable scene in the window is stacked and reduced to a per-pixel median, not a mean. A median rejects surviving bright cloud as an outlier; a mean averages it in.

Cloud masking is where this most easily goes wrong, because the standard cloud masks have a high false-positive rate over bright snow — they flag the glacier as cloud. Masking naively deletes the subject. So:

The snow/ice flag is never masked. It is a keep class, not a reject class.
Only high-confidence cloud is rejected, along with cloud shadow and cirrus. Medium-confidence cloud is allowed into the stack, where the median deals with it.
The number of valid observations is tracked per pixel and carried through as data, not discarded.

Four ways this can mislead you

1. Observation depth, not ice

This is the most serious one, and it is measured rather than hypothesised. A year built from more observations measures less area, because a deeper stack lets the median reject seasonal snow that a thin stack cannot. Observation depth also rises over time as satellites accumulate. The two together manufacture a downward trend out of Landsat's launch history.

At Austfonna the correlation between observation depth and measured area is −0.79, depth against year is +0.86, and the error against the published reference area falls monotonically from +62% at 2.3 observations per pixel to +2.7% at 7.8. Most of the apparent 24-year decline there is artifact.

Filtering to good years alone does not fix it — the correlation holds at −0.75 among those. Where the app detects this pattern it says so on the chart rather than drawing a trend line.

2. Terrain shadow

Low sun casts deep shadow, and NDSI reads shadow as not-snow. Every Arctic site is exposed to this, and the steep ones — Karakoram, St. Elias, Stikine — most of all. The design calls for a DEM-based illumination correction so shadowed pixels are excluded rather than counted as ice loss. It is not applied in this build.

3. Debris cover

Where a glacier is mantled in rock, the measurement stops seeing ice that is still there. This is severe at Karakoram and significant at St. Elias, Chugach and Stikine. A shrinking clean-ice area at these sites can mean advancing debris cover rather than retreating ice.

4. A flat line is a result

Karakoram is documented in the published literature as near-balance or slightly advancing since the 1990s — the Karakoram anomaly. When this app shows a flat line there, that is the finding, not a broken chart.

Quality ratings

Every year on every timeline carries a rating, encoded by pattern as well as colour. A year is refused publication rather than quietly shipped when its AOI-mean observation count falls below the site's threshold (3 by default).

good: Enough clear observations for the median to do its job.
degraded: Thin but usable. Treat the number as weak evidence — this is where the observation-depth bias above bites hardest.
interpolated: Below the observation threshold; not a measurement.
missing: No usable imagery. Shown as a gap, never filled in.

Where the imagery comes from

Era	Sensor	Res	What to know
1982–1993	Landsat 4 TM	30 m	Sparse; high-Arctic coverage very thin.
1984–2012	Landsat 5 TM	30 m	Primary source 1986–1998. Before 1999, high-latitude acquisition was limited by ground-station reach and no onboard recorder.
1999–2022	Landsat 7 ETM+	30 m	The 2003 scan-line corrector failure costs ~22% of each scene in stripes. Usable only in medians of several scenes at differing stripe phase.
2013–	Landsat 8 OLI	30 m	Best radiometry of the series.
2021–	Landsat 9 OLI-2	30 m	With Landsat 8, an effective 8-day revisit.

Only Tier 1 scenes are used for any published number. Tier 1 carries a measured geolocation error (median 4.4–7.8 m across these sites, sub-pixel); Tier 2 carries none at all, at any site. The difference is quantified against unquantified error, which is why Tier 2 is permitted for visual context only and never for a metric.

Two consequences worth stating plainly. Four sites — Grant, Agassiz, Prince of Wales and Mueller — have no usable imagery before 1999, so their baseline is 1999 and is labelled as such; there is no extrapolation back to 1986. And Landsat does not image above 82.7° N, so the sliver of Greenland beyond it is excluded from every area total rather than counted as ice loss.

The picture is not the measurement

The true colour view is a presentation frame: a picture assembled so that a bad year is legible, which is a different job from measuring one. It is allowed to do things the measurement is not. Where Tier 1 saw nothing it falls back to Tier 2, then to a median of the five years either side. It removes atmospheric haze separately for each of those sources, because they do not share an atmosphere. It levels the sea against a fixed reference for the site rather than letting a hazy year render its ocean pale grey. Small holes are diffused shut.

None of that reaches a number. Every area, uncertainty and ice mask on this site is measured from Tier 1 imagery before any of it happens — and the false colour view is that imagery, with nothing done to it. If you want to see what the measurement saw, that is the button.

A frame that borrows does not look like it borrows, so each one says so instead. The note under the True colour button gives that year's composition: how much of the picture Landsat actually saw in the year on the label, how much is borrowed from the years either side, and how much was never observed by anything and is filled. Where a year could not be built without inventing more than 15% of it, no frame exists and that year still shows the plain render. In practice those are years the satellites barely covered at all, which render close to black either way.

Why every site uses a different map

Web Mercator — the projection behind almost every online map — is unusable here. Its scale distortion grows without bound toward the poles, and it cannot represent the poles at all. Fifteen of these twenty sites are Arctic. A Mercator map of Austfonna is stretched by a factor of five, and any area measured on it is wrong by that factor.

So no site is measured, stored or drawn in Web Mercator. Each site is stored, tiled and drawn in a projection chosen for its own latitude, stated in the footer of every map. Greenland is drawn in EPSG:3413, Antarctica in EPSG:3031, and most other sites in a Lambert azimuthal equal-area projection centred on the site itself. Areas are computed per pixel on the ellipsoid rather than read off the projected grid.

The one Mercator input is the background. The imagery around each picture is Esri World Imagery, which is published only as Web Mercator tiles. Each tile is reprojected into the site’s own projection before it is drawn, so everything on the screen is in that projection; nothing is measured on the background, and it is not one of the years. It is there so the picture sits in its landscape rather than in a black box.

The visible cost is that two sites cannot be laid side by side on one grid, and that north is only exactly up at the centre of each map. Both are honest consequences of measuring area on a sphere. The graticule, scale bar and north arrow on each map are orientation aids for this reason, not decoration — in a polar projection a reader has no Mercator intuition to fall back on.

What this build does not do yet

This is a pilot. 18 of 20 missions have imagery, and it is built to demonstrate the architecture on real data — not to publish a measurement.

Site	Grid	Years	Rated good	Obs/px
Grant Ice Cap	60 m	28	24	0.0–12.5
Wrangell–St. Elias	300 m	41	30	0.7–6.8
Agassiz Ice Cap	180 m	28	26	0.5–16.1
Severny Island	240 m	41	25	0.3–14.8
Prince of Wales Icefield	180 m	30	22	0.0–14.3
Southern Patagonia	300 m	30	20	0.9–7.2
Devon Ice Cap	120 m	41	39	1.5–11.5
Karakoram	300 m	39	23	0.0–6.4
Chugach	120 m	40	26	0.0–11.0
Spitsbergen	300 m	40	18	0.2–9.4
Vatnajökull	120 m	41	25	0.9–6.8
Austfonna	90 m	40	21	0.1–7.8
Mueller Ice Cap	60 m	27	23	0.8–10.0
Stikine Icefield	180 m	41	24	0.5–9.9
Penny Ice Cap	120 m	41	37	1.2–7.4
Barnes Ice Cap	120 m	41	40	2.6–10.2
Academy of Sciences	60 m	40	26	0.3–12.5
Juneau Icefield	120 m	41	23	1.0–7.8

The pipeline records these limits against the sites it has built:

Areas are clipped to the GLIMS via Ice-Fields project glacier outline, so they measure this glacier rather than whatever ice falls inside the AOI box. The clip is also a ceiling: a measured area can never exceed the outline it was cut with, so agreement with the published extent is not independent confirmation.
Areas are measured at 30 m; the imagery is shown at 90 m. The picture is coarser than the measurement. — Austfonna only
Areas are measured at 60 m; the imagery is shown at 120 m. The picture is coarser than the measurement. — Devon Ice Cap, Chugach, Vatnajökull, Penny Ice Cap, Barnes Ice Cap, Juneau Icefield only
Areas are measured at 60 m; the imagery is shown at 180 m. The picture is coarser than the measurement. — Agassiz Ice Cap, Prince of Wales Icefield, Stikine Icefield only
Areas are measured at 60 m; the imagery is shown at 240 m. The picture is coarser than the measurement. — Severny Island only
Areas are measured at 60 m; the imagery is shown at 300 m. The picture is coarser than the measurement. — Wrangell–St. Elias, Southern Patagonia, Karakoram, Spitsbergen only
Clean-ice area only: debris-covered ice is spectrally invisible.
No illumination correction applied yet.
Pilot resolution 60 m, not the 30 m production grid. — Grant Ice Cap, Mueller Ice Cap, Academy of Sciences only

The most consequential of those is the bounding box, and it can now be given a size. Glacier outlines for Svalbard put Austfonna itself at 8,067 km² in 2010, while every glacier inside its bounding box comes to 9,246 km² — so about 1,200 km² of the figure is neighbouring ice. At Spitsbergen the gap is far worse: 8,362 km² for the named icefield against 29,317 km² inside the box, which is three and a half times as much. On top of that sits whatever seasonal snow the year happened to leave behind. Treat every number here as an architecture demonstration on real imagery, not as a measurement.

Attribution

Landsat Collection 2 surface reflectance — USGS / NASA. Courtesy of the U.S. Geological Survey.
Coastlines — Natural Earth, public domain, via world-atlas.
Glacier outlines and inventory — RGI / GLIMS, used for AOI seeding and classifier validation.
The twenty missions and the expedition behind them — Ice Legacy.