ASTER, Sentinel-2, PRISMA or EnMAP: a practical decision table for alteration mapping

Sensor choice for alteration mapping is decided by four things and almost never by the one people argue about. The four are: how finely the sensor samples the 2.0–2.5 micrometre region, whether the instrument still works, whether you can get a scene over your licence area at all, and what you are legally allowed to do with the result. Spatial resolution — the specification everyone quotes — is the least discriminating of the lot, because ASTER, PRISMA and EnMAP all land within a factor of two of each other in the SWIR.

This is the decision table, followed by the reasoning behind each column.

The decision table

Sources: ASTER characteristics, NASA JPL; ASTER SWIR anomaly, NASA Earthdata; Sentinel-2 mission, Copernicus; PRISMA, eoPortal; EnMAP mission; EMIT, eoPortal.

What “enough bands” means at 2.2 micrometres

The 2.0–2.5 micrometre region is where alteration mapping lives, and the number of bands a sensor places in it determines what question you can ask. Aluminium-hydroxide bonds absorb near 2.20 µm, which is the signature of muscovite, illite and sericite — phyllic alteration — and of kaolinite. Alunite absorbs near 2.17 µm. Iron-magnesium-hydroxide and carbonate features sit near 2.33–2.35 µm, marking chlorite, epidote and calcite, which is the propylitic halo.

ASTER samples that region with five bands roughly 40 nanometres wide. That is enough to answer is there an Al-OH absorption here, which is why the classic ASTER band ratios work: band 4 over band 6 and band 5 over band 6 respond to the phyllic signature, band 4 over band 5 separates alunite from kaolinite, and band combinations around bands 7, 8 and 9 pick up carbonate. Those ratios have found alteration zones for twenty years and continue to.

What five 40-nanometre bands cannot do is tell you which Al-OH mineral, or how the mineral chemistry varies. A hyperspectral sensor sampling the same region every 8 to 12 nanometres resolves the shape and exact position of the absorption minimum, and the position shifts systematically with composition — white mica moves toward longer wavelengths as it becomes more phengitic. That shift is a vector, and it points toward the hydrothermal centre. This is the actual reason to pay for hyperspectral: not a better map of the same thing, but a different measurement.

ASTER’s SWIR has been unusable since April 2008

If you are planning to acquire new ASTER SWIR data, stop — there is none. On 23 April 2008 the ASTER SWIR detector temperature rose sharply and bands 5 through 9 saturated, most likely because the detector decoupled from its cold finger. A recycling procedure on 7 May 2008 failed to bring the temperature down, all subsequent recovery attempts failed, and the subsystem was switched off entirely in August 2012. NASA’s own advisory is unambiguous: ASTER SWIR data acquired since April 2008 are not usable and show saturation and severe striping.

The consequences for planning are specific:

• ASTER SWIR is an archive, covering roughly 1999 to early 2008. If your area was imaged cloud-free in that window, the data are excellent and free. If it was not, no amount of budget changes that.
• ASTER VNIR (15 m) and TIR (90 m) continue to operate. The TIR subsystem is genuinely useful and under-exploited — it maps silica content, which no VNIR-SWIR instrument does, and silicification is an alteration product in its own right.
• Any tutorial, textbook chapter or vendor deck that recommends ASTER for SWIR alteration mapping without dating the imagery is describing a workflow that has been partly unavailable for eighteen years.

ASTER has no blue band, and for some projects that decides it outright

ASTER’s shortest-wavelength channel, band 1, begins at 0.52 µm — in the green. There is no blue band. That single gap has two consequences that are easy to miss when comparing specification sheets.

The first is spectral: the standard iron-oxide discrimination ratios that separate goethite, hematite and jarosite rely on the blue-to-red slope in the visible, and ASTER cannot compute them. Iron-oxide mapping with ASTER is therefore a weaker exercise than with a sensor carrying a blue channel, which matters because gossans and supergene iron caps are frequently the surface expression you are looking for.

The second is more mundane and, in one of our projects, decisive. ASTER cannot produce a true-colour composite. When ZetaMine built a labelled imagery corpus for alteration-related classification, the target was a natural-colour RGB image that a human annotator — or a vision model — could look at and reason about. ASTER was rejected on that basis alone. The corpus was built instead from 10,348 Landsat-7 ETM+ GLS2005 scenes, composited as bands 30/20/10 to red/green/blue for natural colour. Of those, 491 scenes were rejected for exceeding 50% no-data and 125 more as edge slivers, leaving 9,732 images totalling 741 MiB, every one content-hash unique.

The lesson generalises. The sensor decision is driven by the downstream consumer of the image, and a human eye or a vision model is a consumer with a hard requirement for blue. Mineral-mapping suitability and annotation suitability are different criteria, and a sensor can be excellent at one and disqualified from the other.

Sentinel-2 is a change detector that happens to have a SWIR band

Sentinel-2 has exactly one band in the 2.0–2.5 µm region: B12, centred near 2190 nm, at 20 m resolution and roughly 180 nm wide. One broad band cannot discriminate alteration mineralogy — it can only tell you that something in that region is absorbing.

That is not a criticism, because Sentinel-2 is not competing for the same job. Its 5-day revisit from the two-satellite constellation, 290 km swath, 10 m visible bands, free and open licence and complete global archive make it the correct choice for everything around the alteration map: base mapping, vegetation masking, snow and cloud screening, seasonal change, and monitoring disturbance at a licence area over time. Use it for the questions that need repetition, and use a hyperspectral sensor for the question that needs spectra.

The one genuinely useful mineralogical role for Sentinel-2 is as a regional filter. A coarse B11/B12 ratio across a whole belt costs nothing and narrows down where to spend a PRISMA or EnMAP tasking request — and tasking requests are the scarce resource, not the imagery.

PRISMA and EnMAP have the same resolution and different constraints

PRISMA and EnMAP both deliver 30 m hyperspectral data over a 30 km swath, and on spectral grounds they are close substitutes. The differences that will actually decide your project are elsewhere.

Read the PRISMA licence row twice if the output is going to a client. ASI provides the data free to anyone who registers, with restrictions on redistributing products to third parties and on commercial use. A derived alteration map delivered to a paying client is exactly the case that needs checking before the work starts, not after.

EnMAP’s off-nadir pointing is the specification that changes project plans. A 27-day nadir revisit means a cloud-affected acquisition costs you a month; the ability to acquire at 30° off nadir in four days turns that into a manageable delay. The cost is geometric — off-nadir acquisition degrades the pixel geometry and complicates atmospheric correction over terrain — but a usable scene with a harder correction beats a perfect scene next month.

Published band counts vary slightly by product level and source: the EnMAP mission page states 224 contiguous bands, while some published analyses of Level-2A products report about 246 channels. Check the product specification for the level you are ordering rather than the mission summary.

A second table: choose by question, not by sensor

Overall classification accuracy is not exploration success

Published alteration-mapping studies report overall accuracy and kappa, and those numbers do not mean what a non-specialist reader assumes. They measure agreement between a classified map and a reference map — usually derived from known geology, field spectra or an existing deposit’s documented alteration halo. That is a hindsight test. It tells you the sensor and algorithm can reproduce something already known. It does not tell you they would have found it.

Two published patterns make the point concretely. First, the algorithm matters as much as the sensor: a 2024 study in the European Journal of Remote Sensing mapping alteration zones over a porphyry copper system with PRISMA data reported overall accuracies of 71.9%, 87.5%, 81.3% and 71.9% — a spread of more than fifteen percentage points — for selective principal component analysis, band ratios, matched-tuned matched filtering and linear spectral unmixing applied to the same scene. A headline accuracy figure without the algorithm and validation design attached carries almost no information.

Second, the strong published results cluster at exceptionally well-characterised sites. Cuprite and Goldfield have served as spectroscopy calibration sites since the 1980s, which is why EnMAP was evaluated there, and why PRISMA validation work concentrates on documented districts such as Aktogay and the Chilean Andes. These are the right places to characterise a sensor and the wrong places to estimate discovery rates.

The honest framing for a client is this: hyperspectral alteration mapping produces a defensible ranking of targets, reduces the ground area a field team has to walk, and provides mineralogical evidence for why a target is ranked where it is. It does not find deposits. Alteration is not ore, mapped alteration is surface alteration, and a blind zone under fifty metres of cover is invisible to every sensor in the table above.

Four things that will degrade your map regardless of sensor

Vegetation. Even sparse canopy dominates the VNIR-SWIR signal. Green vegetation cover above roughly 30% makes mineral mapping unreliable, and the standard responses — vegetation-corrected indices, spectral unmixing — reduce the problem rather than removing it.

Atmospheric correction. Hyperspectral mineral mapping depends entirely on surface-reflectance accuracy in narrow bands. A correction artefact at 2.2 µm is indistinguishable from an Al-OH absorption. Use the mission’s own Level-2 product where one exists, and compare against field spectra when you can.

Mixed pixels. A 30 m pixel over an alteration halo contains altered rock, unaltered rock, soil, shadow and often vegetation. Sub-pixel methods estimate abundance; they do not resolve the mixture.

Look-alike minerals. Gypsum, ammonium-bearing minerals and some evaporites produce features near the diagnostic alteration wavelengths. Playa margins and agricultural areas generate confident false positives, and no accuracy statistic computed against a reference map in a mineralised district will warn you about them.

Where this stops applying

This table is built for hydrothermal alteration mapping over exposed, arid to semi-arid terrain from spaceborne sensors. Outside that envelope the ranking changes.

Under significant vegetation or in the humid tropics, the entire VNIR-SWIR approach weakens, and radar, geophysics and geochemistry carry more of the load than any optical sensor.

At outcrop and drill-core scale, spaceborne 30 m data is the wrong tool by two orders of magnitude — that is a field spectrometer and hyperspectral core-scanner problem.

For commodities without a distinctive surface alteration expression, or for targets under transported cover, the whole method is inapplicable and no sensor choice rescues it.

And every specification in the table is a mission specification as published at the time of writing, 27 August 2026. Instruments degrade — ASTER’s SWIR is the definitive example — missions end, and data policies change. Verify the current status on the operator’s own page before committing a project to a sensor, because the most expensive way to discover an instrument failure is in the middle of an acquisition plan.