What does data fusion mean in geological remote sensing?

It is combining datasets with different strengths — optical spectral bands, a DEM, geophysics — into co-registered layers so each compensates for the others' limits. Typical examples are pan-sharpening, stacking ASTER SWIR ratios over a Sentinel-2 base, and draping spectral indices on hillshade.

Which ASTER and Sentinel-2 band ratios map alteration minerals?

ASTER band 4/6 highlights argillic/phyllic clays via the 2.2 um Al-OH absorption, and 5/8 targets advanced argillic alteration. With Sentinel-2, band 11/12 (SWIR) approximates clay content and 4/8 (red/NIR) is the basis for vegetation masking. Validate ratios against field samples before interpreting.

How do I co-register Sentinel-2, ASTER, and a DEM before fusing them?

Reproject everything to one projected CRS (a UTM zone) with gdalwarp, resample to a common grid and pixel size, and check sub-pixel alignment against ground control points. Misregistration of even one pixel creates false edges that look like geological contacts.

Should I pan-sharpen Landsat before interpretation?

Pan-sharpening (e.g. Landsat band 8 at 15 m) sharpens visual interpretation of structure and texture, but it alters pixel spectra. Do spectral analysis like band ratios on the original multispectral bands, and use the pan-sharpened product for delineation and presentation only.

Data Fusion for Geological Interpretation

Short answer

Data fusion in geological interpretation means combining datasets with complementary strengths — Sentinel-2 for spatial detail and frequent revisit, ASTER for shortwave-infrared (SWIR) mineral discrimination, Landsat for a multi-decade archive, and a DEM for structure — into co-registered layers where each compensates for the others' weaknesses. Done well, it turns scattered scenes into an evidence stack that narrows field targets. Done carelessly, sub-pixel misregistration and uncorrected spectra manufacture "contacts" that do not exist.

This guide covers the corrections that must precede fusion, the specific band ratios that map alteration, the fusion techniques (pan-sharpening, PCA, band stacking, DEM draping), and how to validate before anyone treats the result as geology.

Know what each sensor can and cannot see

Fusion only helps if you respect each sensor's physics.

Sentinel-2 (Copernicus). 13 bands; 10 m in visible/NIR (bands 2,3,4,8), 20 m in red-edge and SWIR (bands 5,6,7,8a,11,12), 5-day revisit with two satellites. The 20 m SWIR bands 11 (~1.6 um) and 12 (~2.2 um) carry useful clay and carbonate information, though far coarser spectrally than ASTER.
Landsat 8/9 (OLI). 30 m multispectral, 15 m panchromatic, ~16-day revisit, and crucially a continuous archive back to the 1980s (TM/ETM+). Its value is temporal: baselines, mine-footprint change, vegetation trends.
ASTER. Historically the workhorse for alteration mapping because of three SWIR bands across the 2.0–2.4 um region where Al-OH, Mg-OH, and carbonate absorptions sit, plus five thermal-infrared bands for silica. Note the SWIR detectors failed in 2008, so archive scenes are the usable SWIR source.

No optical sensor sees lithology directly. It sees reflected and emitted radiation modulated by surface mineralogy, vegetation, soil, moisture, and illumination. Fusion manages those confounders; it does not remove the need for field truth.

Corrections before fusion

Fusion of poorly prepared inputs produces confident nonsense. Three steps are mandatory.

Atmospheric correction to surface reflectance

Work in surface reflectance, not top-of-atmosphere radiance, especially when comparing dates or sensors. Use Sentinel-2 L2A products (Sen2Cor-processed) and Landsat Collection 2 Level-2 surface reflectance. Band ratios computed on TOA data drift with haze and sun angle.

Co-registration to a common grid

Reproject every input to one projected CRS — a UTM zone for the area, e.g. EPSG:32633 for UTM zone 33N — and resample to a shared pixel grid.

# bring ASTER SWIR to the Sentinel-2 grid: same CRS, 20 m, aligned origin
gdalwarp -t_srs EPSG:32633 -tr 20 20 -tap \
  -r bilinear aster_swir.tif aster_swir_20m.tif

Use -tap (target aligned pixels) so the two rasters share grid origin; otherwise stacked layers are offset by a fraction of a pixel. Verify alignment against ground control points or a road network — misregistration of even one pixel along a slope generates a linear artefact that reads like a fault.

Masking confounders

Mask clouds, shadow, water, and dense vegetation before interpreting mineral signals. A quick vegetation mask from Sentinel-2: NDVI = (B8 − B4) / (B8 + B4); exclude pixels above ~0.3 where canopy hides the substrate. Cloud and shadow come from the Sentinel-2 scene classification layer (SCL) in L2A.

Fusion techniques that earn their place

Band ratios for mineral indicators

Ratios suppress topographic shading (the slope term largely cancels) and isolate absorption features.

ASTER 4/6 highlights Al-OH clays (argillic/phyllic alteration) via the 2.2 um absorption.
ASTER 5/8 targets advanced argillic alteration (alunite, kaolinite).
ASTER 4/8 or (5+7)/6 is used for muscovite/illite groups.
Sentinel-2 11/12 approximates clay abundance at coarser spectral resolution; 2/11 is a ferric-iron style proxy.

Compute these in QGIS Raster Calculator or with gdal_calc.py:

gdal_calc.py -A aster_b4.tif -B aster_b6.tif \
  --calc="A.astype(float)/B" --outfile clay_ratio.tif --NoDataValue=0

Treat ratios as candidate maps. Anomalies cluster where alteration or where soils, lichen, or shade mimic the spectrum — that is why validation is non-negotiable.

Pan-sharpening for delineation, not spectra

Pan-sharpening fuses a high-resolution panchromatic band with lower-resolution multispectral bands (Landsat 30 m → 15 m using band 8). It sharpens structural and textural interpretation but alters per-pixel spectra. Rule: do spectral analysis (ratios, classification) on the original multispectral pixels; use the pan-sharpened product only for visual delineation and figures.

gdal_pansharpen.py landsat_b8.tif landsat_mss.tif pansharp.tif

Principal component analysis

PCA on a multiband stack concentrates variance into a few components and can isolate mineralogical contrast that is buried across many correlated bands (the Crósta directed-PCA technique is the classic application to iron-oxide and hydroxyl mapping). Inspect eigenvector loadings to learn which component carries the absorption of interest before interpreting it.

DEM draping and structural fusion

Drape spectral indices over hillshade derived from a DEM (Copernicus GLO-30 or SRTM 30 m) to read mineralogy in structural context. Compute hillshade with gdaldem hillshade dem.tif hillshade.tif -z 1.5, then overlay the ratio layer at partial opacity. Lineaments that align with ridge crests and drainage often correspond to faults and joints, fusing terrain morphology with the spectral signal.

Worked example: phyllic alteration screening

Acquire Sentinel-2 L2A (10/20 m) and an archive ASTER L2 SWIR scene over the prospect.
gdalwarp both to EPSG:32633, 20 m, -tap aligned; co-register to GCPs.
NDVI mask from Sentinel-2 B8/B4 to drop vegetation > 0.3; apply SCL cloud/shadow mask.
Compute ASTER 4/6 (Al-OH) and 5/8 (advanced argillic); compute Sentinel-2 11/12 as an independent clay check.
Stack the ratios as RGB, drape over Copernicus GLO-30 hillshade.
Rank anomalies, generate field targets, and label every layer with its confidence and the masks applied.

Common pitfalls and the reasons behind them

Reading false colour as lithology. RGB band combinations are interpretive aids; soils and shade produce the same colours as target minerals. The cause is treating a visualization as a classification.
Comparing dates without correcting illumination. Different sun angle, season, and moisture change reflectance even where geology is identical, so apparent "change" is atmospheric, not geological. Use surface-reflectance products and similar acquisition geometry.
Stacking misregistered rasters. A one-pixel offset on a slope produces linear edges that mimic contacts. Always -tap align and verify against GCPs.
Ratios on TOA data. Haze biases numerator and denominator unequally, shifting anomaly thresholds. Correct to surface reflectance first.
Using ASTER SWIR scenes after 2008 without noticing the detector failure, yielding empty or interpolated bands. Check acquisition date.

Validation

Compare anomalies against any existing geological map and known mineral occurrences before trusting them.
Field-check a sample of high-ranking pixels with a portable spectrometer or hand sample; spectral mapping without ground truth is a hypothesis, not a result.
Cross-check the same mineral signal in two independent sensors (ASTER and Sentinel-2) — agreement raises confidence, disagreement flags an artefact.
State the masks, corrections, and CRS in the metadata so a reviewer can reproduce the stack.

Bathyl perspective

We use fusion to expand what a geologist can see, not to replace the field. The strongest deliverable is explicit about what each sensor contributed, what was masked, and which anomalies are still hypotheses awaiting ground truth. That honesty is what lets a project manager weight the evidence correctly.