Appearance
Multispectral imaging of a manuscript means photographing the same page many times, each time under a different narrow band of light from ultraviolet (around 365 nm) through visible to near-infrared (around 940 nm). Because inks, parchment and pigments reflect and absorb each wavelength differently, comparing the resulting frames exposes faded text, scraped-off (palimpsest) writing and material damage that white light hides. You do not need a laboratory: a monochrome camera, a set of narrowband LEDs and free software get you a usable result.
What problem does multispectral imaging actually solve?
White light averages everything your eye sees into three broad channels. A faded iron-gall annotation and the parchment under it may reflect nearly identically in visible light but diverge sharply at 365 nm or in the near-infrared. By isolating the band where the contrast is largest, you recover legibility that no amount of contrast-stretching a single RGB photo can produce. This is why libraries image the Archimedes Palimpsest, charred Herculaneum scrolls and water-damaged charters this way.
How does the capture chain fit together?
Think of five linked stages, each of which can ruin the others if skipped:
- Illumination — narrowband LEDs fired one band at a time, or broadband light through filters.
- Sensor — a monochrome camera with the IR-cut/hot-mirror removed.
- Geometry — rock-solid copy stand or repro rig so all bands register pixel-for-pixel.
- Calibration — a spectral reference target (e.g. a Spectralon or X-Rite chart) and flat-field frames.
- Processing — stack alignment, normalisation, and statistical separation.
A weak link anywhere shows up as colour fringes, drift between bands or noise that swamps the faint text.
Which wavebands matter for manuscripts?
A defensible starter set covers the responses where document materials differ most:
| Band (nm) | Region | Typical use |
|---|---|---|
| 365 | UV-A | Fluorescence of parchment, retouching, sizing |
| 450 | Blue | Carbon vs. iron-gall separation |
| 535 | Green | General legibility baseline |
| 625 | Red | Faded red lead / vermilion contrast |
| 730 | Far red | Iron-gall fading reveal |
| 850 | NIR | Carbon ink stands out, stains drop away |
| 940 | NIR | Penetrates surface grime and foxing |
Eight to twelve bands across this span handle the vast majority of European codices.
What does a minimal processing run look like?
After capture, align and stack the bands, then let statistics find the contrast for you. A typical Python starting point:
python
import numpy as np
from skimage import io, exposure
from sklearn.decomposition import PCA
# stack: list of 16-bit single-band TIFFs, already registered
bands = [io.imread(f"band_{nm}.tif").astype(np.float32) for nm in
(365, 450, 535, 625, 730, 850, 940)]
cube = np.stack(bands, axis=-1) # H x W x B
flat = cube.reshape(-1, cube.shape[-1])
pca = PCA(n_components=4).fit_transform(flat)
pc2 = pca[:, 1].reshape(cube.shape[:2]) # PC2 often isolates faint text
out = exposure.rescale_intensity(pc2, out_range=(0, 255)).astype(np.uint8)
io.imsave("pc2_reveal.png", out)Principal component 1 usually captures overall brightness; the faint writing tends to surface in PC2 or PC3. Always inspect every component rather than trusting the first.
How do I keep results honest and repeatable?
Record paradata: every band's wavelength, exposure, aperture, lamp, target reading and software step. Without it, a striking false-colour image is unverifiable and another researcher cannot reproduce or challenge it. Treat the calibrated 16-bit raw stack as the archival object; the false-colour reveal is only an interpretation of it.
What are the common beginner mistakes?
- Using a colour DSLR and assuming the "IR photo" is multispectral — it is not.
- Skipping flat-field correction, so lamp hotspots masquerade as features.
- Letting the page shift between bands, producing rainbow edges on text.
- Over-stretching one band and declaring victory instead of comparing all of them.
- Throwing away raw frames after exporting a JPEG.
Key Takeaways
- Multispectral imaging captures one page under many narrow bands (≈365–940 nm) to reveal what white light hides.
- A monochrome sensor plus narrowband LEDs is the core; a colour DSLR cannot substitute.
- Eight to twelve well-spaced bands handle most manuscripts; 16 add finesse, not magic.
- Rigid geometry, flat-fielding and a spectral target are non-negotiable for trustworthy results.
- PCA or band-ratio methods surface faint text — check every component, not just PC1.
- Archive 16-bit raw bands and full paradata; false-colour images are derivatives, never the record.
Frequently Asked Questions
What is multispectral imaging of a manuscript?
It is capturing the same page under a series of narrow wavebands from ultraviolet to near-infrared, then comparing those frames to reveal ink, erasures or damage invisible under white light.
How many wavebands do I actually need?
A practical entry set is 8 to 12 bands spanning roughly 365 nm to 940 nm. Sixteen bands give finer control, but 8 well-chosen bands already separate most iron-gall and carbon inks.
Do I need a special camera?
You need a monochrome sensor without a Bayer filter or hot-mirror, plus narrowband LED illumination or filters. A colour DSLR with an IR-cut filter cannot do true multispectral capture.
Why monochrome instead of a colour camera?
A monochrome sensor records the full spectral response of every pixel, so each waveband is captured at native resolution. A colour sensor interpolates across a Bayer mosaic and discards most of the UV and IR signal.
What file format should I keep the captures in?
Archive 16-bit linear TIFFs of every raw band plus flat-field references. Derivatives such as false-colour JPEGs are fine for sharing but must never replace the calibrated raw stack.
Is multispectral imaging safe for fragile parchment?
Yes, when you use LED illumination with no UV-C, keep exposure brief and monitor surface temperature. LEDs emit far less heat and harmful UV than older xenon or tungsten lamps.