Liang, H. Advances in Multispectral and Hyperspectral Imaging for Archaeology and Art Conservation. Applied Physics A: Materials Science & Processing

Abstract

Multispectral imaging has been applied to the field of art conservation and art history since the early 1990s. It is attractive as a non-invasive imaging technique because it is fast and hence capable of imaging large areas of an object giving both spatial and spectral information. This paper gives an overview of the different instrumental designs, image processing techniques and various applications of multispectral and hyperspectral imaging to art conservation, art history and archaeology. Recent advances in the development of remote and versatile multispectral and hyperspectral imaging as well as techniques in pigment identification will be presented. Future prospects including combination of spectral imaging with other non-invasive imaging and analytical techniques will be discussed.

Full text

Appl Phys A (2012) 106:309–323
DOI 10.1007/s00339-011-6689-1
INVITED PAPER
Advances in multispectral and hyperspectral imaging
for archaeology and art conservation
Haida Liang
Received: 19 May 2011 / Accepted: 7 November 2011 / Published online: 19 November 2011
© Springer-Verlag 2011
Abstract Multispectral imaging has been applied to the
field of art conservation and art history since the early 1990s.
It is attractive as a non-invasive imaging technique because
it is fast and hence capable of imaging large areas of an ob-
ject giving both spatial and spectral information. This paper
gives an overview of the different instrumental designs, im-
age processing techniques and various applications of mul-
tispectral and hyperspectral imaging to art conservation, art
history and archaeology. Recent advances in the develop-
ment of remote and versatile multispectral and hyperspec-
tral imaging as well as techniques in pigment identification
will be presented. Future prospects including combination
of spectral imaging with other non-invasive imaging and an-
alytical techniques will be discussed.
1 Introduction
Multispectral and hyperspectral imaging collect images of
an object in a series of spectral windows. They are efficient
methods for collecting millions of spectra since a spectrum
is measured for each spatial pixel (Fig. 1). The distinction
between multispectral and hyperspectral imaging is rather
blurred and very much discipline dependent. In general, hy-
perspectral imaging consists of more finely divided spectral
channels than multispectral imaging. Multispectral imaging
can sometimes refer to a set of images taken at vastly differ-
ent parts of the electromagnetic spectrum, e.g. three visible
images in red, blue and green, an infrared image and an X-
ray image of an object [1, 2]. We will not use this definite of
H. Liang (

)
School of Science and Technology, Nottingham Trent University,
Clifton Lane, Nottingham NG11 8NS, UK
e-mail: haida.liang@ntu.ac.uk
multispectral imaging here. For the rest of the paper, we will
refer to multispectral and hyperspectral imaging together as
spectral imaging or imaging spectroscopy.
Multispectral and hyperspectral imaging were first devel-
oped for remote sensing, which can include planetary sci-
ence and astronomy [3]. More recently, spectral imaging has
found applications in terrestrial laboratories for applications
in biology, medicine, chemistry, industrial sorting, quality
control and surveillance [4, 5]. Multispectral imaging has
been applied to the field of cultural heritage since the early
1990s. As a non-invasive imaging technique, it has the ad-
vantage over invasive techniques in that investigations can
be carried out on any object (even on intact and fragile ones
where samples cannot be taken) and anywhere on an object.
It was first applied for qualitative band to band comparison
in order to identify areas of different material composition,
natural degradation of material, past conservation interven-
tion, preparatory sketches, and quantitatively for improved
precision in colour measurement. Later, with increased num-
ber of bands and speed of acquisition, it was used to extract
spectral reflectance information for pigment identification.
It can operate in the UV, visible, the near infrared and in
UV-fluorescence mode.
Multispectral and hyperspectral imaging are in increas-
ing demand in the field of art conservation, art history and
archaeology judging by the number of recent reviews on
the subject from the conservation and archaeology commu-
nity [6, 7].
Spectral imaging has mostly been applied to paintings
and manuscripts. In the case of paintings, multispectral
imaging was first developed to increase the colour fidelity
of the images. In the past 20 years, a number of EU projects
has been dedicated to the design and implementation of
high colour fidelity, high resolution scanning systems for the

310 H. Liang
Fig. 1 A schematic diagram illustrating a spectral cube obtained from
multispectral imaging, the colour image derived from the spectral cube
and a spectrum for a point on the blue colour which can be identified
with the pigment azurite
recording of museum paintings and other objects of art (e.g.
[8–13, 55]).
2 Instrument design considerations
A spectral imaging system needs the following essential
components: lighting, focusing optics, detector and most im-
portantly a means of wavelength selection. There is a diverse
range of methods for wavelength selection which determines
the design of the illumination system and the spatial and
spectral scanning strategy.
2.1 Illumination requirements
One of the special requirements for imaging of heritage ob-
jects is minimum light exposure to ensure that light-induced
ageing of the objects is kept at a minimum. There is a vast
body of work done on light-induced ageing which has in-
formed display strategies in museums (e.g. [14]). In gen-
eral, UV and thermal radiation are eliminated from a white
light source before it is used to illuminate objects in muse-
ums. For example, for oil paintings the recommended level
is ∼200 lux and for manuscripts and other paper-based art-
works the lighting level is kept at ∼50 lux. It is generally as-
sumed that the reciprocity principle holds, which basically
says that light-induced damage is determined by the accu-
mulated total energy incident on a material rather than the
intensity of the incident light. The reciprocity principle has
been widely used to justify high intensity illumination for
fast imaging (e.g. [11]). To achieve the same signal-to-noise
ratio in an image, one can either illuminate the object with
a low intensity light for longer or a high intensity light for
a shorter period of time. Since it is the total energy incident
on an object that determines the damage, it appears sensible
to use a higher intensity light to increase the imaging effi-
ciency without causing extra damage. However, extra cau-
tion must be exercised at high intensity levels since the reci-
procity principle must break down at some level. Saunders
et al. [15] conducted a survey of 21 light sensitive pigments
and found that they all follow the reciprocity principle up to
8 × 10
3
lux. Our recent studies using a microfading spec-
trometer [16] have found that some light sensitive pigments
such as orpiment, where the light-induced degradation is due
to competing reactions, reciprocity does not hold [17]. It was
found that for these very sensitive pigments illuminated with
2 ×10
6
lux of incident light, it takes ∼30 seconds illumina-
tion for any measurable damage to occur and minutes for
any damage that is noticeable by the naked eye. The rate of
degradation is always greatest within the first ∼30 seconds.
Therefore, care must be taken to choose the right illumina-
tion level. It might be sensible to do spot tests for light sen-
sitive material using microfading spectrometry prior to high
intensity illumination.
2.2 Wavelength selection
Wavelength selection can be achieved either on the illumi-
nation light path such that only a selected wavelength range
of light is incident on the object at a time, or on the reflected
light path before the detector such that the light reflected
from the object can be separated spectrally.
2.2.1 Wavelength selection through illumination
The first multispectral imaging system designed for paint-
ings was through filtering the illumination and using a
monochrome digital camera to collect the reflected light [9].
Interference filters were placed in front of a halogen–
tungsten light source. The advantage of such a system is eco-
nomic light exposure since only a narrow wavelength range
is incident on the object at a time. This is very important
for light sensitive materials especially paper-based works of
art or manuscripts. The requirements for the filtering system
are high throughput and low out-of-band response within the
detector sensitive wavelength range, and good thermal sta-
bility as they are placed close to the light source. The other
advantage is that any off-the-shelf monochrome camera and
lens system can be used without modification. However, in
practical terms an average camera lens has significant chro-
matic aberration such that images collected at widely differ-
ent wavelengths will have different focal length resulting in
significant difference in magnification [18] which will then
need to be corrected in a post-processing software. While
such a system is compatible with a wide range of cameras,
it can be less flexible and portable in terms of the light-
ing component of the system. The contribution from back-
ground light can be significant when there is no spectral

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 311
filtering in front of the camera, which restricts the use of
such systems outside studios. With the development of lu-
minous LEDs, this option is becoming more attractive since
LEDs are energy efficient [19]. One of the potential disad-
vantages of such a system is that they cannot be used for
UV-fluorescence imaging because there is no spectral filter-
ing in front of the detector.
2.2.2 Wavelength selection in the reflected light
Alternatively, wavelength selection can be achieved through
either filtering or dispersing the reflected light. In the snap-
shot mode, the entire spectrum is collected simultaneously
per spatial point and a spatial area is collected through scan-
ning in a time sequence. In the sequential mode, the spa-
tial field of view is imaged through one wavelength chan-
nel simultaneously and the full spectral cube is collected
through sequential spectral filtering. The snapshot mode has
its advantage when imaging objects that are time varying
over a time scale shorter than the imaging time. These can
be of concern in remote sensing, astronomy or biomedical
imaging where there are time dependent signals. In imaging
of cultural artefacts, this is rarely a problem which is why
only recently that imaging in the snapshot mode has been
adopted.
2.2.3 Sequential spectral filtering
Sequential spectral filtering usually involves placing the
filtering system between the lens and the detector, which
means that specialised lens may have to be used to adapt
to the increased distance between the lens and the detector.
Placing the filtering system in front of the lens is practically
difficult because of the large aperture of the filtering system
required to avoid vignetting. Filtering systems with such a
large aperture is either costly or not available.
The simplest method for wavelength selection is to place
bandpass interference filters between the lens and the de-
tector [12, 13, 20, 21, 47]. Filter selection can be com-
puter controlled through a motorised filter wheel. Custom
designed filter sets for imaging purposes that have the same
optical thickness will ensure that the image scale stay the
same for different wavelengths. Such a filter set combined
with a high quality lens designed for minimum chromatic
aberration ensures that no image scaling is necessary be-
tween channels [13]. However, a tilt of the position of the
filter by a fraction of a degree can result in noticeable shifts
in the image. As will be discussed in Sect. 4, image shift-
ing and re-scaling can easily be dealt with automatically
in post-processing. The advantage of interference filters is
large aperture, large field of view, good optical quality and
low cost. The disadvantage is the fixed number of filters and
slow speed in filter change since it takes time for the filter
wheel to move and settle into a new position. Linear/circular
variable interference filters can also be used to mechanically
move through the wavelength channels.
Tunable filters such as Liquid Crystal Tunable Filters
(LCTF) and Acousto-Optical Tunable Filters (AOTF) of-
fer rapid tuning between wavelength channels and have no
moving parts. Both offer random access for wavelength se-
lection. By taking care in the optical design, it is possible
to produce high quality images that are not shifted between
wavelength channels.
LCTF can be tuned by changing the voltage applied to
a birefringent liquid crystal. It is essentially an electrically
tuned Lyot filter. The tuning speed is given by the relaxation
time of the crystal and is typically of the order of tens to hun-
dreds of milliseconds. The bandwidths of the filters are fixed
and increases with central wavelength. The transmission ef-
ficiency decreases with decreasing wavelength resulting in
relatively low blue response [22]. LCTFs are polarisation
sensitive which means that the maximum overall through-
put is less than 50%. Hyperspectral imaging using LCTF
has been explored for heritage applications in the 400 nm–
700 nm range [23, 24] and the 650 nm–1040 nm range [25].
Since the reflectance spectra of pigments in the 400 nm to
1000 nm range is fairly smooth, the increased spectral reso-
lution of a hyperspectral system (at the expense of decreased
signal-to-noise ratio) did not provide much additional infor-
mation.
AOTF is tuned by changing the frequency of acoustic or
Radio Frequency (RF) waves applied to a birefringent crys-
tal, typically TeO
2
for wavelength range between 350 nm
and 5000 nm. The RF waves travelling across the crystal es-
sentially set up a volume diffraction grating by producing a
periodic modulation of the refractive index within the crys-
tal through the photo-elastic effect [26]. AOTF is polarisa-
tion sensitive. The output from an AOTF consists of the un-
diffracted ray and the two orthogonally polarised diffracted
beams angularly separated. There is a one-to-one correspon-
dence between the RF frequency injected and the central
wavelength of transmission. The bandwidth can also be tun-
able through injection of multiple RF frequencies [27]. The
minimum achievable bandwidth also increases with wave-
length. The speed of tuning is given by the travel time of
the acoustic wave in the crystal and is typically of order of
hundred microseconds [28]. While it is possible to recom-
bine the orthogonally polarised diffracted beams to increase
overall throughput, in practice it is difficult to maintain high
optical quality through such combination.
Both LCTF and AOTF can tune up to an octave. AOTFs
in general have higher throughput compared with LCTF and
do not have a decreasing throughput with decreasing wave-
length. On the other hand, LCTF have larger aperture and
angular field of view. While AOTFs have faster tuning speed
than LCTF, the tuning time in either case is insignificant

312 H. Liang
compared with typical integration times and readout speed
of an average 1000 ×1000 pixels CCD detector.
Other tunable devices include the Fabry–Pérot etalon and
the Fourier transform Michelson interferometer. In the case
of the Fabry–Pérot etalon, wavelength tuning is achieved
through mechanically changing the distance between the
mirrors using piezo-electric transducers [29, 30]. These de-
vices have large aperture but small angular field of view.
The Fourier transform Michelson imaging interferometer
can also be used through scanning one of the two mirrors
and collecting the two-dimensional interferogram on the de-
tector. The Fast Fourier Transform (FFT) of the interference
pattern of a pixel gives the reflectance spectrum. This is
essentially performing FTIR spectroscopy for a number of
spatial pixels simultaneously.
2.2.4 Simultaneous spectral collection
One of the most common methods of spectral collection in
remote sensing is to use a slit combined with a diffraction
grating to disperse the light. A two-dimensional detector can
collect a series of spectra corresponding to spatial points
aligned with the slit simultaneously. A full 3D spectral cube
can be collected by scanning spatially in the direction per-
pendicular to the slit [31–33]. Such devices can offer much
higher spectral resolution than the tunable filters described
above but with reduced flexibility. The full spectral range
with a fixed spectral resolution is always collected. How-
ever, high spectral resolutions of less than 10 nm are gener-
ally not needed since spectral reflectance of pigments have
fairly smooth spectral shape. Similarly, in remote sensing
applications, it is well recognised that the maximum spec-
tral resolution necessary is between 5 nm and 20 nm [34].
Since the full spectral range is collected at once, it places
high demands on the chromatic aberration tolerance of the
lens. In the sequential filtering designs discussed previously,
there is more flexibility in adjusting the focus and integra-
tion time for each individual spectral channel for best focus
and optimum signal-to-noise ratio.
Sagnac imaging interferometer is also used for the simul-
taneous spectral collection in a line of spatial pixels similar
to the grating spectral imager, except a FFT is needed to re-
cover the spectrum per pixel. Such a device has been used
by Casini et al. [35] for UV-fluorescence imaging.
An alternative method of simultaneous wavelength col-
lection has been used for multispectral imaging of paint-
ings [36–38] where the reflected light from a single point
is collected and distributed using a fibre bundle to a series
of detectors with different interference filters. Such a device
has the potential to reduce light scattering if only a small
region around the point being measured is illuminated (sim-
ilar to a fibre optic spectrometer setup). The drawback is the
limited speed of capture.
Finally, another simple method of simultaneous spectra
collection can be achieved by sacrificing spatial resolution
of a CCD camera by having different filters placed in front
of individual pixels similar to the concept of a RGB colour
camera except more than 3 filters are involved (e.g. [39]).
Such an instrument involves modifying the CCD sensor de-
sign which could be expensive. It is possible to recover the
full spatial resolution by dithering the camera and taking
multiple shots. The spectra measured simultaneously are
not from exactly the same spot unlike previous methods.
A hybrid design where a small number of interference fil-
ters are placed in front of a RGB colour camera rather than
a monochrome camera has been applied to imaging paint-
ings [40]. In this case there is no need for a special detector
and a simple RGB camera can be used. The various combi-
nation of filters in front of the detector and on the detector
pixels provides extra spectral resolution. However, the dis-
advantage is the attenuation of collected photons by going
through two sets of filters as well as loss of spatial resolu-
tion.
2.3 Detectors
In the wavelength range between 300 nm–1000 nm, silicon
CCD detectors are used. For low noise detection, Peltier
cooled systems are preferred. In the Short Wave Infrared
(SWIR) range between 900 nm–1700 nm, mechanically
cooled InGaAs detector arrays are the most commonly used
due to their high sensitivity and low cost. While extended
InGaAs can reach beyond 2 µm, they cannot compete with
HgCdTe in sensitivity. HgCdTe detectors are sensitive to a
broad wavelength range from ∼1µmto∼10 µm. Currently,
spectral imaging have mostly been conducted in the range
between 400 nm and 1700 nm, mainly because of the huge
jump in cost for detector arrays that are sensitive beyond
1.7 µm. Some early works have pushed this up to 2.2 µm us-
ing low sensitivity analogue detectors such as the PbO-PbS
vidicon. The only recent device used in art applications that
managed to reach 2.3 µm is the one that uses single element
InGaAs detectors [38].
3 Calibration and sources of error
The method of calibrating multispectral or hyperspectral
imaging involves the normal procedure of calibrating CCD
images (e.g. [41]) as well as the calibration of the spectral
response of the imaging system.
Thermal noise associated with CCD detectors include
dark noise and readout noise [41]. The dark current and the
noise associated with it can be reduced by using a cooled de-
tector. The readout noise is also determined by the speed of
the readout. The shot noise or photon noise is determined by

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 313
the Poisson statistics of the quantum arrival of photons and
is given by the square root of the number of photo-electrons
detected.
Calibration of spectral images involves taking the follow-
ing calibration images:
– dark images of the same integration time as the object
frames but with the lens cap on (or illumination off) to
correct for the accumulated thermal dark current;
– flatfield images per channel of a uniform white target to
correct for the spatial inhomogeneity of illumination and
pixel-to-pixel gain variation of the CCD;
– images of a spectral standard through each wavelength
channel to correct for the spectral response of the imaging
system.
The reflectance at a pixel (the ith pixel) captured with a
spectral channel of central wavelength λ is given by
R
i
(λ) =R
W
(λ)g
(I
i
(λ) −D
i
(λ))f
i
(λ)

n
[(W
i
(λ) −DW
i
(λ))f
i
(λ)]/n
(1)
where R
W
is the true spectral reflectance of a spectral stan-
dard, I
i
is the counts for the light reflected from the object,
D
i
is the dark counts for the same integration time as the ob-
ject frame, f
i
is the flatfield correction factor for that pixel,
W
i
is the counts for the spectral standard and DW
i
is the
corresponding dark counts for the same integration time as
the standard, g is the scale factor to adjust the integration
time of the object to that of the spectral standard and n is the
number of pixel over which to average the response of the
spectral standard.
While CCD detectors are highly linear, non-linearity is
still observed close to saturation and at very low expo-
sure times. Optimum exposure time for maximum signal-
to-noise is determined by the maximum counts over which
the CCD is linear. For best quality images and spectral re-
flectance measurements, exposure times should be adjusted
per channel for maximum signal-to-noise ratio images.
It is best to capture images after the light has been
switched on for 10–20 minutes for a stable intensity of illu-
mination and to avoid temperature gradients along the imag-
ing pathlength which can degrade the spatial resolution of
the system (similar to the ‘seeing’ effect in astronomy).
The measured spectral reflectance of an object depends
on the geometry of the illumination/collection setup which
determines the relative amount of surface reflection to that
of the diffuse reflection. For comparison between materials,
it is best to image the objects in the same setup as the ref-
erence material. This is particularly important for systems
with moderate spectral resolution.
Finally, the cross-talk due to scattered light (similar to the
adjacency effect in remote sensing) can result in inaccuracy
in the measured spectral reflectance. A dark area surrounded
by bright regions would appear to be brighter than if it had
been surrounded by dark regions, because of light scattered
from the surrounding regions. Martinez et al. [56]showed
that this scattering can be modelled as a linear function
I
m
(i) =I
t
(i) +αI (2)
where I
m
is the measured intensity of a pixel, I
t
is the true
intensity,
I is the mean intensity of the image and α is the
cross-talk scattering coefficient which can be measured by
imaging a set of calibration targets with a neutral grey sur-
rounded by white, grey and black backgrounds.
4 Image alignment and mosaicing
As mentioned above, for some instrument designs it is ex-
pected that the images from different spectral channels need
to be shifted or even spatially re-scaled in order to have an
aligned image cube. There are various algorithms associated
with image registration (e.g. [42]). Image registrations in-
volving only linear shifts are relatively simple and can be
found by performing cross-correlation. Spatial image scal-
ing involving re-sampling could result in some loss of infor-
mation, therefore it is best to design the optics of the system
to avoid scaling of the images.
High resolution imaging of large objects inevitably in-
volves mosaicing of images. Adjacent images need to be
taken with sufficient overlap to allow automatic image reg-
istration. Since the shifts are linear, simple cross-correlation
algorithm can be used for image registration [43–45].
5 Applications
5.1 Qualitative inter-band comparisons for revealing
hidden information
Applications of spectral imaging in art conservation include
the detection of damages and past intervention through inter-
band comparisons. Figure 2 shows that the infrared spectral
bands reveal areas of damage (black spots and a crack) that
had been repaired and retouched such that they are invisible
in the colour image and in many of the other visible spec-
tral bands. Near infrared bands are particularly useful for
this purpose and for revealing the underdrawings (prepara-
tory sketch) beneath the paint layers. Figure 3 shows that
the 880 nm image reveals clearly the preparatory sketches
beneath the paint layer. The sketches were drawn with a
solid material and some of it were made by pouncing. Such
information is invaluable to art historians in studying the
techniques of painting and for attribution and authentication.
Ever since the pioneering work of van Asperen de Boer in
the 1960s [46] in developing a vidicon infrared camera for
imaging underdrawings, it has been known that pigments

314 H. Liang
Fig. 2 Multispectral images
(400–1100 nm) taken with an
upgraded VASARI system using
the CRISATEL set of
interference filters placed
between the lens and the
detector [13], and the derived
colour image (see Sect. 5.2)
rendered under the D65
illuminant of a region of a
painting by Carlo Crivelli, Saint
Catherine of Alexandria
(NG 907.2) from the National
Gallery, London collection
Fig. 3 Colour photo of a
painting, The Virgin and Child
with an Angel (NG 3927) after
Francesco Francia (image
copyright The National Gallery,
London) and the 880 nm image
of a region around the eye of the
angel taken with PRISMS (see
Sect. 6)[90]) at a distance of
6 m showing the preparatory
drawings under the paint layer
are most transparent to infrared radiation in the 1 µm to
2 µm range. Earlier near infrared multispectral/hyperspectral
imaging systems for the 1 µm to 1.5 µm were developed us-
ing infrared to optical converters or vidicon tube technology
which is not linear and has low sensitivity [20, 47]. Nowa-
days InGaAs detectors are much more affordable and there
are a number of SWIR spectral imaging systems developed
for heritage applications [32, 38, 92].
Traditionally, UV-fluorescence imaging is considered to
be the standard method for revealing erased or faded writing
on manuscripts. However, it appears that other non-visible
bands such as reflected light from the UV (rather than fluo-
rescence) [49] and infrared bands [100] can also be effective
at revealing ‘hidden’ writings. Figure 4 shows that a page of
a prayer book that looked blank in the visible is revealed in
the 880 nm channel to have a signature written on the top
part of the page [100]. Multispectral imaging has been suc-
cessfully used to separate the erased writing of Archimedes
palimpsest from the later over-written text [48] and hyper-
spectral imaging has been shown to be effective in visual
enhancement of old documents corrupted by ink bleed and
ink corrosion [50].
5.2 Colour rendering
Multispectral imaging enables rendering of colour accurate
images of paintings under any lighting conditions, unlike a
normal tri-colour image which can only capture an accu-

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 315
Fig. 4 (a) A colour photo of
folio 202r of MS Lat. Liturg.
g.1. from the Bodleian Library
collection (courtesy of the
Bodleian Libraries, University
of Oxford); (b) near infrared
image at 880 nm taken with a
modified PRISMS (see Sect. 6)
[90]ofthetopregionofthe
page showing a clear
signature [100]
rate colour image under the specific illumination used at the
time. The tri-stimulus values, X, Y and Z can be derived
from the spectral reflectance through
X =k

∞
0
R(λ)S(λ)x(λ)dλ,
Y =k

∞
0
R(λ)S(λ)y(λ)dλ,
Z =k

∞
0
R(λ)S(λ)z(λ) dλ, (3)
k =
100

∞
0
S(λ)y(λ)dλ
(4)
where R is the spectral reflectance of the material, S is the
spectrum of the illuminant normalised to 100 at 560 nm and
x, y, z are the colour matching functions for the standard 2
◦
observer [51]. Once the spectral reflectance of a material has
been determined, the colour of the material under illumina-
tion of any spectral density distribution can then be uniquely
derived (Fig. 5). The tri-stimulus values are often converted
to the CIE standard colour coordinates L,a, b for a uniform
colour-space [51, 52]. Discussions on the various CIE colour
difference formula are given in [53, 54]. The colour differ-
ence ΔE is used to describe small colour differences and in
conservation they are used to describe the fading of coloured
material.
5.3 Conservation monitoring
Accurate recording of the spectral and therefore colour of
paintings has been used for monitoring the change due to
natural degradation, before and after transport, exhibition
and conservation treatment as detailed in a review given by
Martinez et al. [56]. In addition, due to the increased use
of laser cleaning, multispectral imaging has been used for
the on-line monitoring of laser cleaning of marbles [57, 58],
paper and parchment [59]. Both UV reflectance and UV-
fluorescence images were found to be particularly useful in
Fig. 5 Colour images derived from multispectral images taken with
the upgraded VASARI system using CRISATEL set of interference fil-
ters [13, 55]ofHeads of Angels (NG 1842), an Italian fresco in the
National Gallery, London: (a) rendered in daylight (D65) and (b)ren-
dered in candle light (illuminant A) [55]
monitoring the cleaning process and investigating potential
damages caused by laser cleaning.
One of the greatest problems with conservation of wall
paintings and walls of historical structures is moisture. The
water absorption bands around 1.4 µm and 1.9 µm can be
used to monitor moisture content. Our group has used the

316 H. Liang
PRISMS SWIR hyperspectral imager (see Sect. 6) to moni-
tor moisture in walls using the 1.4 µm band.
For monitoring purposes, good absolute calibration is
necessary especially because the changes are likely to be
very small. The earlier these changes can be detected, the
more valuable the monitoring exercise is.
5.4 Pigment identification
Pigment identification using spectral reflectance has a long
history (e.g. [60]), however, non-invasive spectral pigment
identification in the visible and near infrared has not been
met with a lot of enthusiasm in the conservation commu-
nity so far. This is partly because the technique is per-
ceived as not always yielding definitive identifications, es-
pecially where there is a complicated pigment mixture or
that if the paint has degraded. In addition, for oil paint-
ings, tiny paint samples can be taken from paintings pro-
vided that the sampling sites are near cracks and edges. His-
torically, this became possible when micro-chemical analy-
sis became viable. Invasive chemical analyses always yield
more chemically specific information, and often times mul-
tiple chemical analysis methods can be used on one sam-
ple. For paintings where multispectral imaging was first ap-
plied to, there has not been an overwhelming need for non-
invasive pigment identification. This is, however, not the
case for manuscripts where sampling is not possible. In par-
ticular, for illuminated manuscripts, non-invasive pigment
identification is particularly valuable both for assisting con-
servation decisions and for historical studies [100].
One of the main obstacles with pigment identification
using spectral reflectance is the lack of comprehensive
databases of reference pigment and paint. A number of
groups have their own reference pigment sets but few have
a comprehensive set that has been systematically prepared
and measured (e.g. [61, 62]). The only one that is publicly
available is the CNR-IFAC on-line database of spectral re-
flectance from 270 nm–1700 nm of various pigments in
common binding media found on Western European paint-
ings [62].
In the case of mixture of pigments, the identification of
the pigments involves the ‘unmixing’ of the spectral com-
ponents of the constituent pigments. Spectral unmixing in
this context is very different from those in remote sens-
ing [64, 65]. In remote sensing, spectral unmixing is neces-
sary mainly because the spatial resolution is relatively low
due to the great distances involved and hence the spectral re-
sponse from a pixel is a combination of various material that
are spatially separated. Linear unmixing is therefore suffi-
cient in such cases to separate out the spectra of the con-
stituent material. For paint mixtures, the small pigment par-
ticles are uniformly dispersed in the binding medium and
the reflectance spectrum of the paint mixture is not simply a
linear mixture of the spectral reflectance of the paints of sin-
gle pigments. Light transport in a medium is described by
the radiative transfer (or transport) equations first developed
in astrophysics to describe the effect of stellar atmosphere
and interstellar medium on light propagation [70, 71]. Spec-
tral unmixing for paint is best performed by modelling the
physics of light transport in a turbid medium.
The Kubelka–Munk (KM) theory [72, 73] involving just
two diffuse fluxes in the forward and backward direction
which successfully describes the spectral reflectance of paint
layers is a simple approximation to the full radiative trans-
fer model. The main assumptions of the KM theory are
(1) the paint layers or any turbid medium is homogeneous
in the sense that the particle sizes are much smaller than
the thickness of the layer; (2) the transverse extent of the
layer is much greater than the thickness; (3) the illumination
is diffuse and the light propagation in both directions are
uniformly diffuse. Limitations of the simple two-flux KM
theory and comparisons with simple improvements are de-
scribed in many papers (e.g. [74, 75]). The KM theory re-
lates the diffuse reflectance (R) of the layer without inter-
faces to the effective absorption (K) and effective scattering
coefficients (S) through
R =
1 −R
g
[a −b coth(bSh)]
a +b coth(bSh) −R
g
(5)
where a =(K +S)/S, b =
√
a
2
−1, h is the layer thickness
and R
g
is the reflectance of the substrate. In the limit of a
paint layer with infinite optical thickness, the ratio between
the absorption and scattering coefficients is related to the
reflectance by
K
S
=
(1 −R
∞
)
2
2R
∞
(6)
where R
∞
is the spectral reflectance of a layer with infi-
nite optical thickness. It should be noted that K and S do
not always have a linear relation to concentration. KM the-
ory assumes diffuse illumination and collection, however, in
practise, measurements are often made with collimated nor-
mal illumination [75]. Saunderson’s correction [76] is nor-
mally applied to correct for the reflection at the air/paint and
paint/air interfaces. The KM theory is so simple and suc-
cessful that it is still being used in the paint industry to work
out the mixing ratios of paints to match a given colour.
It is usually assumed that the effective absorption and
scattering coefficients (K and S) of a mixture is a linear
combination of the constituent coefficients weighted by their
concentrations. In the case where the pigments are all mixed
with a highly scattering white paint, the final pigment mix
can be described as a linear combination of K/S since the
white paint dominates the scattering [77]. The advantage of
this simplification is that only one measurement of the spec-
tral reflectance of a reference paint is necessary. Without

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 317
Fig. 6 The measured reflectance spectra of red earth in egg tempera
(red curve) and azurite in egg tempera (blue curve)aswellastheir
mixture (green curve). A spectrum of a mixture derived from the KM
theory is shown in comparison (dashed black curve)
this simplification, it is then necessary to measure K(λ) and
S(λ) for each reference paint which usually involves either
measurements of a semi-transparent paint layer over white
and black substrates or measurements of the same paint with
different thickness (always semi-transparent) over the same
substrate.
If a mixture of different paint components can be mod-
elled as a linear combination of K/S, then we can predict
the reflectance of the paint mixture. This is sometimes re-
ferred to as the single constant KM theory. Since the aim
here is to identify the pigments in a mixture, the concentra-
tions of the single pigments were set as free parameters to
find the best non-negative least squared fit of the predicted
spectrum to the actual measured spectrum of the mixture for
each combination of reference pigments [66, 78]. Liang et
al. [66] showed that the method worked even for pigment
mixtures without the addition of white pigments, except in
cases where pigments with high absorption or very low scat-
tering are involved. For pigments with high K and low S,it
is necessary to use the spectra of those pigments mixed with
a white pigment to obtain a correct spectrum for the paint
mixture. Figure 6 shows that a mixture of red earth and azu-
rite paint can be correctly identified using the above method
and the measured spectral reflectance of pure red earth paint
and pure azurite paint. Figure 7 shows that a mixture of in-
digo and orpiment can be correctly identified using the spec-
tral reflectance of indigo mixed with lead white and pure or-
piment spectrum but not with pure indigo and pure orpiment
spectrum since indigo has a very high absorption coefficient.
In contrast, Latour et al. [79] prepared a number of ref-
erence paint samples over black and white substrates and
measured their K and S and assumed that the K and S of
the mixture are linear combinations of their respective con-
stituents to deduce the pigment mixture using a method sim-
ilar to the above. This is sometimes referred to as the two
constant KM theory.
Fig. 7 (a) The reflectance spectrum of indigo in linseed oil (blue
curve) and orpiment in linseed oil (red curve)aswellasamixtureof
the two pigments in linseed oil (green curve). The spectrum of indigo
and orpiment mixture is inconsistent with the spectrum of a mixture de-
rived from KM theory using the single pigment paint spectra (dashed
curve). (b) The reflectance spectrum of the indigo and orpiment mix-
ture predicted from KM theory (dashed curve) using the spectra of an
indigo mixed with lead white in linseed oil (blue curve) and an orpi-
ment in linseed oil spectrum (red curve) correctly predicts the spectrum
of an indigo and orpiment mixture (green curve)
It is important to know the effect of the various parame-
ters associated with a paint that can affect pigment identifi-
cation. The effect of particle size, concentration and types of
binding medium on the spectral reflectance have been sys-
tematically studied by Feller [63] and Liang et al. [66]. Fig-
ure 8 shows that the pigment to medium ratio only affects
the peak of the spectrum but not the general spectral shape.
Figure 9 shows that the binding medium have by far the
most dramatic effect on the spectral reflectance, changing
even the ratios between the peaks, but the positions of the
peaks remain unchanged. On the other hand, changes in par-
ticle size of pigments can result in peak shifts by as much as
20 nm as demonstrated in Fig. 10 for azurite and malachite
with mean particle sizes between 30 µm and 3 µm.
Finally, it is important to know how the effect of sur-
face dirt and varnish might affect the measured spectral re-
flectance since historic paintings will have surface dirt and
sometimes degraded yellowed varnish on top of the paint

318 H. Liang
Fig. 8 The effect of pigment concentration in a binding medium on
the reflectance spectra is shown for azurite in egg tempera. The con-
centrations are 89% (red curve), 78% (green curve), 67% (blue curve)
and 56% (black curve)
Fig. 9 The effect of the type of binding medium on the reflectance
spectra is shown for smalt pigment in linseed oil (green curve), egg
tempera (red curve) and acrylic (blue curve)
surface. Figure 11 shows that surface dirt acts like a neutral
density filter and does not change the shape of the spectrum.
Since many Western European paintings are varnished and
old varnish tends to turn yellow, the effect of varnish is ex-
amined. Figure 12 shows that old varnish acts like a yellow
filter which suppresses the blue reflectance, leaving the rest
of the spectrum unchanged. The effect of relatively new var-
nish on spectral reflectance of paint has been studied in de-
tail by Berns and de la Rie [68] and Elias et al. [67]. This
study was also extended to thin layers of old varnish in the
longer wavelength spectral range where the effect of absorp-
tion due to the old yellowed varnish is not significant [69].
Figure 12 shows the significant absorption in the blue part
of the reflectance spectrum of a paint measured through an
old yellowed varnish.
Based on the knowledge of how the various factors such
as particle size could affect the spectral reflectance and the
method of identifying the pigment mixtures using the KM
theory, Liang et al. [66] devised an algorithm to automati-
cally identify pigments by fitting different combinations of
Fig. 10 The effect of particle size on the reflectance spectra for
(a) azurite and (b) malachite in egg tempera (red for grade 1, green
for grade 3 and blue for grade 5). Grade 1 is the largest (∼30 µm) and
grade 5 is the smallest (∼3µm)
reference paint spectra to an unknown spectrum. The best
fit for each combination as well as single pigment spectra
with and without white pigment mixtures are then cross-
correlated with the unknown spectrum. The best identifi-
cation is the one with the highest cross-correlation coeffi-
cient at zero offset. An extra 20 nm range around zero offset
should be allowed to account for the potential shift in peaks
due to particle size differences. Figure 13 shows an exam-
ple of a spectral pigment identification using multispectral
imaging of the painting in Fig. 3. The green paint was iden-
tified with a mixture of chrome yellow and Prussian blue.
This identification was verified by an independent examina-
tion under the microscope of a small sample from the green
region [82].
An alternative method to the KM theory, which solves
the radiative transfer equation to predict the bi-directional
reflectance of semi-transparent stratified paint and glaze lay-
ers, has been validated using known samples by Simonot et
al. [80] and Latour et al. [81].
5.5 Pigment mapping using multivariate statistics
Multivariate statistics [83] is often used in analysing hyper-
spectral image cubes in remote sensing [4]. Similarly it is

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 319
Fig. 11 The effect of surface dirt on the spectral reflectance. A colour
photo of a region of a mural painting from the tomb of Prince
Zhanghuai showing regions that have been test cleaned (copyright
Shaanxi History Museum). The spectral reflectance of a red paint
(from the region marked by a black box) in the area where it has been
test cleaned (green circles) and a neighbouring area that has not been
cleaned (red circles). The uncleaned reflectance spectrum is also scaled
up by multiplying a constant (red dots) to show that the surface dirt
does not change the spectral shape
also used in chemometrics for spectral analysis. In spec-
tral imaging of paintings, it is often used for rapid pigment
mapping. For example, Baronti et al. [20] used principal
component analysis to decompose the spectra into a num-
ber of orthogonal principal components. Similarly, Delaney
et al. [32] used the standard remote sensing software ENVI
to decompose the spectra into mutually independent end-
members. Most of these methods assume linear combina-
tion. Multivariate statistical methods are efficient at distin-
guishing different materials, however, the endmembers do
not necessarily have any physical meaning and therefore
cannot be compared to the spectra of reference material for
the identification of material.
5.6 UV-fluorescence
UV-fluorescence has been used in conservation since the
1920s. It has been used to qualitatively identify old var-
nish and other organic material. Spectral imaging provides a
means of quantitatively studying the fluorescence emission.
In spectral imaging systems where the filtering or wave-
length selection is between the object and the detector, UV-
fluorescence images can be readily obtained by using a low
energy UV light source with a filter that restricts the inci-
dent light to a narrow spectral range around the intended
excitation wavelength and blocking off radiation in the vis-
ible range of the spectrum where the material is likely to
have fluorescence emission. The main difficulties of obtain-
ing UV-fluorescence spectra is the radiometric calibration. It
requires measurement of the spectral response of the entire
imaging system. Early applications were semi-quantitative
(e.g. [35]), and later attempts used either indirect radio-
metric calibration [84] or measured directly the spectral re-
sponse of the filter system and the CCD detector to derive
the total spectral response of the system [85].
An extension of UV-fluorescence is laser induced fluo-
rescence (LIF) which was first applied to aid pigment analy-
sis [86]. LIF is a point-based technique where a spectrometer
is used to obtain the fluorescence signal emitted at the laser
illuminated spot. Recently, it has been used to map regions
of interest at discrete spectral bands [87].
6 Towards remote spectral imaging
There is always a trade-off between field of view and spa-
tial resolution. For high resolution imaging of large ob-
jects such as large wall paintings, it is necessary to image
a number of small areas at high resolution and then mo-
saic the images together. This usually requires imaging at
close range, which means for objects at lofty heights, such
as ceiling paintings, it is necessary to use scaffolding or me-
chanical lifting devices to move the imaging system close
to the areas of interest [88, 89]. Recently a spectral imag-
ing system specifically designed for remote imaging of wall
paintings, PRISMS (Portable Remote Imaging System for
Multispectral Scanning), has been developed in our group
to overcome such difficulties by using a small amateur tele-
scope [66, 90, 92].
PRISMS has a visible/NIR (400–880 nm) multispectral
imaging system as well as a short wave infrared (900–
1700 nm) hyperspectral imaging system (Fig. 14). The

320 H. Liang
Fig. 12 The effect of yellowed old varnish on spectral reflectance.
A colour image of a region of a test painting showing parts of the
blue sky covered with yellowed varnish and parts that have the var-
nish cleaned off, as well as regions that were re-varnished with a new
varnish. The spectrum from the yellowed varnish area (red curve)is
also scaled up by multiplying a constant (red dashed curve)andcom-
pared with the unvarnish area (green curve) to show that the yellowed
varnish absorbed strongly in the blue region
Fig. 13 The pigment mixture used in the green headdress of the virgin
in Fig. 3 is identified with a mixture of chrome yellow and Prussian
blue. The reflectance spectrum of a small area of the green headdress
(green curve) captured with the PRISMS (Sect. 6) multispectral camera
is compared with the spectrum derived from KM theory (dashed curve)
using a reference spectrum of chrome yellow in linseed oil (red curve)
and Prussian blue and lead white in linseed oil (blue curve)
VIS/NIR system is a low budget system that consists of a fil-
ter wheel with 10 bandpass filters each with a bandwidth of
40 nm except for the one at 880 nm which has a bandwidth
of 70 nm and a Peltier cooled Jenoptics CCD camera [90].
The SWIR system consists of a Gooch & Housego AOTF
spectrograph and a Xenics InGaAs camera. The best spec-
tral resolution of the AOTF is 10 nm and the central wave-
length and bandwidth can be rapidly tuned automatically
from a laptop [92]. A Meade ETX-90 telescope is used for
high resolution imaging at distances greater than 3.5 m and
lenses are used for close range imaging below 3.5 m. Spa-
tial resolution of tens of microns per pixel can be achieved
for distances of 10 m. The system also gathers 3D informa-
tion, since the telescope mount is computer controlled with
angular precisions of 1

to 3

and the distance can be deter-
Fig. 14 (a) The PRISMS system in the visible/NIR configuration for
remote imaging at distance greater than 3.5 m and (b) the modified
PRISMS for scanning of manuscripts
mined to millimetre precision using the focus position for
each scene. PRISMS is able to perform spectral imaging in
the VIS/NIR and SWIR, as well as measuring the 3D sur-
face texture of the object as a by-product of the high spatial
resolution imaging. The maximum distance range of opera-
tion depends on the intensity of the illumination. Figure 14
shows that the PRISMS system is flexible and can easily be
adapted to close range imaging of manuscripts by swapping
a lens with the telescope.
Laser induced fluorescence (LIF) has the clear advan-
tage of having a narrow excitation spectral range and a high
intensity and therefore remote imaging potential. LIF has
been used by a number of groups to remotely image build-
ings using a point by point scan to detect bio-degradation on
stones [94] and frescoes [87] and detecting past conserva-
tion interventions and identifying stone types [93].

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 321
7 Combined with other techniques
The combination of multispectral and hyperspectral imag-
ing in the visible and SWIR with other complementary
non-invasive or micro-destructive techniques where possi-
ble, holds the promise of uncovering more detailed informa-
tion about an object. The advantage of spectral imaging is
its speed of capture and therefore it is possible to scan large
areas.
7.1 Combined with traditional micro-destructive
techniques
In cases where detailed chemical information is needed, for
example the exact nature of the binding medium or pigment
identification where the paint has significantly degraded,
it is useful to complement the non-invasive imaging spec-
troscopy with traditional suite of micro-chemical analysis
using tiny samples from the object if it is possible to do so
(e.g. oil paintings). While the micro-destructive techniques
give detailed chemical information from isolated tiny sam-
ples which might not be representative of a painting, the
non-invasive spectral imaging technique can verify how rep-
resentative the point is by comparing the spectrum at the
point with other areas on the object [13, 91].
In addition, multispectral imaging combined with a mi-
croscope has been used to examine paint cross-sections to
give spectral reflectance information in addition to the con-
ventional colour information through visual examination of
a sample cross-section under a microscope [95].
7.2 Combined with laser scanning and photogrammetry
Any object can be visually reconstructed if both its 3D spa-
tial structure and spectral reflectance per point are charac-
terised. Even seemingly 2D objects like paintings have 3D
texture information. For wall painting, 3D mapping is par-
ticularly important because of the surface roughness and the
continuous motif that covers large areas of the interior of an
architectural structure. There has been a number of projects
where multispectral imaging has been combined with 3D
capture using laser scanning and photogrammetry [96, 97].
7.3 Combined with X-ray fluorescence (XRF)
X-ray fluorescence is a non-invasive, in situ point analy-
sis technique (in the portable versions) for the identifica-
tion of elements with atomic number higher than sodium. It
complements the large area scans of imaging spectroscopy.
While XRF is an elemental analysis technique, imaging
spectroscopy is sensitive to the chemical composition of the
compound (e.g. the paint mixture). As described in Sect. 5.4,
the pigment, the binding medium, the pigment particle size,
concentration can all affect the spectral reflectance of a
paint. On the other hand, XRF is able to detect not only the
main elemental composition but also trace elements which
can be useful to study the source of pigments. While the
complementary nature of the two non-invasive techniques
have been known for a long time, few studies have used the
two in combination [21, 32, 98, 99].
7.4 Combined with optical coherence tomography (OCT)
OCT is a relatively new non-invasive technique for the imag-
ing of subsurface microstructure of transparent and semi-
transparent material in the near infrared (see review on OCT
in this special issue). OCT gives not only the subsurface
layer structures but also reveals the scattering and absorption
properties of the material [61]. OCT complements spectral
imaging in the identification of the material and providing
information on the surface topology and subsurface layer
structure. A study combining OCT and multispectral imag-
ing was conducted by our group on an easel painting [66]
and a wall painting [100]. Combined spectral imaging and
OCT images of virtual cross-sections of paint layers has the
potential to match the information content given by micro-
scopic examinations of real sample cross-sections.
8 Future
Hardware development of multispectral and hyperspectral
imaging has reached maturity with many commercially
available instruments. However, it is important to realise that
the right instrument with the appropriate software must be
chosen for a specific application. An instrument designed for
microscopic hyperspectral imaging may not be suitable for
remote spectral imaging without hardware/software modi-
fication. It would be useful to extend the spectral imaging
capabilities to the mid-IR, the wavelength range tradition-
ally used for FTIR where more chemically specific infor-
mation can be obtained. Some very recent progress has been
made in this direction [101]. Given the heterogeneous nature
of heritage applications, it is best to build modular systems
that are versatile and flexible.
Future research should concentrate on further develop-
ment of material identification methods using spectra col-
lected with hyperspectral imaging systems that cover a
broad wavelength range. One critical ingredient is the de-
velopment of comprehensive databases of reference ma-
terial. More efficient methods of presenting the data are
needed, since spectral imaging is soon to become one of
the tools routinely used by the heritage community for the
recording and the scientific examination of artefacts. Most
of the applications so far have concentrated on paintings and
manuscripts. Applications on other types of object, espe-
cially intact objects where invasive methods are not allowed,

322 H. Liang
should be further explored. Closer collaborations with prac-
titioners from the heritage community is needed for spectral
imaging to solve real problems and hence further demon-
strate its relevance to art conservation and archaeology.
Acknowledgements Contributions from past and current research
fellows, students and technicians at Nottingham Trent University are
gratefully acknowledged: Rebecca Lange, Simon Godber, Tom Vaj-
zovic, Kafing Keita, Andrei Lucian, Borislava Peric, David Parker and
Sammy Cheung. Special thanks to Rebecca Lange for making last
minute illustrations. I would like to thank my collaborators John Cupitt,
Helen Howard, Sophie Julien-Lees, Sarah Neate, Chris Pannell, David
Saunders, Marika Spring, Jane Spooner, Jon Ward and Qunxi Zhang
for valuable contributions to the various spectral imaging projects.
Some of the work presented here was funded by UK Engineering and
Physical Sciences Research Council (EPSRC) grant (EP/E016227/1),
EU FP5 supported CRISATEL project (IST1999-20163), Royal So-
ciety Research Grant (2005–2006), Leverhulme Trust Project Grant
(2006–2009) and a Stimulating Innovation for Success grant from Not-
tingham Trent University (2006–2007).
References
1. A. Aldrovandi, D. Bertani, M. Cetica, M. Matteini, A. Moles, P.
Poggi, P. Tiano, Stud. Conserv. 33, 154 (1988)
2. M. Hain, J. Bartl, V. Jacko, Meas. Sci. Rev. 3, 9 (2003)
3. A. Goetz, G. Vane, J. Solomon, B. Rock, Science 228, 1147
(1985)
4. H. Grahn, P. Geladi (eds.), Techniques and Applications of Hy-
perspectral Image Analysis (Wiley, New York, 2007)
5. P. Yuen, M. Richardson, Imaging Sci. J. 58, 241 (2010)
6. C. Fischer, J. Kakoulli, Rev. Conserv. 7, 3 (2006)
7. M. Kubik, in Physical Techniques in the Study of Art, Archaeol-
ogy and Cultural Heritage, vol. II. (Amsterdam, 2007), p. 199
8. A. Burmester, J. Cupitt, H. Derrien, N. Dessipris, A. Hamber, K.
Martinez, M. Müller, D. Saunders, in 3rd International Confer-
ence on Non Destructive Testing, Microanalytical Methods and
Environmental Evaluation for Study and Conservation of Works
of Art, Rome, 1993, p. 201
9. D. Saunders, J. Cupitt, Natl. Gallery Tech. Bull. 14, 72 (1993)
10. K. Martinez, J. Cupitt, D. Saunders, Proc. SPIE 1901, 25 (1993)
11. C. Lahanier, G. Alquié, P. Cotte, C. Christofides, C. de Deyne, R.
Pillay, D. Saunders, F. Schmitt, in International Council of Mu-
seums, Committee for Conservation (ICOM-CC), 13th Triennial
Meeting Rio de Janeiro (James & James, London, 2002), p. 295
12. A. Ribés, H. Brettel, F. Schmitt, H. Liang, J. Cupitt, D. Saunders,
in Proceedings of the PICS Conference. The Digital Photography
Conference, IS&T, Springfield, VA, 2003, p. 215
13. H. Liang, D. Saunders, J. Cupitt, J. Imaging Sci. Technol. 49, 551
(2005)
14. G. Thomson, The Museum Environment, 2nd edn. Butterworth
Heine Series in Conservation and Museology (Butterworth, Ox-
ford, 1986)
15. D. Saunders, J. Kirby, in ICOM Committee for Conservation
11th Triennial Meeting Edinburgh (James & James, London,
1996), p. 87
16. A. Lerwill, J. Townsend, H. Liang, J. Thomas, S. Hackney, E-
Preservation, Science 5, 17 (2008)
17. H. Liang, R. Lange, A. Lucian, P. Hyndes, J. Townsend, S. Hack-
ney, in International Council of Museums, Committee for Con-
servation (ICOM-CC), 16th Triennial Conference Lison, Criterio
Produo Grica Lda., Almada, 2011
18. M. Klein, B. Aalderink, R. Padoan, G. de Bruin, T. Steemers,
Sensors 8, 4476 (2008)
19. L. Fauch, E. Nippolainen, V. Teplov, A.A. Kamshilin, Opt. Ex-
press 18, 23394 (2010)
20. S. Baronti, A. Casini, F. Lotti, S. Porcinai, Appl. Opt. 37, 1299
(1998)
21. J.K. Delaney, E. Walmsley, B.H. Berrie, C.F. Fletcher, in Sack-
ler NAS Colloquium Scientific Examination of Art: Modern Tech-
niques in Conservation and Analysis (National Academy of Sci-
ence, 2005), p. 120
22. N. Gat, Proc. SPIE 4056, 50 (2000)
23. J. Hardeberg, F. Schmitt, H. Brettel, Opt. Eng. 41, 2532 (2002)
24. R. Berns, L. Taplin, F. Imai, E. Day, D. Day, Stud. Conserv. 50,
253 (2005)
25. J. Mansfield, M. Attas, C. Majzels, C. Collins, E. Cloutis, H.
Mantsch, Vib. Spectrosc. 28, 59 (2002)
26. N. Gupta, Proc. SPIE 6940, 69401P-1 (2008)
27. C. Stedham, M. Draper, J. Ward, E. Wachman, C. Pannell, Proc.
SPIE 6889, 68891C (2008)
28. D.N. Stratis, K.L. Eland, J.C. Carter, S.J. Tomlinson, S.M. Angel,
Appl. Spectrosc. 55, 999 (2001)
29. J. Bland-Hawthorn, in Encyclopaedia of Astronomy & Astro-
physics (MacMillan and Institute of Physics Publishing, London,
2000)
30. W. Marinelli, C. Gittins, A. Gelb, B. Green, Appl. Opt. 38, 2594
(1999)
31. A. Casini, M. Bacci, C. Cucci, F. Lotti, S. Porcinai, M. Picollo,
B. Radicati, M. Poggesi, L. Stefani, Proc. SPIE 5857, 177 (2005)
32. J.K. Delaney, J.G. Zeibel, M. Thoury, R. Littleton, M. Palmer,
K.M. Morales, E.R. de la Rie, A. Hoenigswald, Appl. Spectrosc.
64, 584 (2010)
33. M. Picollo, M. Bacci, A. Casini, F. Lotti, M. Poggesi, L. Stefani,
in Conservation Science 2007, Milan (Archetype Publications,
2007), p. 162
34. J. Kerekes, J. Schott, in Hyperspectral Data Exploitation—
Theory and Applications (Wiley, New York, 2007), p. 20
35. A. Casini, F. Lotti, M. Picollo, L. Stefani, A. Aldrovandi, in Con-
servation Science 2002, Edinburgh, 2002, p. 248
36. P. Carcagni, A.D. Patria, R. Fontana, M. Greco, M. Mastroianni,
M. Materazzi, E. Pompaloni, L. Pezzati, Opt. Lasers Eng. 45, 360
(2007)
37. C. Bonifazzi, P. Carcagni, R. Fantana, M. Greco, M. Mastroianni,
M. Materazzi, E. Pampaloni, L. Pezzati, D. Bencini, J. Opt. A,
Pure Appl. Opt. 10, 064011 (2008)
38. C. Daffara, E. Pampaloni, L. Pezzati, M. Barucci, R. Fontanna,
Acc. Chem. Res. 43(6), 847 (2010)
39. J.M. Eichenholz, N. Barnett, Y. Juang, D. Fish, S. Spano, E.
Lindsley, D.L. Farkas, Proc. SPIE 7568, 75681L (2010)
40. R.S. Berns, L.A. Taplin, M. Nezamabadi, M. Mohammadi, Y.
Zhao, in ICOM Committee for Conservation, 14th Triennial
Meeting The Hague (James & James, London, 2005), p. 743
41. S. Howell, Handbook of CCD Astronomy, 2nd edn. (Cambridge
University Press, Cambridge, 2006)
42. B. Zitova, J. Flusser, Image Vis. Comput. 21, 977 (2003)
43. J. Cupitt, K. Martinez, Proc. SPIE 1663, 19 (1996)
44. D. Saunders, J. Cupitt, Natl. Gallery Tech. Bull. 16, 61 (1995)
45. R. Billinge, J. Cupitt, N. Dessipris, D. Saunders, Stud. Conserv.
38, 92 (1993)
46. J. van Asperen de Boer, Stud. Conserv. 14, 96 (1969)
47. C. Balas, V. Papadakis, N. Papadakis, A. Papadakis, E. Vaz-
giouraki, G. Themelis, J. Cult. Heritage 4, 330 (2003)
48. D.J. Walvoord, R.L. Easton Jr., IEEE Signal Process. Mag. 25(4),
100 (2008)
49. K. Knox, R. Easton, W. Christens-Barry, in 16th European Sig-
nal Processing Conference (EUSIPCO 2008), Lausanne, 2008,
on-line publication
50. S.J. Kim, F. Deng, M.S. Brown, Pattern Recognit. 44, 1461
(2011)

Advances in multispectral and hyperspectral imaging for archaeology and art conservation 323
51. CIE Colorimetry, 3rd edn. CIE Publication 015:2004 (Central
Bureau of the CIE, Vienna, 2004)
52. K. McLaren, J. Soc. Dyers Colour. 92, 338 (1976)
53. CIE Technical report: Improvement to Industrial Colour-
difference Evaluation, CIE Publication 142-2001 (Central Bu-
reau of the CIE, Vienna, 2001)
54. M.R. Luo, G. Cui, B. Rigg, Color Res. Appl. 26(5), 340 (2001)
55. H. Liang, D. Saunders, J. Cupitt, C. Lahanier, in Conservation of
Ancient Sites on the Silk Road: The Second International Con-
ference on the Conservation of Grotto Sites, DunHuang, 2004
(Getty Publications, Los Angeles, 2010), p. 267
56. K. Martinez, J. Cupitt, D. Saunders, R. Pillay, Proc. IEEE 90(1),
28 (2002)
57. P. Pouli, V. Zafiropulos, C. Balas, Y. Doganis, A. Galanos, J. Cult.
Heritage 4, 338 (2003)
58. V. Papadakis, A. Loukaiti, P. Pouli, J. Cult. Heritage 11, 325
(2010)
59. W. Kautek, S. Pentzien, D. Müller, K. Troschke, R. Teule, Proc.
SPIE 4402, 130 (2001)
60. M. Bacci, Sens. Actuators 29, 190 (1995)
61. H. Liang, B. Peric, M. Hughes, A. Podoleanu, M. Spring, D.
Saunders, Proc. SPIE 6618, 661805 (2007)
62. CNR-IFAC, Fibre Optics Reflectance Spectra Database of
Pictorial Materials in the 270 to 1700 nm Range (2011).
http://fors.ifac.cnr.it/
63. R. Johnston-Feller, Color Science in the Examination of Mu-
seum Objects—Nondestructive Procedures (Getty Publications,
Los Angeles, 2001)
64. J. Bowles, D. Gillis, in Hyperspectral Data Exploitation—
Theory and Applications (Wiley, New York, 2007), p. 77
65. J. Zhang, B. Rivard, A. Sanchez-Azofeifa, Remote Sens. Envi-
ron. 95, 57 (2005)
66. H. Liang, K. Keita, B. Peric, T. Vajzovic, in Proc. OSAV2008,
The 2nd Int. Topical Meeting on Optical Sensing and Artificial
Vision, St. Petersburg, 2008, p. 33
67. M. Elias, L. Simonot, M. Menu, Opt. Commun. 191, 1 (2001)
68. R.S. Berns, E.R. de la Rie, in ICOM Committee for Conservation,
13th Triennial Meeting Rio de Janeiro (James & James, London,
2002), p. 211
69. M. Elias, L. Simonot, M. Thoury, J. Frigerio, Opt. Commun. 231,
25 (2004)
70. A. Schuster, Astrophys. J. 21, 1 (1905)
71. S. Chandrasekhar, Radiative Transfer (Dover, New York, 1950)
72. P. Kubelka, F. Munk, Z. Tech. Phys. 12, 593 (1931)
73. P. Kubelka, J. Opt. Soc. Am. 38, 448 (1948)
74. J. Nobbs, Rev. Prog. Color. 15, 66 (1985)
75. W. Vargas, G. Niklasson, Appl. Opt. 36, 5580 (1997)
76. J.L. Saunderson, J. Opt. Soc. Am. 32, 727 (1942)
77. R. Berns, J. Kreuger, M. Swicklik, Stud. Conserv. 47, 46 (2002)
78. Y. Zhao, R. Berns, L. Taplin, J. Coddington, Proc. SPIE 6810,
681007 (2008)
79. G. Latour, M. Elias, J. Frigerio, Appl. Spectrosc. 63, 604 (2009)
80. L. Simonot, M. Elias, E. Charron, Appl. Opt. 43, 2580 (2004)
81. G. Latour, M. Elias, J.M. Frigerio, J. Opt. Soc. Am. A 24, 3045
(2007)
82. A. Roy, G. Mancini, Natl. Gallery Tech. Bull. 31, 64 (2010)
83. M. Baxter, Statistics in Archaeology (Arnold Applications of
Statistics, London, 2003)
84. A. Pelagotti, L. Pezzati, A. Piva, A. Mastio, in 14th European
Signal Processing Conference, Florence, Italy, 2006
85. D. Comelli, G. Valentini, A. Nevin, A. Farina, L. Toniolo, R.
Cubeddu, Rev. Sci. Instrum. 79, 086112 (2008)
86. A. Anglos, M. Solomidou, I. Zergioti, V. Zafiropulos, T. Papa-
zoglou, C. Fotakis, Appl. Spectrosc. 50, 1331 (1996)
87. F. Colao, R. Fantoni, L. Fiorani, A. Palucci, I. Gomoiu, J. Opto-
electron. Adv. Mater. 7, 3197 (2005)
88. E. Lange, Occas. Pap. - Br. Mus. 114, 1 (1996)
89. L. Macdonald, Proc. SPIE 6069, 60690I (2006)
90. H. Liang, K. Keita, T. Vajzovic, Proc. SPIE 6618, 661815 (2007)
91. H. Liang, K. Keita, T. Vajzovic, in ICOM Committee for Con-
servation, 15th Triennial Conference (Allied Publishing, New
Delhi, 2008), p. 353
92. H. Liang, K. Keita, C. Pannell, J. Ward, in IX CONGRESO NA-
CIONAL DEL COLOR, Alicante, Spain, 2010, p. 189
93. L. Palombi, D. Lognoli, V. Raimondi, G. Cecchi, J. Hällström, K.
Barup, C. Conti, R. Grönlund, A. Johansson, S. Svanberg, Opt.
Express 16, 6794 (2008)
94. D. Lognoli, G. Lamenti, L. Pantani, D. Tirelli, P. Tiano, L.
Tomaselli, Appl. Opt. 41, 1780 (2002)
95. J. van der Weerd, M. van Veen, R. Heeren, J. Boon, Anal. Chem.
75, 716 (2003)
96. N. Brusco, S. Capeleto, M. Fedel, A. Paviotti, L. Poletto, G.M.
Cortelazzo, G. Tondello, Mach. Vis. Appl. 17, 373 (2006)
97. L. Barazzetti, F. Remondino, M. Scaioni, M. Lo Brutto, A. Rizzi,
R. Brumana, in International Archives of Photogrammetry, Re-
mote Sensing and Spatial Information Science, Vol. XXXVIII,
Part 5, Commission V Symposium, 2010, p. 62
98. A. Casini, F. Lotti, M. Picollo, L. Stefani, E. Buzzegoli, Stud.
Conserv. 44, 39 (1999)
99. R. Lange, H. Liang, Q. Zhang, Proc. SPIE 8084, 80840Y (2011)
100. H. Liang, R. Lange, H. Howard, J. Spooner, Proc. SPIE 8084,
80840F (2011)
101. C. Daffara, L. Pezzati, D. Ambrosini, D. Paoletti, R. Di Biase,
P.I. Mariotti, C. Frosinini, Proc. SPIE 8084, 808406 (2011)

Copyright of Applied Physics A: Materials Science & Processing is the property of Springer Science &
Business Media B.V. and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print, download, or email
articles for individual use.