Analytical & Fusion
Technique Candidates
for MA-XRF Data
by Laurens van Giersbergen, MSc
Supervisor: Dr. K. Keune
i
Abstract
This thesis first presents an overview of the mechanisms that enable Macroscopic X-Ray Fluorescence (MA-XRF) scanning and outlines the development of XRF spec-trometers. The development of a commercial spectrometer allowed MA-XRF scanning to become a widespread technique for paintings research.Current research practice is moving towards increased collaboration between experts, combining various analytical techniques. An overview of the complimentary spectro-scopic techniques that are often employed in combination with MA-XRF is presented in this thesis. Especially hyperspectral imaging techniques show potential for combi-nation with MA-XRF, as both techniques produce data cubes of a scanned image. Because MA-XRF is a sensitive technique but only allows for elemental identification, while hyperspectral imaging is less sensitive but provides chemical information, the two techniques are highly complementary.
However, with steadily increasing amounts of collected data, there is a call for auto-mated integration of information from these various techniques. Data processing and data fusion techniques have therefore been developed or co-opted from other research disciplines to automate the integration process. Several of the most used and most promising data fusion techniques are reviewed and discussed in this thesis. Principal Component Analysis (PCA) is found to already be widely used as a tool to get quick results and allows for easy data exploration. Newer is t-Distributed Stochastic Neighbourhood Embedding (t-SNE), which has been identified as a technique that shows great promise as a slower but more informative and customisable data fusion and exploration technique, compared to the established PCA technique. Supervised systems that require training data such as Artificial Neural Networks (ANN) and many Endmember Extraction Algorithms (EEAs) are promising but currently suffer from a lack of standardised data and the limited availability of reference libraries.
Table of Contents
ii
1 Introduction
1
2 X-Ray Fluorescence
3
2.1
Primary X-rays
. . . .
3
2.2
Secondary X-rays
. . . .
5
2.3
Characteristic Fluorescence
. . . .
7
2.3.1 Shells & Holes
. . . .
7
2.3.2 Transitions
.. . . .
8
3 X-Ray Fluorescence Modalities
10
3.1
Detection Geometries
. . . .
10
3.1.1
Energy and Wavelength Dispersive XRF
. . . .
10
3.1.2 Total Reflection XRF (TXRF)
. . . .
11
3.2
Point-based XRF
. . . .
13
3.2.1 Benchtop-based XRF
. . . .
13
3.2.2 Handheld XRF
.. . . .
13
3.2.3 μ-SR-XRF
. . . .
14
3.2.4 SEM-EDX
.. . . .
16
3.3
Scanning XRF
. . . .
16
3.3.1 Scanning SR-XRF
. . . .
16
3.3.2 Laboratory Scanning MA-XRF
. . . .
18
3.4
Depth-Resolved XRF
.. . . .
21
3.4.1 Confocal XRF (CXRF)
.. . . .
21
3.4.2 Computed Tomography (CT)
. . . .
23
4 Complementary Techniques
24
4.1
Photography
. . . .
24
4.1.1
Visible Light Photography
. . . .
25
4.1.2 Ultraviolet (UV) Fluorescence Imaging
. . . .
25
4.1.3 Infrared Reflectography (IRR)
. . . .
26
4.1.4 X-Radiography
. . . .
26
4.2
Hyperspectral Imaging (HSI)
. . . .
27
4.2.1 Reflectance Imaging Spectroscopy (RIS)
. . . .
27
4.2.2 Fluorescence Imaging Spectroscopy (FIS)
. . . .
27
4.2.3 Visible-induced luminescence (VIL)
. . . .
28
4.3
Scanning techniques
. . . .
28
4.3.1 Macroscopic Fourier Transform Infrared (MA-FTIR)
. . . .
28
4.3.2 Macroscopic X-Ray Diffraction (MA-XRD)
. . . .
29
5 Data Processing
30
5.1
Spectral Pre-Processing
.. . . .
30
5.1.1
Normalisation
. . . .
31
5.1.2 Savitzky–Golay (SG) Filtering
. . . .
31
5.2
Background Removal
. . . .
32
5.3
MA-XRF Image Processing
. . . .
33
5.3.1 Region-of-Interest (ROI) Imaging
.. . . .
34
v
ii
5.3.3 Dynamic Analysis (DA)
. . . .
34
5.4
Hyperspectral Image Processing
. . . .
35
5.4.1 Database Matching
. . . .
36
5.4.2 Spectral Endmember Extraction Algorithms (EEAs)
.. . . .
37
5.4.3 Kubelka–Munk (KM) Theory
. . . .
39
5.5
Multivariate Curve Resolution (MCR)
. . . .
40
5.6
Wavelet Transform (WT) & Wavelet Prism (WP)
. . . .
41
6 Data Clustering & Fusion
43
6.1
Framework
. . . .
43
6.2
Low- and Mid-Level Data Fusion
. . . .
47
6.2.1 Co-Registration
. . . .
47
6.2.2 Principal Component Analysis (PCA)
. . . .
49
6.2.3 Partial Least Squares (PLS)
.. . . .
51
6.2.4 Stochastic Neighbour Embedding (SNE)
. . . .
53
6.2.5 Support Vector Machine (SVM) & Regression (SVR)
. . . .
54
6.2.6 Artificial Neural Network (ANN)
. . . .
55
6.3
High-Level Data Fusion
. . . .
59
6.3.1 Majority Voting
. . . .
59
6.3.2 Bayesian Network (BN)
. . . .
59
7 Discussion
61
7.1
Promising Complimentary Techniques
. . . .
61
7.1.1
Hyperspectral Imaging (HSI)
. . . .
62
7.1.2
Macroscale Scanning Spectroscopy
. . . .
63
7.2
Promising Data Fusion (DF) Techniques
. . . .
64
1
X-ray fluorescence (XRF) has become a much-used analytical technique to investigate canvas paintings due to its non-invasive nature and pigment iden-tification capabilities. Especially since the advent of commercial scanning systems, it has become an invaluable tool in paintings research and conser-vation1. To overcome part of the limitations of the technology, such as the lack
of chemical information, XRF has increasingly been used in combination with other analytical techniques. Those techniques are often based on different physicochemical processes and therefore produce complimentary information to XRF. However, the interpretation of the different techniques also requires a collaboration of a multitude of experts on the different processes.
In order to make these technologies accessible to more laboratories and to make the interpretation of these data more accessible to experts, automation and data fusion methods are being developed. Data fusion is the process of combining outputs from different analytical instruments or techniques. The aim of data fusion is to combine the data from complementary techniques to produce more accurate knowledge compared to when a single technique is employed or when multiple techniques are used in isolation2.
This thesis aims to provide an overview of the principles behind XRF and the various XRF measurement modalities that are in use for paintings research. Additionally, analytical techniques complimentary to XRF that are often used
in tandem with XRF are examined. Data processing regimes are evaluated and finally, from the data fusion techniques that are used with XRF data, recom-mendations are given to: (i) identify promising analytical techniques and (ii) search for promising data fusion methodologies to complement macroscopic
1 Introduction
2
XRF scanning in canvas paintings research. The focus has been put on canvas paintings to limit the scope, but applications in other areas of cultural heritage are also mentioned in passing, where relevant. Macroscopic XRF scanning is an XRF scanning technique that is widely used in the field. Therefore, this was chosen as the basis for data fusion comparisons.
2
In this chapter the creation of fluorescent X-rays is examined. First the crea-tion of primary X-rays is laid out. Then the interaccrea-tion of those primary X-rays with matter to form secondary X-rays is discussed and finally characteristic fluorescence radiation that is used for elemental determination is examined in more detail by looking at the atomic model that underpins X-ray fluorescence (XRF) theory and the transition that follow from it. This then serves as a base
for chapter 3, which elaborates on different XRF instrumentation.
2.1
Primary X-rays
Primary X-radiation is radiation emitted directly from a source. Primary X-rays are often emitted due to energetic particle collisions, e.g. protons or electrons. Primary radiation can be used directly for spectroscopy, using techniques such as particle-induced X-ray emission spectroscopy (PIXE)3,4, and energy
disper-sive (EDS or EDX, section 3.2.4), and wavelength dispersive (WDS) electron microscopy. Synchrotron radiation (SR, sections 3.2.3 and 3.3.1) and radioactive element decay emissions are also considered primary X-radiation, but they do not result from collisions.
Most commonly, primary X-radiation is created with X-ray tubes. The original X-ray tube, as developed by the discoverer of X-rays, Wilhelm Röntgen, was based on low pressure gas ionisation. Nowadays, filament-based tubes are most commonly used (Figure 1). A filament is heated by an electric current in a near-vacuum tube, resulting in electron cloud formation. The electrons are
2 X-Ray Fluorescence
4
then accelerated via a potential difference, until they are rapidly decelerated in a collision with a target material. The rapid deceleration results in the emis-sion of continuum X-rays known as ‘Bremsstrahlung’ and characteristic X-rays
(section 2.3) of the target material due to interaction with the target. The
emit-ted X-rays exit the tube through a window, usually made from beryllium, which can be placed at the end (Figure 1) or the side of the tube, depending on the tube geometry5. Primary X-rays Anode Target Electrons Potential difference Filament
Figure 1. Simplified design of an X-ray tube. The red filament is heated by electric current, which produces electrons (orange).
The electrons are then accelerated towards the anode (turquoise) by a strong potential difference. When the anode is hit by the electrons, the electrons are suddenly decelerated and produce X-rays (purple) as a result. The X-rays exit the tube through a – usually beryllium – window (dashed line).
Source: Figure based on Brouwer (2003)6
The anode target material can be made of different metals but high melt-ing point, specific heat, vapour pressure, and heat conduction are consid-ered. Good heat resistance properties are requires as the energy efficiency of X-ray production is generally around 1%, with the rest being converted to heat. Although the conversion efficiency depends linearly on the atomic number (Z)
and the applied voltage7. Common materials include W, Rh, and Mo and to a
lesser extent Ag, Cu, Pd, and Re but many more metals are occasionally used5,7,8.
The choice of anode material, selection of operating voltage (V) and current (I) depends on what type of analysis is done (Figure 2). Continuum radiation intensity (Ψ) increases linearly with atomic number and current but with the square of the voltage (Ψ ∝ Z · I · U²). So for a high continuum intensity, high atomic number anode materials should be used, such as W (Z=74) or Au (Z=77). Higher atomic number anodes also have higher energy characteristic radiation
(section 2.3) peaks. Because these peaks are narrow and high intensity, they can be preferable for measurements to continuum radiation. In that case, a lower atomic number anode should be chosen, such as Rh (Z=45) or Mo (Z=42)5,8.
60kV 50kV 40kV 30kV 20kV Wavelength (nm) Intensity (a.u.) (a) Wavelength (nm) Intensity (a.u.) 20mA 30mA 40mA 50mA 60mA (b) Wavelength (nm) Intensity (a.u.) Cr Z=24 Mo Z=42 Rh Z=45 W Z=74 Au Z=79 (c)
Figure 2. Plots showing the effect on the continuum intensity when varying different tube parameters. The characteristic
radi-ation that would be superimposed on these spectra has been omitted for clarity. (a) Square power increase in intensity due to
increasing the tube voltage, while keeping the current and anode material constant. (b) Linear effect of varying the tube current,
while keeping the voltage and anode material constant. (c) Linear effect of varying the anode target material, while keeping the
voltage and current constant.
2.2
Secondary X-rays
Secondary X-rays are any X-rays that are emitted as a result from primary X-ray absorption. As mentioned in the previous section (2.1), most X-rays are created using X-ray tubes. These X-rays are then directed onto a sample, after which they interact with the sample, alter the X-rays and produce secondary X-rays9.
When an electron interacts with X-rays, the radiation can (i) be scattered but not lose energy, called ‘Rayleigh scattering’ (ii) be scattered and lose some energy to the electron, called ‘Compton scattering’ or (iii) be absorbed and
2 X-Ray Fluorescence
6
potentially cause X-ray fluorescence (XRF)6. In XRF experiments the recorded
results show primary continuum and primary characteristic radiation that is scattered by the sample. Superimposed on this spectrum is the sample’s own secondary characteristic fluorescence (section 2.3) peaks5.
Reducing Rayleigh and Compton peaks is an analytical challenge that can improve the legibility of the XRF spectrum. Reducing these peaks can be achieved through instrumental design or data processing. However, these scat-ter peaks can also be utilised; Compton scatscat-ter is mostly a result of inscat-teractions with light elements and was used by Thurrowgood et al.10 to selectively produce
a clear image of the carbon-based (Z=6) canvas of Portrait of a Woman by Edgar Degas. Rayleigh scatter, instead, is produced more by interactions with heavy atoms and was used to visualise the lead-based ground. The scattered images combined with the XRF maps then revealed a different female subject beneath the painting surface that was later overpainted by Degas (Figure 3)10.
The good contrast carbon Compton map of the canvas could in the future perhaps be employed for canvas thread analysis (section 4.1.4), especially as they are automatically recorded when doing XRF analysis.
Figure 3. Portrait of a Woman by Edgar Degas, c.1876–80, canvas, 46.3 × 38.2 cm, National Gallery of Victoria, Melbourne. (a)
Visi-ble light image. The boxed region highlights the XRF scan area. (b) X-radiograph. The obscured portrait is rotated 180 degrees
relative to the upper portrait. The face and ear of the obscured sitter are the primary source of contrast. (c) Reflected infrared
image (detail). A partial outline of the obscured sitter’s face is indicated with a dotted line. The extensive use of highly infra-red-absorbing black paint in the final composition provides a limited view of the underlying figure.
Source: Thurrowgood et al. (2016)10
2.3
Characteristic Fluorescence
2.3.1
Shells & Holes
Terminology in the XRF-field is based on a simplified atomic model, called the Sommerfeld-Bohr model (Figure 4). It simplifies the atom as a positively charged nucleus made up of protons and neutrons, with negative electrons circling the nucleus in elliptical orbits at set distances. Electronic orbits that are close together are called ‘shells.’ The shells are given letters corresponding to their principal quantum number (n). The innermost shell (n=1), nearest to the nucleus, is called the K-shell, then followed by the L-shell (n=2), the M-shell (n=3), and so on. Every shell has subshells, numbered according to their energy level within the shell, starting with the lowest energy, e.g. LI, LII and LIII (also L1, L2 and L3). The L-shell has 3 sub-shells and the M-shell has 5. Every shell, with principal quantum number n, can have a maximum of 2⋅n2 electrons5,6.
Incident X-ray Fluorescent X-ray Ejected (photo) electron Nucleus Core hole K-shell L-shell
Figure 4. Diagram showing the creation of KL fluorescent X-rays. The incident X-ray (curved purple line) causes the ejection of a
K-shell electron (the photoelectron, purple), creating a core hole (dashed orange). Subsequently an L-shell electron (turquoise) fills the K-shell core hole, simultaneously emitting fluorescent X-radiation (turquoise curved line).
Another corollary from the energy dependent emission, requiring excitation energies higher than the ionisation energy, is that emission of high energy characteristic fluorescence can be suppressed if the excitation energy is low enough. This creates a possibility for more selective measurements5. The
2 X-Ray Fluorescence
8
lead L-lines from the extensive use of lead white paint can overshadow minor compounds of lower atomic number in a painting. Smieska et al.11 showed that
using excitation energies below the lead L-line showed a marked improvement in the detection of these lighter elements.
2.3.2
Transitions
According to the electric-dipole selection rules, the filling of a hole or ‘transi-tion’ can only occur from a higher shell (minimally n+1) and a subshell with Δℓ = ±1 and Δj = 0 or ±1, where ℓ is the ‘azimuthal quantum number’ and j the ‘total angular momentum quantum number’. These are called the ‘diagram transitions’
(Figure 5). Other transitions (‘satellite transitions’) that are ‘forbidden’ according
to these rules are still present but are less intense and often not visible.
There are two transition notations in use (Figure 5). The officially sanctioned International Union of Pure and Applied Chemistry (IUPAC) notation is struc-tured as [inner electron]-[outer electron], i.e. L3-M5. Although this is the official IUPAC notation since 199112 and is arguably more straightforward, the Siegbahn
notation stems from 191613 and due to its long history of use is still widespread
in literature.
The Siegbahn notation is structured as [inner electron shell letter][Greek alpha-betical order][subscript], e.g. Lα1. The Greek letters are assigned to transitions
on a descending order of intensity. If a transition between the same n- and ℓ-levels but at a different energy level occurs, these transitions are called ‘multiplets’ and are given Arabic numeral subscripts, again, in descending order of intensity. Notably, a specific transition (e.g. L3-M5) always has fixed, corre-sponding Siegbahn notation (e.g. Lα1), regardless of the atom. As a result, the
intensity ordering is not always preserved for atomic numbers (Z) far away from the element in which the transition was first observed and after which the transition was named5.
K L I II III M I II III IVV Lβ4 L1M2 Lβ3 L1M3 Lη L2M1 Lβ1 L2M4 Lι L3M1 Lα2 L3M4 Lα1 L3M5 3 n j 5/2 2 3 2 3/2 3 1 3 1 3/2 1/2 3 0 1/2 1 1 3/2 1/2 0 1/2 0 1/2 2 2 2 1 l Kβ3 KM2 Kβ1 KM3 Kα2 KL2 Kα1 KL3 K series L series L1 L2 L3 Energy
Figure 5. Transition diagram showing the transitions between the K, L, and M shells. Below each transition, its name is stated
in the Siegbahn notation in italics and the IUPAC notation is below it. The shell sub-levels are indicated, as well as their associ-ated n, ℓ, and j quantum numbers. The diagram illustrates that the order of the Greek alphabetic and roman numeral characters are not evident from the diagram in the Siegbahn notation. It should be noted that some numbers appear to be missing, this is because the N or O shell transitions are sometimes of higher intensity than their LM-transition but they are not depicted here. The energy differences are not to scale.
3
This chapter will outline the various types of XRF instrumentation to get an overview of the various types of measurement possibilities, allowing for the contextualisation of the macroscopic XRF (MA-XRF) scanning technique within the XRF field. Over the last two decades, many advances have been made in the development of XRF instruments. The innovation led to more extensive use by the research community, which in turn sparked innovation. As a result, some new techniques have become widely adopted, whereas some classical techniques have been all but abandoned. This chapter will present the different techniques and their historic and current usage in the field. As such, it provides an insight into the type of XRF data that can produced, to later combine it with other analytical techniques using data fusion.
3.1
Detection Geometries
3.1.1
Energy and Wavelength Dispersive XRF
There are two types of XRF spectrometers: energy dispersive XRF (ED-XRF) and wavelength dispersive XRF (WD-XRF) types. In ED-XRF the entire range of energies is recorded at once, whereas WD-XRF uses a diffraction crystal with a collimator to scan over the energies one at a time by rotating the crystal or detector to capture Bragg reflections6.
ED-XRF machines are generally more affordable and are faster due to the simul-taneous recording of all wavelengths. However, the detection of lower atomic
number elements (below Na, Z=11) is normally not feasible and overall light element limits of detection are poor. WD-XRF spectrometers can detect lower atomic number elements (from Be, Z=4) and produce a higher spectral reso-lution, especially for the lighter elements14. However, the technique is generally
more costly, slower and the use of moving parts (rotating crystal or detector) is a possible point of failure6.
WD-XRF is not often used in cultural heritage research but Gil et al.15 showed
the use of a WD-XRF spectrometer on limewashing paintings. Taking advantage of its excellent sensitivity, WD-XRF was used to perform semi-quantitative analysis of inorganic pigment makeup. It allowed them to detect impurities in the lime washing, which granted an insight into the manufacturing process. Kaszowska et al.16 used SEM-WDX mapping to investigate whether Ba, Zn, and
S were co-located or not, to prove the presence of lithopone (BaSO4 + ZnS), as opposed to a mixture of barium white (BaSO4) and zinc white (ZnO). Holakooei et al.17 used the high spectral resolution of WD-XRF to produce quantitative
information on Persian tiles. Combined with a PCA analysis (section 6.2.2), this allowed them to elucidate that Safavid tile manufacturing was carried out locally. All other XRF spectroscopy mentioned from here is ED-XRF based, unless
specified otherwise.
3.1.2
Total Reflection XRF (TXRF)
Total reflection XRF (TXRF) can be either ED-XRF or WD-XRF but has a divergent excitation geometry from standard instrumentation. Conventional XRF is done at a primary beam incident angle above the total reflection angle (also ‘Bragg angle’). Instead, TXRF (Figure 6) employs an incident angle close to or below the Bragg angle with a flat (mean roughness <1 nm, flatness: λ/20) reflector surface upon which the sample is placed18.
Virtually all X-rays are totally reflected off the reflector (passing through the sample) but XRF signal is detected on the reflector surface normal. This results in a signal enhancement as fluorescence photons are virtually exclusively recorded and a large solid angle for detection combined is achieved due to close detector proximity. Picogram detection limits are reached using X-ray tube excitation and femtogram limits are feasible using synchrotron radiation (SR)18.
3 X-Ray Fluorescence Modalities
12
Reflector
Detector
Incident
X-ray ReflectedX-ray
Sample
Fluorescent X-rays θ
Figure 6. Schematic representation of a TXRF spectrometer. The primary X-rays (purple) pass through the sample (blue) and
reflect totally off the reflector, due to the incident angle (θ) being below the Bragg angle. As the primary X-rays pass through the sample, fluorescent X-rays (green) are produced. They are emitted in all directions, with the rays moving towards the detec-tor pictured in the figure. Due to the smoothness of the reflecdetec-tor, virtually no primary radiation is scattered and they are thus barely detected. Because the incident angle is low, the detector can also be close to the sample, resulting in a large solid angle of detection.
Only small sample amounts are required to produce a usable signal due to the high sensitivity. However, TXRF spectrometers cannot measure samples of >10 μm thickness due to auto-absorption at larger thicknesses. TXRF spectrometers are also often prohibitively priced19. For these reasons, the technique cannot
be readily used in paintings mapping.
In 2000 Klockenkämper et al.20 presented a procedure for micro-invasive TXRF,
allowing for elemental characterisation of pigments by swabbing the surface with a cotton swab and depositing the material on a carrier. The technique does require minimally invasive sampling by cotton swab and can exclusively be done on paintings without varnish – potentially during conservation treat-ment. However, due to the small sample, no matrix effects occur and reliable semi-quantitative analysis is possible, with limited sample processing required20.
The laid-out procedure is possibly applicable for analysis during conservation treatments, when pigment filled cotton swabs are produced as a by-product of varnish removal. Nonetheless, the approach does not seem to have garnered mainstream use. Presumably due to the advent of non-invasive XRF modalities and because paintings conservation generally does not benefit from the lower limits of detection.
Molari and Appoloni21 do include TXRF results alongside handheld XRF (section
3.2.2) in 2019, but they do not make any significant additional discoveries by
including the TXRF data that could not be made with the handheld spectrom-etry alone.
3.2
Point-based XRF
3.2.1
Benchtop-based XRF
The most straightforward form of XRF analysis is a point-based measurement. It forms the basis upon which the other technologies are built and is still in use today on its own merit. Conventional point-based XRF (point sample, station-ary, benchtop XRF) still has its place in the laboratory as it generally produces better spectra than handheld instruments (section 3.2.2), it is faster and more affordable than MA-XRF scanning (section 3.3.2) and more accessible than synchrotron facilities (section 3.2.3). The technique is thus in a solid middle ground to be of use.
Benchtop measurements are still done for quick point scans of paintings of lesser importance in the conservation process but has mostly been replaced in published articles since the beginning of the 2010s by either the hand-held spectrometer for flexibility or the MA-XRF scan to a more representative sampling. In other areas of cultural heritage, such as glass or metals research, the technology is still seeing much use19,22–25. In these fields sample-taking is
more accepted, a non-flat geometry complicates the use of MA-XRF, or the sample is approximated as having a nearly uniform elemental distribution, making the use of MA-XRF superfluous. It has been used for brick provenance determinations19, analysis of silver coins22, investigations of Late Bronze Age
gilding technology24, and forensic glass analysis25 to name a few of the
appli-cations for benchtop XRF in cultural heritage.
3.2.2
Handheld XRF
Portable XRF (PXRF) systems were developed to allow for more flexible anal-ysis. Cultural heritage objects can often not be moved as it may damage the object. The object may be too heavy, too large, or too fragile to move into a laboratory, or a museum may not want to take a major work off display. Some of the benchtop instruments can also be considered portable as they can be
3 X-Ray Fluorescence Modalities
14
packaged and moved. However, in this section, only handheld spectrometers are discussed.
Because of the space constraint in portable devices, sacrifices in performance are made. As a result, performance is generally inaccurate and unreliable. Brand and Brand26 showed that measurements with the same model of
spectrome-ter produced different results on the same sample and even the same spec-trometer produced a fluctuating spectrum over time. They state that every machine produces unique data and should thus not be combined with another spectrometer’s data without careful processing or calibration. Despite these shortcomings, they highlight the benefits of instant results, which can take some number of days with other spectrometer designs and which need to move a cultural heritage object. Although peak heights are not completely consistent, the sensitivity is generally good enough to detect trace elements, especially in samples containing few elements.
Newer silicon drift detector (SDD)-based spectrometers allow for low atomic number analysis (down to magnesium, Z=12)26. Handheld devices are suitable
for the analysis of homogeneous samples27 with their significantly larger spot
size (~5 mm2) compared to conventional instruments (~25 μm2)28. As such, the
technique is often employed in (largely uniform) sculpture29–31, ceramics23,32
and metal analysis22. It is also a valuable tool for mural painting analysis, as
these ‘samples’ cannot be transported to the laboratory. An investigation into Pompeiian Roman mural paintings showed the capability of XRF to differen-tiate between different types of red ochre pigments, some of which used to be yellow in colour33. Nevertheless, handheld XRF is used in canvas paintings
research as well, when a higher resolution or scanning machine is not available or instantaneous results are preferred34–36.
3.2.3
μ-SR-XRF
Synchrotrons are large circular electron accelerators that produce high brilliance, coherence and intensity synchrotron radiation (SR), exploiting the radiation emission of accelerating electrons37. Besides these preferable beam
charac-teristics, beamline optics and detectors are often of higher quality compared to laboratory X-ray instrumentation, as expensive optics and detectors are a more worthwhile investment in a multi-million or billion euro synchrotron facility that is used nearly round the clock. However, the beamtime available at facili-ties is often limited and samples are not always permitted to travel, limiting the use of SR techniques. Therefore SR-XRF is especially suited for high spatial or spectral resolutions or high excitation energy spectroscopy.
μ-SR-XRF spot measurements were carried out up until the end of the 2000s to do high-resolution pigment and paintings cross-sections studies20. However,
these point-based measurements have now been completely replaced by sufficient resolution laboratory (section 3.2.1) and portable (section 3.2.2) XRF measurements or with fast enough μ-SR-XRF scanning methods (section 3.3.1)38.
3.2.4
SEM-EDX
Energy dispersive X-ray spectroscopy (EDX or EDS) can be done in combination with scanning electron microscopy (SEM). The microscope provides excitation electrons for imaging, by definition, which can induce XRF photon emission in the sample that can be detected. A high spatial resolution, primary radiation XRF spectrum is obtained as a result. Due to the vacuum conditions inside the microscope, sample size is limited, however. Therefore, SEM-EDX’s main use in paintings research is the investigation of pigments in painting cross-sec-tions39–41. Raven et al.42 used SEM-EDX in the study of the formation of zinc soaps.
The localisation of zinc within different cross-sections in the SEM-EDX images showed that zinc soaps had formed at the interface of delaminating paint layers.
3.3
Scanning XRF
The XRF technique that has garnered the most use over the last decade is scanning macroscopic XRF (macro XRF or MA-XRF). This scanning technique produces hyperspectral XRF images by taking full spectral measurements at every pixel, with the detector and beam scanning over a painting. These images show elemental distributions over the entire scanned area, making the identi-fication of pigments based on elemental composition easier and more robust.
3.3.1
Scanning SR-XRF
Scanning MA-XRF started with scanning SR-XRF of an overpainted composition in a painting by Vincent van Gogh in 2008 by Dik et al.43 (Figure 7). The
scan-ning was done at 500 μm resolution in a 17.5 × 17.5 cm region. It took around 2 days to scan at 2 s dwell times. The proof of concept showed the power of the technique, but also highlights why it was not routinely done at the time due to the prohibitive analysis time. Improved sources, detectors and data process-ing regimes made XRF scannprocess-ing more feasible and have now made it a popular research technique44. The dwell time can be as short as about 3.5 ms, an almost
3 X-Ray Fluorescence Modalities
16
Figure 7. Van Gogh’s “Patch of Grass” scanned using MA-XRF. The MA-XRF elemental maps in black-and-white are overlayed
on the visible light colour image and show an underlaying compostition. The SR-based XRF images have red outline around the elemental maps. The black-and-white regions of (a) and (b) show antimony (Sb) levels and (c) and (d) the mercury (Hg) levels.
The comparison between the left ((a) and (c)) and right ((b) and (d)) images was shown by Alfeld et al.45 to illustrate that the
developped laboratory scan (in 2011) produced sufficient quality results compared to the original SR scan from 2008 to be a viable replacement for routine XRF scanning.
Source: Alfeld et al. (2011)45.
Although μ-SR-XRF point measurements (section 3.2.3) have declined in popularity over the last decade, micro-scale SR-XRF scanning is an oft-used technique to study pigment particle distribution and degradation in paint-ing cross-sections38,46,47. Large improvements were enabled in 2010 by the
development of a large array of detectors in backscatter geometry, called the Maia detector48 and especially by the version using 384 solid drift detectors
(SDDs) in tandem49. The Maia detector is capable of high-resolution
SR-MA-XRF mapping10,11,50,51 and simultaneous XRF and X-ray absorption near-edge
structure (XANES) mapping on microscale samples46,47,52. Although the Maia
detector started as a SR-only detector, its innovative design has started to be implemented in laboratory equipment53.
The development of laboratory-sized scanning MA-XRF spectrometers (section
3.3.2) has lowered the need for SR-XRF scanning. Nonetheless, SR-based
scan-ning XRF still has its place as a high resolution technique50. In their 2013 review
of SR-based pigment characterisation Janssens et al.38 mention that XRF is
often not the end goal of SR research nowadays but the combination with X-ray diffraction (XRD, section 4.3.2) and X-ray absorption spectroscopy (XAS) is valued and recording these three modalities at once has become standard procedure at SR facilities.
3.3.2
Laboratory Scanning MA-XRF
The pioneering use of SR-XRF mapping in the late 2000s43 and early 2010s45
lead to the development of commercial scanning MA-XRF systems such as the Bruker M6 Jetstream1 and made MA-XRF a mainstay in paintings research.
It has become one of the most valued tools in conservation (research) prac-tices. Due to the near-planar geometry of paintings and the two-dimensional nature of MA-XRF scanning, the technique is a natural way of investigating paintings. It overcomes the issue of potentially anomalous results that could result from a point-measurement on objects as heterogeneous as paintings. With the advent of the laboratory scanning MA-XRF machines, the potentially harmful and involved transportation of artworks to synchrotron facilities is no longer necessary either and many paintings have been subjected to MA-XRF investigation as a result. Using MA-XRF scanning new pigment use in the seven-teenth century palette was discovered54, pentimenti have been detected
under the painted surface1,55,56, entire overpainted compositions have been
visualised10,43,57,58, and features that had become invisible due to the ageing of
a painting could be retrieved59.
MA-XRF scanning has matured with the development of the commercial system. The yearly review of XRF papers in the atomic spectroscopy update in the
Jour-nal of AJour-nalytical Atomic Spectroscopy mentions the shift from fundamental and instrumentation development XRF papers to more application-focused ones since 201760–62. There has nonetheless been development in XRF
instru-mentation over the past decade. Most in-house developed MA-XRF scanners have brought marginal improvements over the M6 Jetstream but the system developed by the LANDIS laboratory in Catania, Italy brings many interesting innovations to the table (Table 1).
3 X-Ray Fluorescence Modalities
18
Table 1. Comparison between the most commonly used Bruker M6 Jetstream MA-XRF scanner and a new scanner developpend
by the LANDIS team. Some values are estimated. No rights can be derived from this table.
Bruker M6 Jetstream
1LANDIS scanner
63Scanning dimensions 80 × 60 cm2 110 × 70 × 20 cm2
Z-translation No Yes
Real time visualisation Per line, ROI-based Per pixel, PLS-based
Positioning Inductive proximity switch Wire sensor
Standard spot size 500 μm 500 μm
Standard spot dwell time 10 ms 5 ms
Standard scanning speed 5 mm/s 100 mm/s
μ-XRF pixel size 50 μm 25 μm
μ-XRF dwell time 10 ms 10 ms
The LANDIS system stage is capable of scanning a large 110 × 70 cm2 area (M6
Jetstream: 80 × 60 cm2) at a speed of 100 mm/s (M6: ~5 mm/s), when carrying
out MA-XRF measurements with a 500 μm spot size and fast 5 ms dwell times. μ-XRF scanning is also possible at a pixel size as low as 25 μm (M6: 50 μm) at a dwell time of 25 ms. Additionally, the scanning head position is determined by absolute wire sensor positioning, where previous machines used relative inductive proximity switches that must be recalibrated when the system is interrupted. Also novel is the system’s laser-controlled, motorised Z-axis that can automatically move the measurement head to maintain a consistent meas-urement distance on a non-flat surface, like that of a warped panel-painting63.
A development into a different direction was taken by Walter et al.64. They
adapted an XRF spectrometer design marked to go onto a NASA Mars rover to create a full-field XRF (FF-XRF) imager. The apparatus creates 13.3 × 13.3 mm2 area XRF image using a deep-depleted CCD camera that detects photon
energies as pixel intensity of short periods of time, that are then integrated. So it creates an image of the entire area at once, instead of scanning point for point. Although no clear figures of merit are presented, a rough calculation of the system’s speed puts it in the same order of magnitude as the LANDIS scanner. The authors remark that the omission of a scanning stage makes the machine more portable, as it can be mounted on a standard camera tripod. With further development and using this machine with a scanning stage, perhaps XRF scanning could be done even faster at 100 μm resolution. It is nonetheless interesting to see a different approach being used and the portability of the system may already make it useful for quick scans of smaller areas without needing any more development.
3.4
Depth-Resolved XRF
Although paintings are largely in one plane, they still consist of many layers stacked on top of each other. This layer build-up is essential for the appearance of the painting and knowing the paint stratigraphy is a must for conservation work. Nowadays, microscopic samples are taken to obtain cross-sections to study painting stratigraphy. However, the goal is to be able to do this charac-terisation non-invasively someday. As such, interest has grown in technologies that allow for the differentiation between pigments at different depths. This section will examine how confocal XRF (CXRF) and computer tomography (CT) XRF can provide depth-resolved XRF information for paintings.
3.4.1
Confocal XRF (CXRF)
Confocal spectrometry works by focusing the excitation beam to excite a small volume. Then a half-lens before the detector selects an even smaller sub-region of the excited region, thus selecting a minute volume of the sample
(Figure 8). The detection volume can then be scanned through the sample to
get depth-dependent XRF information. Because only a tiny volume can be in focus for the depth profiling to work, CXRF is generally done using SR, where ~30 μm depth resolution is generally achieved65–69. Nonetheless, laboratory confocal
setups have been developed and showed that depth profiling was of sufficient resolution (~50 μm) to resolve some different paint layers70–72, although
reso-lution is not high enough for measuring exact layer thicknesses yet, which can be as thin as ~15 μm73.
The adoption of the technique, on the road to the eventual replacement of cross-sections, has started to pick up in the last decade. Institutes such as the Centre de recherche et de restauration des musées de France (C2RMF)74,75 and
the National Gallery in Prague76 have now outfitted their laboratory with a CXRF
set up. For a list of cultural heritage CXRF spectrometers and a comparison between these setups, see the review by Laclavetine et al. (article in French)77.
Due to the inherently low collection efficiencies of confocal setups and the lower brilliancy of tube x-ray sources, dwell times in laboratory setups are often high (100 s)71,75 so mapping can only be done on small regions-of-interest
(ROIs)70,71. As such, CXRF mapping experiments are still uncommon compared
to point-based measurements but advances in technology are promising for CXRF scanning78,79. Due to the long exposure, concerns about X-ray dosage on
the painting have also been raised. With many research technologies that use X-radiation, these concerns have grown in general80.
3 X-Ray Fluorescence Modalities
20
Polycapillary full-lens Polycapillary half-lens Detector Multi-layered sample Analysed volume X-ray sourceFigure 8. Schematic representation of the workings of confocal XRF. The purple incident X-ray beam comes from the X-ray
source and is focused by a polycapillary full-lens into the sample. In front of the detector is a polycapillary half-lens (or a poly-capillary conical collimators) that focuses the detected fluorescent X-rays (turquoise). The focal points of the polypoly-capillary full- and half-lenses intersect in a small volume (yellow) within the sample (salmon), ensuring that only that small volume is detected. By moving the analysed volume through the sample, measurements can be done in three axes.
XRF depth resolution is always dependent on fluorescent energy, due to the different penetration depths of different energies. On top of that, low energy photons are especially poorly transmitted compared to high energy photons when using polycapillary lenses in CXRF equipment. Care should thus be taken to correct for photon energy when interpreting the depth profile in CXRF data, especially when using polycapillary lenses instead of polycapillary conical collimators (poly-CCCs).
A new germanium collimating channel array (CCA) enabled 2 μm depth resolu-tion using SR with virtually energy-independent detecresolu-tion81,82. This would likely
be infeasible for mapping purposes as a result of prohibitive measurement times but has the potential for point-based, true painting stratigraphy deter-mination.
Another approach to depth-profiling was outlined by Bártová et al.35 from the
Kβ-lines. Depth was modelled using the lines’ ratio and a Monte Carlo simulation of the XRF travel depth given the detected elements. This approach requires no additional spectrometer complexity, like CXRF, but the depth-modelling is not nearly as straightforward.
3.4.2
Computed Tomography (CT)
Computed tomography is the process of imaging an object by rotating it through radiation and then reconstructing a three-dimensional image from that data using a computer. X-ray tomography using X-ray absorption is a widely used technique in the medical field and has been applied to cultural heritage successfully. However, it has mostly been applied to objects that have more depth, such as sculpture or pottery83,84, as the resolution required for discerning
paint layers is too high. Nevertheless, it can provide useful information on the state of a wooden panel that a panel painting is painted on83.
To combat this partially, X-ray laminography was developed. It is a sub-tech-nique of tomography where a virtually flat object is angled with respect to the beam and the detector. In its tilted orientation, the object is rotated so as to not have the flat object’s long edge obscuring the detector85–87. The
tech-nique promises a 2 μm spatial depth resolution, which enables layer thickness determination and detailed painting studies without requiring cross-sections87.
However, the technique provides only minimal atomic information, in the form of pigment radio-opacity.
XRF CT is based on the same principle of sample rotation as X-ray (absorption) CT, but records fluorescence instead of absorbance. The technique has mostly been employed using the high-energy, well-defined beams of SR facilities67,88–90.
However, Laforce et al.73,91 present an integrated μ-CT, CXRF, XRF-CT benchtop
setup that allows small samples to be imaged in high detail using μ-CT and provides lower resolution elemental data via one of the other two techniques. XRF CT has not been applied to cultural heritage objects. Its use is presumably hampered by the small sample volume that can be detected, resulting in low scanning speeds with extensive use of ionising radiation. Nonetheless, imaging of cross-sections could give insight into the distribution of particles within the sample using XRF CT. However, it is possible that the highly absorbing elements in artist paints inhibit proper detection, making XRF CT unsuitable for paint-ings research. Signal loss of lower energy photons due to auto-absorption was already found to problematically decrease signal when analysing automotive paint samples using SR89.
4
Not only has XRF technology made steady progress over the last two decades, it has also become more common to investigate paintings using a multi-analytical approach. Many spectroscopic techniques have been used in tandem with XRF during research campaigns, as combining complementary techniques allows for discoveries that would not be possible if either technique was employed alone. Because these techniques have been elected for use by experts in the field, it follows that eventual data fusion on these techniques will be the most useful. To be able to evaluate their place in data fusion with XRF, it is good to have a basic understanding of these techniques and the data they produce. The most often combined techniques are presented in this chapter. Because the focus of this thesis is on MA-XRF, only the most frequently combined imaging or scanning methods will be elaborated on.
4.1
Photography
‘Photographic’ techniques are generally considered to be those techniques that (i) create a full field image without scanning and (ii) produce images with a low spectral resolution (1-3 bands). In contrast, ‘hyperspectral imaging’ techniques
(section 4.2) produce images with a high spectral resolution, using many
differ-ent wavelength bands and are more often created using a scanning approach.
Visible light, raking light, and UV fluorescent photography are among the most well-established photographic techniques. Many of these photographic tech-niques are the oldest tools in the field, whereas hyperspectral imaging (HSI) techniques are currently in development. Photography in its various forms is standard practice because the machines (often simply commercial cameras) are much more accessible than spectrometers. Additionally, their use is often more geared towards documentation than a way to gain added information.
4.1.1
Visible Light Photography
Visible light photography is the type of photography that is commonplace not just in research but in daily life. It generally records three colour bands (red, green and blue) and is valued as an inspection and documentation tool, as the resulting image comes close to the natural observation of an object with human eyes.
Therefore, high resolution photographs are generally the first step in cultural heritage research. It allows for inspection up close, by zooming in, as well as an overview if the image is of sufficient resolution. This makes planning research and conservation easier. On top of that, photographs provide an anchor for the contextualisation of more sophisticated analytical results92 and serve as
documentation of the ‘before’ state if alterations, such as sample taking or conservation treatments, are undertaken93,94.
4.1.2
Ultraviolet (UV) Fluorescence Imaging
UV-induced fluorescence photography is carried out by shining ultraviolet (UV) light onto a painting and recording its visible light fluorescence using a conven-tional camera at a long exposure. Care is taken to not cut off the camera above the wavelength of the emitting light source, as this would drown out the dim fluorescence. However, this is often already the case in standard cameras as normal photographs also should not record UV light. A filter to reduce some blue light is occasionally employed.
Several types and ages of varnish and many pigments will fluoresce differently. This makes identification of restoration work and separation of a select number
4 Complementary Techniques
24
The drawback is that it does not provide as much information a its hyperspectral equivalent: FIS (section 4.2.2). Moreover, the calibration of UV photography is an ongoing field of research, which makes pigment identification on the basis of visible and UV fluorescence photography alone slightly subjective. Nonethe-less, fluorescent photography is expected to be used alongside hyperspectral imaging for a long time due to the nearly century long history of usage as an established technique known by many conservators, short acquisition times and the inexpensive instrumentation94,95.
4.1.3
Infrared Reflectography (IRR)
Infrared reflectography (IRR), introduced by J.R.J. van Asperen de Boer, allows for the visualisation of carbon underdrawings using infrared (IR) radiation96.
Special cameras are used that image in the infrared, which makes the technique more costly than visible and UV-induced fluorescence photography. Nonethe-less, it is often still cheaper than a hyperspectral IR imaging spectrometer. As such, it has also become a much-used tool in the conservator’s toolkit and is therefore slated to stay in use for a long time, especially for smaller museum conservation studios.
4.1.4
X-Radiography
X-radiographs are photographs made using X-radiation. The technique is widely used to image heavy-element paints (most importantly; lead white) and to allow a researcher to see through a painting’s stratigraphy. On top of that, with enough resolution and data processing, it has been shown that the imprint of the canvas in lead white grounds allows for canvas weave analysis. Canvas weave patterns are matched between paintings and it becomes possi-ble to determine whether paintings come from the same canvas roll. This can contribute to a painting’s positioning within an artist’s creation timeline and contributes towards authentication efforts97–99.
Because X-ray sources are generally expensive and operators are required to follow an X-ray safety training, this technique is employed less than the other three discussed above. It is also more common to hire an outside specialist to do this imaging, as the information gained is valuable but a museum may not employ it enough to warrant the abovementioned investments.
4.2
Hyperspectral Imaging (HSI)
Hyperspectral imaging (HSI) techniques are somewhat like photography, as they are also generally full field (i.e. every pixel is captured simultaneously, as opposed to scanning per pixel). However, HSI captures many more spectral bands than the one to three in photography. As such, every pixel has a full spec-trum associated with it, allowing for more thorough investigation of every pixel.
4.2.1
Reflectance Imaging Spectroscopy (RIS)
Hyperspectral visible light reflectance imaging spectroscopy (RIS) has been developed to extract more information from light capture. Although it is gener-ally not considered this way, conventional photography is in essence spectro-scopic imaging with few wavelength bands. Because there is more information in the light at the visible light range about a compound than three bands, RIS records smaller wavelength bands. In doing so, RIS enables the discrimination between compounds based on the spectrum gained from reflectance of visi-ble light. Many RIS imaging spectrometers record the visivisi-ble-to-near-infra- visible-to-near-infra-red (VNIR) range, extending the measurement slightly into the near infravisible-to-near-infra-red (NIR)100–102.
There are also true IR RIS hyperspectral imaging systems, which operate in the NIR and short-wavelength IR (SWIR). The technology has shown to be selective and sensitive103,104. The hyperspectral systems produce images with as good, if
not often better, contrast than IRR instruments (section 4.1.3)102. Additionally, the
increased spectral width and resolution allow for spectroscopic identification of pigments and binding media100,101,104,105.
4.2.2
Fluorescence Imaging Spectroscopy (FIS)
Fluorescence imaging spectroscopy (FIS) HSI is the hyperspectral variant of UV-induced fluorescence photography. FIS generally uses the same imaging devices as VNIR HSI (section 4.2.1) but with UV lamps for illumination, akin to UV fluorescence photography. As such, it is a cost-effective technology to employ alongside VNIR RIS.
FIS again provides more spectral information than UV fluorescence photogra-phy. This extra information can be used for more rigorous interrogation of the data and makes red lake discrimination possible100, for example.
4 Complementary Techniques
26
4.2.3
Visible-induced luminescence (VIL)
Visible-induced luminescence (VIL) is a technique similar to FIS (section 4.2.2) as luminescence is induced by light. Instead of ultraviolet light, visible light is used for excitation. The visible excitation light induces infrared luminescence that is often detected using modified commercial cameras with the appropriate filters to remove the visible excitation light. Some pigments are incredibly sensi-tive to this technique. It has been used to identify minute amounts of Egyptian blue on visually clean Egyptian stonework106,107 and in Byzantine fresco108.
The development of tuneable excitation LEDs has shown the potential to image 70 different artists pigments when combined with an infrared HSI spectrometer. The variable excitation expands the number of pigments that luminesce and
using HSI as detection instead of a single band modified camera increases the discrimination power of detection. These two improvement allow VIL to detect more varying pigments109.
4.3
Scanning techniques
Contrasting the imaging techniques that collect information from multiple pixels at once, as described in the previous sections (4.1 & 4.2), scanning tech-niques acquire spectra on a per-pixel basis. Besides MA-XRF scanning there are more macroscale scanning techniques, which are described in the following subsections.
4.3.1
Macroscopic Fourier Transform Infrared (MA-FTIR)
Fourier transform infrared (FT-IR) spectroscopy is a long-established110,111
labo-ratory point-based technique to investigate vibrational molecular states, mainly employed for binding medium analysis110,111 and pigment identification112–114. With
the advent of attenuated total teflection FT-IR (ATR-FTIR) microscale imaging of cross-sections was made possible16,59.
A novel MA-FTIR reflectance instrument was presented by Legrand et al.115 in
2014, using a point scanning FTIR instrument, to collect full spectral data for every pixel. The mid infrared (MIR) scanner provided comparable information to an IRR, when similar wavelength bands were summed to create a virtual IRR reflectogram (section 4.1.3). On top of that, the MA-FTIR was able to more selec-tively visualise underdrawings by investigating different wavelength regions. It was additionally capable of providing pigment-specific information in accord-ance with XRD data (section 4.3.2).
There are some differences between this spatial scanning approach when compared to the hyperspectral imaging spectral scanning approach; while this spatial scanning approach produces a higher spectral resolution, it also requires longer acquisition times. Moreover, the MA-FTIR scanner uses mid infrared (2500 – 7500 nm / 4000 – 1333 cm-1), whereas the hyperspectral
imaging systems generally record spectra in the VNIR (400 – 950 nm / 25 000 – 10 500 cm-1) and SWIR (950 – 2500 nm / 10 500 – 4000 cm-1) ranges. Imaging
in the MIR region is beneficial, as it is the range employed by traditional FT-IR instruments, meaning more knowledge and reference libraries are available for this region. However, these measurements are usually done in a transmission geometry and care must be taken to correct the reflectance spectra in the right way to use these measurements.
In 2020, an automated, PCA-based (section 6.2.2) data processing methodology was established, which Sciutto et al.116 claim would allow the spectrometer to
be used by researchers with little chemometric domain knowledge. The focus was placed on a real-time algorithm, that would allow FTIR experts to quickly interpret the data without requiring them to have intimate knowledge of the data processing involved.
4.3.2
Macroscopic X-Ray Diffraction (MA-XRD)
XRD enables the investigation of crystal lattice structures using the diffraction pattern of X-rays coming from the sample. Many pigments can be distinguished in this way and certain pigments with multiple crystallographic forms can be separated (e.g. cerussite and hydrocerussite in lead white)117,118.
Akin to XRF, XRD imaging started as a SR-based technique119,120 but has recently
been developed as a laboratory macroscopic scanning technique. In transmis-sion mode it provides bulk chemical information117, whereas reflection mode
provides surface chemical information118. Due to this difference, the two can
be combined to carry out a depth-dependent XRD investigation of different pigment mineral forms on a painting121. Moreover, the angular diffraction
differ-ence has been used to map paint stroke direction117 and differentiate manuscript
5
To make data fusion possible, the data from different spectroscopic techniques and apparatus must be processed in such a manner that it can be compared and combined. Therefore, this chapter investigates different types of common spectroscopic data processing regimes that are applied to XRF (Ch. 3) and its complementary techniques (Ch. 4).
5.1
Spectral Pre-Processing
Before spectral data is truly processed to extract chemical or elemental infor-mation, pre-processing steps are often carried out. The aim of pre-processing is not to extract information directly but to facilitate extraction using the process-ing algorithms later in the pipeline (Figure 9). Although many such techniques exist, a few of the most common ones are discussed here.
Data pre-processing
HSI data acquisition Endmember extraction Pigment identification Pixel classification
Figure 9. Schematic example of a hyperspectral data processing pipeline. First hyperspectral imaging data is collected, then
pre-processing is carried out. The pre-processed data is fed into an endmember extraction algorithm and the endmembers are then matched to a pigment library to identify pigment and classify the image. Data fusion is possible between each step in the pipeline.
Source: Example created by Laurens van Giersbergen, generalised from various papers, such as 58,99–101,115
5.1.1
Normalisation
The group of normalisation techniques is important for spectral processing. It seeks to make data more comparable and often allows for easier data extraction. Mean-centring is the process of subtracting the mean from all measurements of a spectrum. This can be done to remove baseline drift effects between different measurements, especially when various operating conditions have been employed115,122.
Another step that is often taken is standardisation, especially when fusing data from different techniques. Here, the mean-centred data is scaled with the standard deviation, equalising the variance of each measurement. Some processing or fusion algorithms are sensitive to a measurement’s magnitude or variance. The standardisation step eliminates these differences and is known to produce better models in Support Vector Regression123 (SVR, section 6.2.5) and
especially in Principal Component Analysis (PCA, section 6.2.2), which seeks to find maximum variance components116,124,125.
Another often used technique is rescaling, min-max normalisation, or apparent reflectance. Here the minimum data value is subtracted from all data points and they are then divided by the largest data point. All data points are then scaled from 0 to 1. This is sometimes done as a pre-processing step but is most often used to produce the final images, as it allows for the full range of image values to be used to display data values58,99,114.
5.1.2
Savitzky–Golay (SG) Filtering
Savitzky–Golay (SG) filtering is a type of digital smoothing filter. It functions by linear least squares fitting a user-selected function to a specified sub-set of data points at a time. For data that is equally spaced (as is often true for spectral data), convolution coefficients can be calculated analytically. These coefficients are widely published and can easily be applied as a filter kernel to move over the data126.
The filter kernel (or function, for unevenly spaced data) is chosen based on the user’s need and produces different results. Moving average and first- and second-order derivative filters are among the most common spectral filters used. The derivative kernels are especially useful as they highlight shape changes in a spectrum122, which is often related to critical spectral information.
5 Data Processing
30
5.2
Background Removal
Background removal is a common spectral data processing technique to isolate potentially useful peaks from the background. In XRF measurements, this back-ground is mainly caused by scattered primary radiation and matrix effects127.
Alfeld and Janssens127 outline three methods for background removal of XRF
data. The simplest and with it, crudest, method is subtracting the background shape of the sum spectrum from every pixel’s spectrum. Although this method is incredibly fast, it does not allow for the modelling of regional variations in the background signal. In their paper they show how this approach incorrectly models manganese content due to the canvas painting’s wooden stretcher (Figure 10)127.
Figure 10. Manganese elemental map of Supper at Emmaus by Caravaggio, showing a t-shaped artifact due to the assumption
of a constant background in the GeoPIXE software suite. The artifact is caused by a difference in primary radiation scattering between the areas of the canvas that are on top of the stretcher frame, compared to areas that only have air behind them. On the left is the original image from Alfeld and Janssens, the right shows outlines to accentuate the t-shaped artifact.
Source: Alfeld and Janssens (2015)127
The second option for background removal outlined by Alfeld and Janssens is background estimation with a filter. Statistics-sensitive non-linear iterative peak-clipping (SNIP) is among the most common ones, where the average of a set number of surrounding spectral channels is compared to the middle chan-nel. If the average is lower than the middle value, the middle value is set to the
average. The procedure slowly ‘erodes’ the peaks by moving over the spectrum until a background shape is estimated over several iterations127. SNIP is one
of the two methods implemented in the near-universally used open source PyMca software, that is produced and maintained by the European Synchro-tron Radiation Facility (ESRF). In the ESRF’s paper, accompanying the release of the first version of the PyMca package, it is noted that the SNIP procedure is often preceded by Savitzky-Golay smoothing (section 5.1.2)128.
The last method described by Alfeld and Janssens, which is again implemented in the PyMca package, entails modelling the background with a higher-order polynomial. The background can be fit with linear or exponential functions during the fitting of elemental peak profiles. They note that generally 5th to 8th
order polynomial functions provided a sufficient model of the background. The modelling approach is generally faster and more applicable for spectra
with low counting statistics. However, the erosion estimation approach is more capable in modelling complex, discontinuous background shapes at the cost of speed. The user thus has to assess the spectrum visually and choose between the limitations of the two methods.
5.3
MA-XRF Image Processing
As mentioned (section 5.2), PyMca is currently the most popular tool for MA-XRF image analysis. For extraction of elemental abundance maps it implements ROI imaging (section 5.3.1) and a least squares fitting procedure (section 5.3.2). The first software package for XRF processing in common use was AXIL, based on a least squares fitting approach. Although it is still used due to its fast non-linear least squares (NL-LS) fitting method, it is not suited for large numbers of spectra such as in MA-XRF imaging. GeoPIXE was originally developed for earth sciences to analyse PIXE data but was soon co-opted for XRF measure-ments, as it is based on the significantly faster dynamic analysis (DA) method
(section 5.3.3)127. The Datamuncher software suite was developed by Alfeld and
Janssens127 to exploit the different strengths from AXIL, PyMca, and GeoPIXE and
is seeing increasing use. It was initially developed to allow batching of AXIL or PyMca fitting, useful for MA-XRF maps that are composed of multiple different scans. Later, they implemented the faster DA method, without GeoPIXE’s restric-tion of only a single background that can be subtracted from all spectra, with the same background subtraction options available as in PyMca (section 5.2)127.
