Visible and Infrared Imaging Spectroscopy of Picasso’s HarlequinMusician: Mapping and Identification of Artist Materials in Situ

JOHN K. DELANEY,* JASON G. ZEIBEL, MATHIEU THOURY, ROY LITTLETON,MICHAEL PALMER, KATHRYN M. MORALES, E. RENE DE LA RIE, andANN HOENIGSWALDScientific Research Department, National Gallery of Art, 4th and Constitution Ave NW, Washington D.C. 20565 (J.K.D., M.T., M.P., K.M.M.,E.R.R.); Night Vision & Electronic Sensors Directorate, 10221 Burbeck Rd, Ft. Belvoir, Virginia 22060 (J.G.Z., R.L.); and Painting Conservation

Department, National Gallery of Art, 4th and Constitution Ave NW, Washington D.C. 20565 (A.H.)

Reflection imaging spectroscopy is a useful technique to remotely

identify and map minerals and vegetation. Here we report on the

mapping and identification of artists’ materials in paintings using this

method. Visible and infrared image cubes of Picasso’s Harlequin

Musician are collected using two hyperspectral cameras and combined

into a single cube having 260 bands (441 to 1680 nm) and processed using

convex geometry algorithms. The resulting 18 spectral end members are

identified by comparison with library spectra, fitting by nonlinear

mixing, and using results from luminescence imaging spectroscopy. The

results are compared with those from X-ray fluorescence spectrometry,

polarized light microscopy, and scanning electron microscopy–energy

dispersive spectrometry (SEM/EDS). This work shows the potential of

reflection imaging spectroscopy, in particular if the shortwave infrared

region is included along with information from luminescence imaging

spectroscopy.

Index Headings: Reflection spectroscopy; Luminescence spectroscopy;

Hyperspectral imaging; Art conservation; Cultural heritage.

INTRODUCTION

Conservation of paintings requires knowledge of the artist’smaterials used, such as colorants, binders, and preparatorylayers. This information can also provide insight into theartist’s working methods. A large number of analyticalchemical techniques have been developed for the identificationof these materials. Some of these methods require a smallsample to be taken, such as light microscopy (polarization andfluorescence), scanning electron microscopy–energy dispersivespectrometry (SEM/EDS), gas chromatography–mass spec-trometry (GC-MS), and high-performance liquid chromatogra-phy (HPLC). Other methods can be applied in situ and thus donot require samples, such as X-ray fluorescence spectrometry(XRF), Fourier transform infrared spectroscopy (FT-IR), fiber-optics reflection spectroscopy (FORS, 350 to 2500 nm), andRaman spectroscopy. All of these methods, however, are pointmeasurement techniques and as such are applied to a limitednumber of sites on a painting, often selected by visualinspection alone. The development of in situ tools capable ofexamining the entire surface of a painting, to at least guide thesite selection for further study using point analytical methods,is of interest to the conservation field.

Some of the in situ techniques in widespread use, such asXRF and FORS, could potentially be used to raster scan anentire painting. However, the long dwell times required makescanning all but small sections impractical. For example, a one

mega pixel raster scan of a painting with an instrumentrequiring a 250 ms dwell time per point, would take ;4000min. Thus, multiplexing methods are required. The FORStechnique can be modified to scan a painting if the single-fiberspectrometer is replaced by an imaging spectrometer. Theimaging spectrometer provides an efficiency improvement overthe single fiber that is related to the number of pixels across theslit. For example, a scanning imaging spectrometer having1024 pixels across the slit reduces the scan time to ;4 min forthe same dwell time. A visible to near-infrared mechanicalscanning imaging spectrometer (400 to 900 nm) has been builtand has demonstrated the advantages of such a methodology toscan paintings at useful spatial resolutions.1 Recently twoportable optical scanning imaging spectrometers, one visible tonear-infrared (417 to 973 nm, 240 bands) and the other near-infrared to shortwave infrared (895 to 1749 nm, 85 bands) havealso demonstrated these advantages, as well as improvedvisualization of compositional paint changes and preparatorysketches in the shortwave infrared range (hyperspectral-infrared reflectography).2

Prior studies of artists’ colorants using diffuse reflectionspectroscopy have shown that the most robust identification ofartists’ materials requires collecting spectra from the near-ultraviolet (UV) to the shortwave infrared (;350 to 2500nm).3–8 For example, many of the inorganic colored and whitepigments can be identified using FORS. This larger spectralrange allows collecting not only the diagnostic electronictransitions that give rise in part to the ‘‘color’’, but alsovibrational overtones and combinations that can provide uniqueidentifying features. The spectral resolution needed for FORSvaries across these spectral regions. In the visible (380–700nm) to near-infrared (NIR, 700–1000 nm) range many of theelectronic transitions (i.e., blues, greens, and earths) varyslowly with wavelength, implying that a low spectral resolutioncan be used, but to discriminate between pigments havingsharp s-shaped transition edges, such as reds and yellows andmixtures, a resolution of ,5 nm may be necessary. In theshortwave infrared (SWIR, 1000–2500 nm) the requiredresolution is ;10–20 nm in order to resolve the narrowvibrational bands,5 similar to what is required for separationand identification of minerals.9

In art conservation the application of reflection imagingspectroscopy, and the collection of images in different spectralbands, for the study of paintings has mostly been carried outusing multispectral cameras having a few to tens of spectralbands that cover the visible range or extend into the NIR. Theseimages are exploited either by creating three-band false colorcomposites or by calibrating to reflectance in order to obtain

Received 1 December 2009; accepted 30 March 2010.* Author to whom correspondence should be sent. E-mail: J-Delaney@NGA.gov.

584 Volume 64, Number 6, 2010 APPLIED SPECTROSCOPY0003-7028/10/6406-0584$2.00/0

� 2010 Society for Applied Spectroscopy

low-resolution reflectance spectra. Studies in which themultispectral imaging has been extended into the SWIR, likeexperiments with Vis-SWIR FORS (350 to 2500 nm), havefound improved identification and separation of many artists’materials.2,5,10,11 For example, visible to SWIR multispectralimaging has been used to separate and map Prussian blue,ultramarine, and cobalt blue in Van Gogh’s paintings LaMousme (1888) and Self Portrait (1889), both in the NationalGallery of Art’s collection, by utilizing the large reflectancechanges in the SWIR.5 The need to extend beyond the visible isnot surprising given that one can mix a variety of artists’pigments to achieve similar colors. Thus, as in remote sensing,it is important to acquire images in a sufficient number ofspectral bands that span the spectral range required to separatematerials of interest. Hence, we have included both the visibleto near-infrared (VNIR) and the SWIR in part to test whetherthis is sufficiently robust to separate, identify, and map artists’colorants in situ.

For some colorants, for example cadmium sulfide (CdS) andcadmium sulfoselenide (CdS1-xSex), the reflectance spectrahave only a simple s-shaped transition edge, making theiridentification difficult, especially since such pigments can bemade in a wide range of hues. The same can be true for organicdyes, such as substituted anthraquinones, used in high opticaldensity, where the blue/green absorption bands may not beapparent in reflectance.12 Here, luminescence from thesematerials can in principle provide additional information13–16

to aid in their identification and spatial mapping. Cadmiumpigments, as semi-conductors, can fluoresce from the conduc-tion band, or from deep traps associated with surface defects orimpurities. The emission from these deep traps is red-shiftedinto the NIR.13,17 Similarly, organic dyes and pigments canfluoresce depending on their chemical structure and themordant used, as well as their microenvironment. Thus,luminescence imaging spectroscopy can be used in combina-tion with reflection imaging spectroscopy to improve therobustness of the identification and mapping. In luminescenceimaging spectroscopy the image cubes are calibrated to spectralradiance, rather than reflectance, to allow comparison ofemission spectra. The luminescence spectral bands aretypically broad (.50 nm), allowing the use of wide spectralfilters, which is advantageous given the low quantum yield ofsuch emissions.

In this paper we report on the identification and mapping ofthe primary colorants used in Pablo Picasso’s HarlequinMusician (1924) in the collection of the National Gallery ofArt, Washington, D.C. (Fig. 1) using Vis-SWIR reflection andVNIR luminescence imaging spectroscopy. Two novel hyper-spectral cameras, recently shown to have sufficient spectralsampling to provide Vis-SWIR FORS quality spectra frompaintings,2 and a multispectral luminescence spectral cameraare used to generate the two image cubes. Since the resultingVis-SWIR image cubes provide a large number of spectra,traditional hyperspectral image processing algorithms are usedto find end members and associated spatial maps. Identificationof the primary colorant or in some case mixtures is done bycomparison with reflectance and luminescence spectral data-bases, as well as by nonlinear reflectance mixing modeling.These results are compared with those from traditional site-specific methods such as in situ XRF and micro-sampleanalysis using polarized light microscopy (PLM), SEM/EDS,and FT-IR.

EXPERIMENTAL

Hyperspectral image cubes are acquired using two hyper-spectral imaging spectrometers in a whiskbroom scan. Thepainting is illuminated diffusely by lamps at 50 degrees fromthe painting normal using two 1000 W Halogen lamps fittedwith UV blocking filters placed 12 ft away. The pixel non-uniformity correction and flat fielding is done with captures ofdark frames and white panels. Four image scans are made witheach camera to cover the painting, and collection time is a fewminutes per scan for each spectral image cube. Two Spectralon(Labsphere, NH) standards (2% and 98% diffuse reflectors) areused as in scene calibration targets to convert the resultingimage cubes to reflectance factor. All images are non-uniformity corrected and calibrated to reflectance factor byempirical calibration through processing in ENVI (ITT)software. Spectral analysis is performed using ENVI (versions4.4 and 4.2) on an Apple MacBook Pro. Reference FORSreflectance spectra were obtained using a fiber spectroradi-ometer (350 to 2500 nm ASD FS3).

Hyperspectral Cameras. The VNIR imaging spectrometeris a model SOC-730VNIR from Surface Optics Corporation(San Diego, CA).2 It is a slit based, scanning hyperspectralsensor that uses a grating spectrometer from Specim Corpo-ration (Finland) to perform the wavelength dispersion. Eachindividual frame records one spatial and one spectraldimension of data. There is a rotating scan mirror to sweepout spatial dimension over time. The sensor uses a 1024 31024 element Dalsa VNIR staring Si charge-coupled device(CCD) camera to perform data acquisition with 12 bits ofdynamic range. To improve signal-to-noise ratio (S/N) anduniformity, the pixels are binned by two, leading to a 512 3512 pixel image, and an integration time of approximately 250ms is used, giving a cube acquisition time of approximatelytwo minutes per image.

In this work, a 70 mm EFL fore optic is used at f/2.2 tomatch the f-number of the spectrometer. The full field of view(FOV) for the sensor is 8.48, and the instantaneous FOV is 0.14milliradians. The sensor is responsive from 417 nm to 973 nmand the spectral resolution used here is 2.25 nm, giving 240distinct colors in the data cube. Before each data acquisition, 25frames of internally shuttered data are acquired. These ‘‘darkframes’’ are then used in the image calibration process asdescribed earlier.

The SWIR imaging spectrometer is a slit based, scanninghyperspectral sensor that uses a grating spectrometer forwavelength dispersion, fabricated by Surface Optics Corpora-tion (SOC-720SWIR).2 The SOC-720SWIR incorporates a 6403 512 element InGaAs array sensor (SUI-640SDV, SensorsUnlimited) with 14 bits of dynamic range. An integration timeof 33 milliseconds is used to acquire the data, giving anacquisition time for each 640 3 640 3 85 element data cube ofslightly over 22 seconds.

A 50 mm EFL fore optic, at f/3 is used, giving an 18.28vertical FOV and a 0.5 milliradian IFOV. The sensor isresponsive from 895 nm to 1749 nm and in this work, thespectral bands are separated by approximately 10 nm, whichgives 85 distinct spectral bands per data cube.

Luminescence Imaging Spectroscopy. A homemademultispectral camera system, operating from 650 to 900 nmin six bands (full width at half-maximum, FWHM, of 40 nm),is used to obtain relative radiance image cubes. Blue/greenexcitation is obtained using two filtered (Kron/Cousins B

APPLIED SPECTROSCOPY 585

FIG. 1. (Left) Visible color image of Pablo Picasso’s Harlequin Musician (1924) (photo by Gregory Williams, Image courtesy of the Board of Trustees of theNational Gallery of Art, Washington, D.C., given in loving memory of her husband, Taft Schreiber, by Rita Schreiber). (Top left) VNIR/SWIR hyperspectral cubeand (bottom left) cluster map used to define the primary end members. (Center) A plot of the reflectance spectra of the end members and (right) their associatedspatial distributions in the painting.

FIG. 2. Results of luminescence imaging spectroscopy after exciting the painting with blue light. (A) False color near-infrared luminescence image (700, 750, and800 nm) obtained from exciting the painting with blue light (380 to 520 nm) after median filtering, (B) map of primary emission spectra (bottom left panel) from theluminescence image cube after median filtering (angles 0.1 except blues, 0.08), (C) visible color image for reference. (Bottom left) Luminescent end memberspectra: Yellow (yellow line), Bright Orange (orange line), Dark Orange (brown line), Green (green line), Blue and White end members (purple lines), Mauve endmember (reddish purple line). (D) Map of the hyperspectral reflectance end members that are luminescent. (Bottom right) End member spectra obtained fromprocessing the reflectance hyperspectral-imaging cube (SAM angles, all 0.1 except green, 0.14). Reflectance end members: Yellow (yellow lines), Bright Orange(orange line), Dark Orange (brown line), and Green (green line).

586 Volume 64, Number 6, 2010

astronomy filter, 380–520 nm) slide projectors (Kodak Ektra-graphic III) set up at ;458 from the normal at 12 feet from theobject. The spectral radiance at the object is 80 milliwatts/sr-m2. The multispectral camera consists of a low-noise 4megapixel Si CCD camera (Qimaging Retiga 4000TR)outfitted with F/D 2, EFL 25 mm, hyperspectral lens (CoastalOptics). The spectral interference pass band filters (FWHM 40nm, Andover Corp) are used to create the six-spectral-bandimage cube. The camera integration time is optimized for eachcapture and varied between 1 and 15 s, with the focal planebinned 23. An in-scene calibration luminescence target is usedto convert the images to relative spectral radiance. The targetluminescence spectrum is measured using a spectroradiometer.

Fiber-Optic Reflection Spectrometer. A fiber-optic spec-troradiometer, FS3 (ASD Inc, Boulder, CO) is used to obtainFORS spectra at various sites in the painting. The spectrometeroperates from 350 nm to 2500 nm with a spectral sampling of1.4 nm from 350 to 1000 nm and 2 nm from 1000 to 2500 nm.The spectral resolution at 700 nm is 3 nm and at 1400 and 2100nm is 10 nm. The light source of a leaf probe head (ASD Inc)was used at a distance of 20 cm to illuminate the painting(;400 lux) and the fiber was placed ;1 cm from the object,giving a ;3 mm spot size at the painting. Typically 16 spectraare averaged and total acquisition time is ,5 s per point.

Polarized Light Microscopy. A Leica Orthoplan polarizinglight microscope configured for transmitted light is used. Thepigment scrapings are mounted as slide preparations usingCargille Meltmount 5870 (nD ¼ 1.662) and examined atmagnifications ranging from 3003 to 5003. Optical andmorphological properties of the unknown particles are notedand comparisons with standard reference pigments are made.

X-ray Fluorescence Spectrometry. A Bruker (formerlyRoentec) ArtTAX Pro XRF Spectrometer, equipped with ahelium flush, a rhodium X-ray tube, and capillary optic lens, isused to obtain elemental composition at sites on the painting.The analysis area at the painting is approximately 75 lm indiameter. The X-ray tube is operated at 50 kV, 200 lA, with alive integration time of 200 seconds.

Scanning Electron Microscopy–Energy Dispersive Spec-trometry. A portion of each dispersed pigment sample isaffixed to a conductive carbon stub using double-side carbontape. Each sample is then examined with a Hitachi S3400variable pressure scanning electron microscope (VP-SEM)fitted with an Oxford Super ATW detector and the INCAenergy dispersive spectroscopy analysis system. Analysis iscarried out in variable pressure mode (40 Pa) at a workingdistance of 10 mm and with an accelerating voltage of 20 kV.Samples are analyzed at magnifications ranging from 15003 to50003. Elemental profiles generated from the SEM/EDSanalyses served as a basis for inferring what pigments arepresent in each of the dispersed samples.

RESULTS

Pablo Picasso’s Harlequin Musician, painted in 1924, was areturn to an earlier theme from the commedia dell’arte and isfrom the synthetic Cubism period (Fig. 1). The colors of thefigure in this version of the Harlequin are more ‘‘heightened’’than in the earlier ‘‘dusty rose period of the Saltimbanques’’.18

The Harlequin Musician is thickly painted and the visiblecomposition is over an existing painting based on visual clueson the surface and from the X-ray image.19 This phenomenonwas not uncommon for Picasso. The earlier painting has yet to

be identified, although there are suggestions that it may havebeen a still life painted in a horizontal orientation. The largefields of bold and varied color represent an ideal test case forthe type of imaging spectroscopy presented here.

The experimental approach to identify and map the primarycolorants consists of first acquiring the VNIR and SWIRreflectance image cubes; second, applying hyperspectral imageprocessing algorithms to obtain end member spectra andassociate maps showing where the end members are found inthe painting; third, the collection, and calibration, of themultispectral luminescence image cube; fourth, the collectionof ‘‘truth’’ reflectance spectra, using FORS, at sites within theend member maps in order to check those spectra obtainedfrom the hyperspectral cube; and finally, the assignment of theprimary pigments, or possible pigments that make up the endmembers, which is done by comparison with spectral databases(reflectance and emission) as well as spectral fitting usingmixtures of pigments. The assignments are then compared withthose obtained from in situ XRF as well as from analysis ofmaterial from dispersed samples examined using polarizedlight microscopy and SEM/EDS.

Collection of Visible–Near-Infrared and Shortwave-Infrared Hyperspectral Image Cubes and Image Process-ing. Reflectance spectral image cubes for the VNIR and SWIRwere collected sequentially using the two cameras described inthe Experimental section and calibrated to reflectance factorusing white reflectance standards. The effective spatialsampling at the painting is 1 to 1.3 mm and four image framesare captured to cover the painting. After flat fielding andcalibration to reflectance factor, the residual root mean square(RMS) reflectance errors ranged from 1.4% to 0.55% on thewhite standard (98% reflective) over the spectral bands. Theresidual errors originate more from limitations on the amountof light that could be safely applied to the painting than frominherent instrumental limits.

Spectral image processing of the reflectance image cubes isdone using the hourglass paradigm developed by Boardman etal.20,21 and implemented in the ENVI image processing program.The two cubes were registered together using tie points selectedin the NIR, giving a single image cube and covering the spectralrange of 441 to 1680 nm (Fig. 1). A Delaunay spatial registrationmethod is used along with a cubic spline interpolator. Because ofnoise in the spectral overlap region of the two cameras, the 901to 968 nm spectral region is removed from the image cube priorto the analysis. To speed data analysis the image was processedat an effective pixel sampling of ;3 mm at the painting. Thisfurther reduces the residual reflectance errors per pixel by ;2-fold (,1%). In the image processing chain the eigen images aregenerated first using the minimum noise fraction (MNF)transform. Next, the pure pixel index is calculated, and finallythe ENVI Nd visualizer is used to find the end-member spectrathrough clustering.

For the VNIR/SWIR image cube, the first 27 MNF eigenimages were retained out of a possible 260. The higher eigenimages were found to have a ‘‘salt and pepper’’ striped noisestructure, attributable to residual non-uniformity of the focalplane geometry. The pixel purity index (PPI) map is obtainedusing 32 000 iterations, at a threshold of one, to find the mostsignificant pixels in the painting. Manual clustering is doneusing the Nd-visualizer algorithm in order to obtain the spectralend members as there is good separation between the clusters(Fig. 1). A summary of the resulting end member reflectance

APPLIED SPECTROSCOPY 587

spectra is given in Fig. 1. For the most part, they are spectrallyunique and span the range of colors. To ensure that theresulting end members’ spectra adequately describe thepainting’s surface, the spectral angle-mapping (SAM) algo-rithm is used. This algorithm identifies pixels in the image cubewhose reflectance spectrum matches those of end members’within a specified tolerance. The resulting end member spectralmaps (Fig. 1, angle tolerance values in Table I) show that thepainting is well described with the exception of the ‘‘gray/blackareas’’ in the painting, which turned out to have near constantreflectance from 441 to 1680 nm.

Luminescence Imaging Spectroscopy. Luminescence im-aging spectroscopy is used to obtain additional information inorder to separate and identify the pigments, especially theyellow, orange, and red pigments whose spectral features aresparse. The painting was illuminated with blue/green light (380to 520 nm) in order to optimally excite the yellow to redpigments and six images were collected between 650 and 900nm at 50 nm (FWHM 40 nm) intervals. The resultingluminescence image cube was calibrated to relative spectralradiance using an in-scene standard. A false color imageconstructed from the 700, 750, and 800 nm images shows thatthe green, yellow, and lighter orange paints are luminescent inthe NIR, whereas the dark orange is less so, and the red paint ofthe sleeves does not emit at all (Fig. 2A). Using similarmethods to those used to analyze the reflectance image cube,

the primary emission spectra and their spatial distribution in thepainting is given in Fig. 2B. These results show that theluminescence from the yellow (Harlequin’s medallions andstockings), green, and lighter orange paint end members allhave a peak at ;750 nm. The emission spectrum for the deeperorange has a shallower peak at ;750 nm. The spectra of theblue emission in the false color image (Fig. 2A) all showdeceasing emission intensity from 650 nm to 900 nm,suggesting that their emission peaks are below 650 nm. Thesespectral components map to some of the blue end members aswell as the white and mauve end members. They are assignableto emission from organic binders and white pigments,17 andmaybe an organic red pigment in the case of the mauve endmember.

Fiber-Optic Reflection Spectroscopy. Site-specific fiber-optic reflectance spectra (FORS) are acquired at sites definedby the end member maps in Fig. 1 with a spatial size of ;3 mmdiameter and spectral sampling of 2 to 3 nm. The FORS spectraare used to validate the spectra obtained from the hyperspectralcameras and referred to as ‘‘FORS reference spectra’’. Since thesolid angle of the fiber and the illumination geometry isdifferent from that of the hyperspectral cameras, differences inscales and offsets are expected.5

Analyses and Assignment of End Members. A tieredanalysis approach was used for the assignment of pigments tothe end members, which includes comparison to spectral

TABLE I. Assignments and verification of the various end members.

Region

HSI reflectance and MSI luminescence X-ray fluorescence (XRF) Micro-sample

Map tol. VNIR/SWIR Elementsa Possible assignment PLM, EDS-SEM

White face 0.08 Lead white Pb, (Zn, Ca) Lead white, zinc white, chalks Lead white, trace zinc whiteWhite around guitar 0.06 Zinc white, gypsum Zinc white, Ca sulfate (gypsum)Blue face 0.1, 0.06 Cobalt blue, lead white Pb, Zn, Co (Ca,

Ba, Mn, Fe,Al)

Cobalt blue, earths, Prussian,ultrmarine, lead white, zinc white,barium white, lithopone

Cobalt blue, lead white, possibletrace zinc whiteb

Blue hat 0.1, 0.1 Ultramarine, lead white Pb, (Ca, Ba, Fe,Zn)

Prussian, ultramarine, earths, leadwhite, zinc white, barium white,lithopone

Ultramarine, lead white, blackpigment (charcoal, bone)

Blue field behind head 0.3 Prussian blue, lead white Pb, Fe, (Cd, Ca,Zn)

Prussian, ultramarine, cadmiumyellow, lead white, zinc white

Oil, lead white, Prussian blueb

Red arms 0.1 Vermilion, white (TBD) Hg, Pb, Ba, (Ca,Fe)

Vermilion, earths, lead white,barium white

Dark red, near guitar 0.1 Unknown organic red dye,non-lead white

Organic red colorant, lithoponewhite, trace vermilion

Beige guitar 0.08 Yellow Fe oxide, gypsum,lead white, zinc white (?)

Iron earth, cadmium (CdS),black, lead white, zinc white,Ca (e.g., gypsum, dolomite,Ca carbonate)

Mauve field behindhead

0.08 Ultramarine, red dye, leadwhite

Pb (Ca, Ba, Fe,Zn)

Earths, lead white, zinc white,barium white, lithopone

Red lake dye, ultramarine blue,whites: lead, barium

Dark brown walls 0.08 Iron oxide, non-lead white Pb, Zn, Fe, Ca,Mn (Si, K, Ti)

Iron earths, umber, red earth, leadwhite, zinc white

Iron earths, quartz, zinc white

Yellow stockings 0.1 Cadmium yellow paint,lead white

Pb, Cd, Ba, (Ca,Fe, Zn, Hg)

Cadmium yellow, vermilion, leadwhite, barium white, lithopone,zinc white

Cadmium yellow (CdS), whites:barium, trace zinc

Orange guitar 0.1 Orange (TBD), cadmiumyellow, non-lead white

Zn, Ba, Cd, (Pb,Fe, Cr)

Cadmium yellow, earths, tracechrome orange, zinc white,lithopone, barium white, leadwhite

Lead chromate (orange),cadnium yellow (CdS),whites: zinc, barium

Green chair 0.14 Cadmium yellow,ultramarine, lead white (?)

Pb, Cd, Ba, (Ca,Fe, Zn)

Cadmium yellow, earths, Prussian,ultramarine, lead white, bariumwhite, lithopone, zinc white

Cadmium yellow (CdS),ultramarine blue, bone black,lead white

Grays NA Black (TBD), white (TBD) Zn, Ca, Pb (P, K,Fe)

Iron earths, bone black, zinc white,lead white

a Major elements are given in bold, minor in plain type, and trace elements in parentheses.b FT-IR results from Ref. 25.

588 Volume 64, Number 6, 2010

FIG. 3. White end members, (left) spatial distribution map (tolerances 0.013 green and 0.012 red) and (right) reflectance plots, obtained from analysis of thehyperspectral image cube. (Inset) FORS SWIR reflectance spectra of reference zinc white (top trace) and lead white (bottom trace) paints (linseed oil). (Middle) Plotof the FORS reflectance spectra in the visible range from sites defined by the white end member maps. The green areas on the map are assigned to lead white and thered to zinc white containing gypsum as a bulking agent (Table I).

FIG. 4. Spatial maps and reflectance spectral plots of the weakly to non-fluorescing spectral end members obtained from the analysis of the hyperspectral imagecube. (Left) Spatial distribution map of (middle) the end members on the painting obtained using the spectral angle mapper algorithm. (Right) Associated referenceFORS reflectance spectra obtained from sites defined by the end member maps and the result of nonlinear mixing reflectance fits using pigments (black line)identified in Table I and lead white. (Top row) Blue end members: Ultramarine blue (blue lines), Cobalt blue (purple and aqua blue lines), and Prussian blue (darkblue line). (Middle row) Red end members: Vermilion (red line), Organic red dye (dark red line). (Bottom row) Beige, Brown, and Mauve end members: Yellowiron oxide (mustard line), Brown earth (brown line), and Red lake and Ultramarine (purple line).

APPLIED SPECTROSCOPY 589

databases, luminescence information, and spectral fitting. Sucha method is found to be necessary because not all the endmember spectra are sufficiently identifiable by reflectancespectral shape alone. A summary of the resulting assignmentsand verification is given in Table I. Assignment of the paints/pigments via reflectance spectra was done by comparison topublished spectra of minerals22 and databases of reflectancespectra, including powder and painted samples, constructed bythe National Gallery of Art and CNR-IFAC.23 Luminescencespectra were compared to published spectra and an in-housedatabase of three-dimensional (3D) luminescence spectra ofwhite pigments and CdS and CdS1-xSex paints.17 In somecases, mock-ups were made of pigment mixtures in order totest possible assignments. Fitting of the spectra was done usinga single-point Kubelka–Munk model, which takes as anassumption that the scattering from the pigments in the paintlayers and the database are the same.

White End Members. The white areas of the painting aredescribed by the two end members found (Fig. 3), which arespectrally distinct in the SWIR (1400 to 1600 nm). The firstend member has a narrow absorption band at 1450 nm (1445nm FORS), whereas the second has a broader absorptionfeature at 1450 nm (1446 nm FORS) with shoulders at 1490and ;1540 nm (1485, 1534 nm FORS). The spectral feature at1206 nm is common to both white end members and is likely avibrational overtone (e.g., CH2 stretch) from organic materialssuch as the binder. Note that the feature at ;900 to 970 nmoriginates from transition between the two imaging spectrom-eters. The maps in Fig. 3 are made using the 1010 to 1680 nmspectral portion of the cube since this is the region of primaryspectral difference. The first end member maps to theHarlequin’s face and upper body, and the second is localizedto the white areas surrounding the guitar. FORS spectra takenat sites defined by these maps confirm the spectral differenceobserved in the SWIR but also show differences in the 350 to400 nm range where the VNIR camera does not operate.Specifically, the second end member has near zero reflectancefrom 350 to ;380 nm, whereas the first end member has.25% reflectance over this region.

The most probable white paints to have been used are leadand zinc white, which can be distinguished by reflectionspectroscopy.3,4 Zinc white, a semi-conductor, has an inflectionpoint at ;384 nm associated with its band gap transition, whilelead white is reflective throughout the 350–400 nm region. Inthe SWIR, lead white has a narrow absorption band at 1446 nm(FORS) associated with a hydroxyl group (Fig. 3), whereaszinc white, in oil, has broad and less defined absorptionfeatures in this region,3,4 and these features are likely from thesubstrate and binding media. Thus, the first end member thatmaps to the face and upper body is assigned to lead whitebecause of the reflectance in the UV/blue and the band at 1446nm (FORS). The second end member can be assigned to zincwhite from the inflection point in the UV at 385 nm, obtainedfrom the first derivative of the FORS spectra, although thepresence of SWIR bands at 1444 nm (FORS), with shoulders at1490 and 1540 nm (FORS) are not characteristic of zinc whitebut of hydrated calcium sulfate (gypsum, peaks at 1445, 1484,and 1531 nm24), a known bulking agent. The XRF measure-ment from a site in the map of the first end member gives alarge amount of lead and some zinc, suggesting the presence oflead white and some zinc white. The SEM/EDS analysis of asample from the white face shows lead white (Table I) and only

trace zinc white, which is consistent with the FT-IR analysis ofa similar site.25 Thus, the primary white pigment present is leadwhite, consistent with the reflection spectral analysis. For thesecond white end member, SEM/EDS of a dispersed samplefrom the lower proper right white field found elementalevidence for zinc white and calcium sulfate, consistent with theresults from reflection spectral analysis.

Blue End Members. The blue regions of the painting arewell described by five end members that have three distinctspectral shapes (Fig. 4). The reflectance differences in the NIR/SWIR allow direct assignment of the primary colorants, as hasbeen shown previously.2,5 The end members and associatedmap show that Prussian blue was used for the darker blue fieldbehind the Harlequin, cobalt blue for the lighter blues aroundthe face and arms, and ultramarine for the gray-blue fields ofthe hat, face, and chair. The spectral assignment of Prussianblue is made based on the low reflectance in the red and theSWIR until approximately 1000 nm, which originates from thecharge-transfer-state absorption, and a rise in reflectancebeyond that until ;1500 nm8 when it levels off. Theassignment to ultramarine blue is made based on the increasein reflectance in the NIR at 710 nm with a shoulder at 770 nmand constant reflectance in the SWIR. The two end membersboth assigned to ultramarine differ primarily in the reflectancelevel reached in the SWIR. Since ultramarine is relativelytransparent, these differences in reflectance are more likely torelate to the material below. As noted earlier, the end memberwith higher SWIR reflectance maps to the lower portion of thepainting (Fig. 1). In the map here (Fig. 4), the two endmembers are not distinguished since they are both assigned toultramarine blue. The spectral signature for cobalt blue is seenin the last two blue end members, specifically, the distinctivespectral shape in the SWIR (1200 to 1500 nm) as well as smallabsorption peaks in the visible (481, 545, and 624 nm (FORS)).These features originate from the ligand-field transitionsbetween the d–d orbitals of Co(II) in a pseudo-tetrahedralconfiguration.8 Note that the presence of cerulean blue isunlikely here as the SWIR feature would have been shifted to1400 to 1800 nm. The difference between the two cobalt blueend members is primarily the level of white mixed in.Examination of the 1400 to 1600 nm portion of the endmember spectra, and associated FORS spectra, shows a narrowband at 1446 nm and indicates that the dominant white pigmentused in combination with ultramarine and Prussian blues is leadwhite (see white end members). The broad absorption in theSWIR of the cobalt blue pigment obscures absorption bands inthis region from white pigments, making assignment of whitepigments problematic. However, in the case of the lightercobalt blue paint there is a small narrow absorption at 1445 nm,consistent with the presence of lead white.

Further confirmation of these assignments comes from thespectral fits to FORS reference spectra, using mixtures of leadwhite spectra and spectra of the blue pigments from thedatabases. These blue pigment assignments are in agreementwith site-specific results from XRF for cobalt blue and Prussianblue and from dispersed sample analysis using PLM forultramarine (Table I). While the XRF instrument used candetect lighter elements down to Na, heavy elements such aslead can absorb the emission from these lighter elements. Thus,the lack of Al in the XRF spectra for the ultramarine endmember is not unexpected.

590 Volume 64, Number 6, 2010

Red End Members. Three red end members were found bythe spectral analysis (Fig. 4) and map to distinct areas on thepainting. The end member that maps to the Harlequin’s redsleeves and patches on the tunic has a reflectance transitionedge at ;596 nm, indicative of vermilion (HgS) or a cadmiumsulfoselenide pigment. The lack of red to NIR luminescence,often observed with cadmium sulfoselenide pigments,17 andhigh opacity in the SWIR compared to the other moretransparent end members, favors the vermilion assignment.The FORS reference spectrum, taken from a site within themapped area, has a sharp s-shaped rise with an inflection at 591nm, as determined by the first derivative (Fig. 5). The FORSspectrum is well described by vermilion mixed with lead white(Fig. 4). XRF analysis at a site within the mapped areaconfirms the presence of mercury (Hg), consistent with thevermilion assignment, and lead, although the elementalanalysis also supports the possible presence of barium whiteand lithopone (barium sulfate and zinc sulfide) (Table I), whichare not observable in the reflectance spectra.

The two other red end members have no features in thevisible and similar transition edges in the red, but a slower risewith a transition at ;655 nm, and they reach different levels ofreflectance in the NIR and SWIR (Fig. 4). These two endmembers map to the dark purplish-red region around the guitar.Since the NIR and SWIR reflectance indicates that the paintlayer is relatively transparent in these spectral regions, thedifferences between these end members may occur from belowthe paint layer, rather than from the primary pigment. In theoverall map these two similar end members map to twoseparate areas (Fig. 1). The FORS spectrum, from a site in themapped area, is similar to the end member having lowerreflectance in the SWIR. The first derivative has an inflectionpoint at 635 nm, ;80 nm FWHM, which is twice that ofpigments such as vermilion and cadmium yellow. Possiblepigments include cadmium red, chrome red, and an organic red

pigment. The slow rise and lack of near-infrared redluminescence make the presence of cadmium red unlikelyand the color and spectral shape are not a good match tochrome red either. The lack of visible absorption bands,normally seen with substituted anthraquinones, would inprinciple rule out organic red pigments. However, at highoptical density, or when painted over an absorbing paint layer,the visible absorption bands are not always apparent in thereflectance spectrum.12

The color of the paint for these end members, the lack ofvisible absorption bands, and the 635 nm inflection point, plusthe 80 nm FWHM in the first derivative, do match theproperties of red lake carmine (carminic acid or cochineal) athigh optical density, as determined from test panel paint outs(our data and Ref. 12), but not those of madder lake. Thus, thereflectance spectrum is not inconsistent with assigning theprimary pigment to an organic red colorant such as carmine,although the presence of visible absorption bands in the blue/green, associated with the n!p* transition, would have madethe assignment more definitive. The broad spectral feature inthe SWIR (1400 to 1600 nm), having an absorption peak at1436 nm (FORS), is not characteristic of lead white, but moreof paint samples containing lithopone or zinc white (see Fig. 3and Refs. 3 and 4). A microscopic examination of a dispersedsample, taken from a site within the map, shows that it iscomposed of an unknown organic cool red colorant andlithopone, which is consistent with the reflectance results(Table I). A small amount of vermilion is found that is not seenin the reflectance spectrum, which is consistent with its lowabundance.

Beige, Brown, and Mauve End Members. The two endmembers that map to the beige guitar and the brownbackground of the wall (Fig. 4), and the FORS spectra fromsites in the mapped region, all have the characteristic spectralfeatures of iron oxides,22 specifically, the absorption from thestrong charge transfer bands in the UV and the Laporte-forbidden transitions in the NIR.22 The reflectance spectrum ofthe beige guitar end member has a transition edge at 545 nmand a peak reflectance at 770 nm and NIR absorption at 921nm. The dark brown end member has a peak reflectance at 760nm and NIR absorption at 861 nm. In the 1400 to 1600 nmspectral region the beige end member has strong absorption at1446 nm and shoulders at 1484 and ;1538 nm, not unlike thewhite end member found around the guitar identified as zincwhite with calcium sulfate (gypsum). However, the depth ofthe 1446 nm band is larger (;23) compared to that of gypsum,indicative of the presence of lead white, with its 1446 nm bandfrom the hydroxyl group.3,4 Hence, the paint used for the beigeguitar contains a yellow iron oxide or earth mixed withgypsum24 and lead white (Table I), although the presence ofzinc white cannot be ruled out, since yellow iron oxide isstrongly absorbing in the near UV. The dark brown endmember of the walls has a shifted and broader peak at 1434 nm(FORS), similar to the deep red end member, suggesting thatlead white is not present but that the presence of zinc white orlithopone is likely. The FORS reference spectrum can be fitusing mars yellow for the beige end member field and burntsienna for the brown mixed in the visible and NIR, but not inthe SWIR, since lead white without gypsum was used in modelmixing. XRF elemental analysis of a site within the dark brownend member gives evidence for iron oxides plus possible leadand zinc whites, although the lead white is not likely in the

FIG. 5. A plot of the first derivatives of the FORS spectra obtained at sites thatdefine the (A) Yellow, (B) Bright Orange, (C) Dark Orange, and (D) first Redend member maps.

APPLIED SPECTROSCOPY 591

dark brown paint, given the results from PLM and SEM/EDSanalysis of the dispersed sample (Table I).

The end member reflectance spectrum that maps to themauve fields behind the Harlequin shows a peak in the blue at470 nm and shoulders at 409, 509, and 528 nm, along with aninflection point at 630 nm. The small absorption bands in thevisible at ;533 nm and ;565 nm along with the blue peak inreflectance and red inflection point are indicative of substitutedanthraquinone (i.e., a red lake pigment). The color of the paint,however, is more of a purple rather than a red; thus, the color isnot likely to originate from a red lake pigment alone,suggesting the presence of a blue pigment as well. Evidencefor a blue pigment comes from the fact that the majorreflectance peak for the paint occurs at 468 nm, which is red-shifted compared to the lake pigments (i.e., carmine at ;430nm and rose madder at ;415 nm). Secondly, the red inflectionpoint at ;630 nm for the mauve end member is also too farred-shifted compared to these lakes, which typically haveinflection points at 580 to 600 nm for non-saturated paints.12

Ultramarine is the most likely blue pigment, given thefeatureless spectrum in the SWIR. Ultramarine blue has areflectance peak at ;470 nm along with the transition to highreflectance at 700 nm, which would help to explain the redshifts observed. The SWIR feature at 1450 nm (1445 nmFORS) seen in the end member spectrum is indicative of leadwhite. Spectral fitting of the FORS reference spectrum usingultramarine blue, along with the spectrum of a near-saturationcrimson red dye and lead white, gives a good fit to the bluepeak and the inflection point in the red and SWIR, but does notdescribe the visible absorptions as expected. Fitting usingPrussian blue or cobalt blue gave poor fits, especially in theNIR and SWIR, supporting the notion that the blue pigment islikely to be ultramarine. The dispersed sample taken from asimilar site provided evidence for a red lake pigment,ultramarine blue, and lead white (Table I). XRF elementalanalysis performed at a site contained in the end member mapis suggestive of the presence of iron earths and white pigmentssuch as lead white, barium sulfate, and lithopone. Since organicdyes are not measurable by XRF and ultramarine can beproblematic to assign, these results are not necessarilyinconsistent.

Yellow End Member. The recovered yellow end memberhas a primary sharp s-shaped transition at 476 nm with asecond smaller transition in the red (Fig. 2). The yellow endmember maps to the yellow stockings and to the yellowmedallions on the Harlequin’s shirt (Fig. 2D). Examination ofreflectance spectra over the yellow end member map shows theprimary transition at 476 nm, which defines the basic color andremains the same, but the second transition edge, occurring at;580 nm, varies in amplitude from ;0 to ;20% across theyellow map. Generally this transition is seen on the yellowmedallions on the Harlequin’s shirt and visual inspectionshows that they lie upon the larger red medallions. Thederivative of the FORS spectra shows a narrow peak at 477 nm(30 nm FWHM) and a smaller peak (;20%) at 583 nm, whichis close to the 591 nm peak of red vermilion (Fig. 5). Thus, theyellow looks to be partially transmitting, giving rise to somered reflectance from the vermilion below. The simple s-shapedreflectance transition edge at 477 nm can be assigned to avariety of yellow pigments, but the emission maximum at 750nm in the luminescence image cube suggests that it is CdS.13,17

The yellow end member reflectance spectrum in the SWIR has

a band at 1450 nm (1445 nm, FORS), attributable to leadwhite. The strength of this band varies over the yellow endmember map, being strongest in the stocking and weak in theyellow medallions, suggesting that the amount varies and thatother whites are present. XRF elemental analysis and PLMexamination of the dispersed sample show that the primaryyellow pigment is cadmium sulfide. Since cadmium yellow isrelatively transparent in the red to infrared range, theunderlying paint layers will contribute to the reflectancespectra. Thus, the small reflectance peaks that are observedin the red/NIR are not unexpected.

Orange End Members. The two end members recoveredfrom the hyperspectral end member analysis map to the outeredges of the guitar (Fig. 2D). Both reflectance spectra have asimple s-shaped rise in reflectance, but it increases less steeplythan in vermilion. The lighter orange end member has aninflection point at 544 nm and the darker orange at 550 nm.Several orange pigments could give rise to the color, such ascadmium orange or lead chromate. The presence of NIRemission from the lighter and to a lesser degree from the darkerorange would suggest the presence of a cadmium pigment.However, the observed emission maximum at 750 nm is notconsistent with a cadmium orange pigment (CdS1-xSex). Arecent study has shown that these NIR deep-trap emissionbands shift with the transition edge of CdS1-xSex mixtures.17

Thus, the 750 nm emission maximum is more consistent withcadmium yellow (CdS), and an emission maximum of ;800nm would be expected for CdS1-xSex orange having thereflectance transition seen here.16 Hence, the observed 750 nmemission in the orange field is likely to originate from cadmiumyellow (CdS) and not from a cadmium orange (CdS1-xSex).Evidence for a small amount of yellow pigment can be seen inthe first derivative of the FORS spectrum taken from a site inthe lighter orange end member map (Fig. 5), in which a smallpeak at ;455 nm is observed. The first derivative of the deeperorange, exhibiting less NIR luminescence, does not have anobservable reflectance peak in the yellow (Fig. 5). Thus, theorange color for both end members must be from a non-cadmium orange pigment, such as chrome orange (leadchromate, PbCrO4 Pb(OH2)).

The derivatives of the FORS spectra of the oranges bothshow primary transition edges at 540 and 550 nm and have abroader (by ;23) asymmetrical peak than cadmium yellowand vermilion. The colorant for the orange is thus likely to havea distribution of hue rather than a specific color. It is worthnoting that chrome orange is obtained by grinding red leadchromate particles more finely and a larger distribution in colormay not be unexpected. In the 1400 to 1600 nm region there isa broad asymmetrical absorption feature at 1440 nm (FORS),whose position and shape is not indicative of lead white but ofzinc white or lithopone. The XRF elemental analysis shows thepresence of cadmium sulfide, but not selenium, thus ruling outCdS1-xSex as the orange pigment, consistent with theluminescence results. Evidence for trace amounts of chromiumis observed in the XRF data, which is suggestive of chromeorange. PLM of a dispersed sample confirmed the presence ofcadmium yellow interspersed with a low amount of chromeorange, appearing as yellow to red small crystals. Thecharacterization of the particles as chrome orange among thecadmium yellow particles was accomplished using SEM/EDS.SEM/EDS analysis also found no evidence for lead white, butBa and S were detected together in some instances, while in

592 Volume 64, Number 6, 2010

others Ba, Zn, and S were all detected. These data furthersupport the presence of barium white and lithopone.

Green End Member. The single green end member shows apeak in the visible at 505 nm (500 nm FORS) that increases inreflectance to about 750 nm and in the SWIR remainsapproximately constant (Figs. 2D and 6). This end membermaps to the green of the Harlequin’s chair and is an unusualcolor. The luminescence image cube shows the green paintemission maximum at ;750 nm, similar to that of the yellowand orange and confirmed as yellow cadmium sulfide. Thus,the simplest hypothesis is that the green is a mixture ofcadmium yellow and a blue pigment. Alternatively, commer-cial mixtures of cadmium yellow with viridian green exist, butthese are a bright green to yellow green color, which isdifferent from the earthy, muddy green here. To determine themost likely blue pigment used in the green paint, single-pointKubelka–Munk mixing was done of cadmium yellow with thethree blues found in the painting (Fig. 6). Good agreement wasfound in the visible when CdS was mixed with ultramarine andcobalt blue, but not with Prussian blue. However, onlyultramarine blue gives a satisfactory fit/shape in the SWIR.Separately, mockups of greens were made using mixtures ofcadmium yellow and ultramarine, Prussian, and cobalt blue todetermine the effects of the blue pigments on the luminescencefrom cadmium yellow. Measurements of the luminescencefrom all the samples showed that mixtures with ultramarine andcobalt blue gave emission spectra most resembling that of thecadmium yellow paint, whereas the mixture with Prussian bluegave lower yield and a distorted emission spectrum shape. XRFelemental analysis confirms the presence of CdS and possiblyPrussian blue or ultramarine. The dispersed sample examina-tion using PLM confirmed that the blue particles areultramarine and not Prussian blue. Some evidence for the useof lead white is seen in the end member spectrum and FORSspectrum (band at 1438 nm), but the spectral feature is weakand broader than expected, which is in part consistent with thepresence of a black pigment observed in the PLM and SEM/EDS analysis of the dispersed samples.

Paint Media Spectral Bands. The FORS spectrometer,

unlike the hyperspectral cameras, extends to 2500 nm, a regionthat has been identified26 as having overtones of bindervibrational bands. The spectra from sites within the endmember maps (Fig. 7), with the exception of vermilion, allhave a series of absorption bands that line up with vibrationalovertone bands that have been assigned to drying oils.26

Specifically, bands at 2347 and 2304 nm are attributed to thesymmetric and anti-symmetric (CH2)þd(CH2) stretching andbending vibrations, and bands at 1754 and 1727 nm areassociated with the symmetric and anti-symmetric firstovertone bending of CH2. These results are consistent withthe primary paint medium being a drying oil, which is notsurprising. The zinc white end member (containing gypsum)and the beige end member, associated with the guitar, bothhave a strong 1944 nm absorption band in this region and aweaker band at ;2217 nm, which is likely from the gypsum.24

CONCLUSION

These results show that many of the primary colorants usedin the construction of Picasso’s Harlequin Musician can beidentified and mapped by imaging reflection spectroscopy, butonly when extended into the SWIR and when results fromluminescence imaging spectroscopy are included. Omittingpart of the NIR and SWIR portion of the reflectance imagecube is found to reduce the degree of separation of thecolorants in the analysis, making the clustering less complete.

FIG. 6. FORS reflectance spectra of green end member (black line) and resultof fitting by mixing cadmium yellow (CdS) paint with various blue pigments(Prussian, ultramarine blue, and cobalt) and lead white. The best fit is achievedusing the ultramarine blue end member (thick black line).

FIG. 7. SWIR FORS reflectance spectra from sites defined by the endmembers maps for (A) lead white, and (B) ultramarine blue with lead white andend members identified as containing gypsum, (C) beige guitar, and (D) zincwhite. The absorption bands at 1727, 1754, 2304, and 2347 nm are assignableto drying oils and those at 1944 and 2217 nm to gypsum.

APPLIED SPECTROSCOPY 593

This is found to result in an incomplete list of end members.This is especially true for the blue, green, earth, and whitepigments. Since the narrow vibrational overtone absorptionfeatures have been important for identification and separation,the higher spectral resolution used here appears to benecessary. Complementing the results of the reflection andluminescence imaging spectroscopy with data from site-specific in situ techniques, such as XRF, improves assignmentof pigments. These results suggest that when imagingspectroscopy (both reflection and luminescence) is combinedwith site-specific in situ analysis, such as XRF, robust pigmentidentification and mapping are possible. Reflection imagingspectroscopy in the visible to shortwave infrared range is apowerful tool for remote sensing of the earth.27 The resultsobtained here, in applying the technique to Picasso’s HarlequinMusician, suggest that this may be an important tool for thestudy of works of art as well.

ACKNOWLEDGMENTS

The authors would like to thank Dr. A. Fenner Milton and Mr. Ross Merrillfor their support, Mr. G. Williams and Mr. A. Newman for help in collectingthe digital images, and Dr. P. Ricciardi, Ms. E. Walmsley, and Dr. M. Picollo ofCNR-IFAC for helpful discussions. The authors also acknowledge theassistance of Mr. C. Kent in the acquisition of the image data. The authorswish to thank Dr. G. Gautier for sharing her FT-IR results. Drs. J. K. Delaneyand E. R. de la Rie acknowledge funding from the Andrew W. MellonFoundation.

1. A. Casini, M. Bacci, C. Cucci, F. Lotti, S. Porcinai, M. Picollo, B.Radicati, M. Poggesi, and L. Stefani, Proc. SPIE-Int. Soc. Opt. Eng. 5857,177 (2005).

2. J. K. Delaney, J. G. Zeibel, M. Thoury, R. Littleton, K. M. Morales, M.Palmer, and E. R. de la Rie, Proc. SPIE-Int. Soc. Opt. Eng. 7391, 739103-1(2009).

3. M. Bacci and M. Picollo, Studies Conserv. 41, 129 (1996).4. M. Bacci, M. Picollo, G. Trumpy, M. Tsukada, and D. Kunzelman, JAIC

46, 27 (2007).5. J. K. Delaney, E. Walmsley, B. H. Berrie, and C. F. Fletcher,

‘‘Multispectral Imaging of Paintings in the Infrared to Detect and MapBlue Pigments’’, in Scientific Examination of Art – Modern Techniques InConservation and Analysis (The National Academies Press, Washington,D.C., 2005), p. 120.

6. M. Picollo, M. Bacci, D. Magrini, B. Radicati, G. Trumpy, M. Tsukada,and D. Kunzelman, Proceedings of the Conference Modern PaintsUncovered (London, 2007), pp. 118–128.

7. P. Ricciardi, J. K. Delaney, L. Glinsman, M. Thoury, M. Facini, and E. R.de la Rie, Proc. SPIE-Int. Soc. Opt. Eng. 7391, 739106-1 (2009).

8. M. Bacci, D. Magrini, M. Picollo, and M. Vervat, J. Cultural Heritage 10,275 (2009).

9. G. A. Swayze, R. N. Clark, A. F. H. Goetz, T. G. Chrien, and N. S.Gorelick, J. Geophys. Res. 108, 5105 (2003).

10. S. Baronti, A. Casini, F. Lotti, and S. Porcinai, Appl. Opt. 37, 1299 (1998).11. A. Casini, F. Lotti, M. Picollo, L. Stefani, and E. Buzzegoli, Studies

Conserv. 44, 39 (1999).12. C. Bisulca, M. Picollo, M. Bacci, and D. Kunzelman, Proceedings of the

9th International Conference of NDT of Art (Jerusalem, 25–30 May,2008), www.ndt.net/article/art2008/papers/199Bisulca.pdf.

13. E. R. de la Rie, Studies Conserv. 27, 1 (1982).14. A. Casini, F. Lotti, M. Picollo, L. Stefani, and A. Aldrovandi, ‘‘Fourier

Transform Interferometric Imaging Spectrometry: a New Tool for theStudy of Reflectance and Fluorescence of Polychrome Surfaces’’, inConservation Science 2002 (Edinburgh, Scotland, 22–24 May, 2002).

15. D. Anglos, M. Solomidou, I. Zerioti, V. Zafiropuolos, T. G. Papazoglou,and C. Fotakis, Appl. Spectrosc. 50, 1331 (1996).

16. D. Anglos, C. Balas, and C. Fotakis, Am. Lab. 31(2), 60 (1999).17. M. Thoury and J. K. Delaney, unpublished data.18. M. McCully, The Burlington Magazine 139, 218 (1997).19. National Gallery of Art, unpublished technical image.20. J. W. Boardman, F. A. Kruse, and R. O. Green, ‘‘Mapping target signatures

via partial un-mixing of AVIRIS data: in Summaries’’, in Fifth JPLAirborne Earth Science Workshop (JPL Publication 95-1, v. 1, 1995), p.23.

21. J. W. Boardman, ‘‘Automated spectral un-mixing of AVIRIS data usingconvex geometry concepts: in Summaries’’, in Fourth JPL AirborneGeoscience Workshop (JPL Publication 93-26, v. 1, 1993), p. 11.

22. R. N. Clark, ‘‘Spectroscopy of Rocks and Minerals, and Principles ofSpectroscopy’’, in Manual of Remote Sensing, Remote Sensing for theEarth Sciences, Volume 3, A. N. Rencz, Ed. (John Wiley and Sons, NewYork, 1999), Chap. 1, p. 3.

23. CNR-IFAC, Fiber Optics Reflectance Spectra Database of PictorialMaterials in the 270 to 1700 nm Range, http://fors.ifac.cnr.it/.

24. G. R. Hunt, J. W. Salisbury, and C. J. Lenhoff, Modern Geol. 3, 1 (1971).25. Dr. Gwenaelle Gautier, Art Institute of Chicago of NGA, DC samples,

unpublished results.26. M. Vagnini, C. Miliani, L. Cartechini, P. Rocchi, B. G. Brunetti, and A.

Sgamellotti, Anal. Bioanal. Chem. 395, 2107 (2009).27. A. F. H. Goetz, G. Vane, J. E. Solomon, and B. N. Rock, Science

(Washington, D.C.) 228, 1147 (1985).

594 Volume 64, Number 6, 2010