Spectral- and Image-Based Metrics for Evaluating Cleaning Tests on Unvarnished Painted Surfaces

1. Introduction

The conservation field is a rich and dynamic profession that has continuously drawn on its interdisciplinary nature to find appropriate remedial treatments for the challenges faced by its practitioners. The surface cleaning of decorative or painted surfaces is one such challenge that occupies conservators [1]. In this paper, surface cleaning will be referred to as the selective removal of deposited soiling upon painted surfaces. In turn, soiling is used to imply atmospheric particulate matter that deposits itself on heritage surfaces [2].

Besides traditional options, the repertoire of surface cleaning treatments available to conservators features adaptions from fields such as green chemistry [3], chemical engineering [4,5], and physics [6,7]. In recent years, notable improvements have been made with gel cleaning [8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24], which is the focus of this work.

Despite these advances, most of these cleaning treatments are still evaluated in a somewhat subjective manner. Scores according to pre-defined criteria are assigned to results of cleaning tests, following an ever more established practice within the profession [10,25,26]. This evaluation method is relatively easy to apply, and can also be visualised as star diagrams to better comprehend outcomes [10,27,28,29,30,31].

On one hand, this practice reflects the inherent nature of conservation practice. Every object requires to be treated differently based on its context [32,33,34,35]; and as specialised professionals, conservators will evaluate risk and success based on their training and experience. From a stakeholder/representative perspective, the perception of what makes a successful cleaning treatment is highly subjective too, and the concept of cleanliness can be said to be values- and/or community-based, as defined by contemporary conservation theory [36,37]. From this point of view, it is possible that a fully objective evaluation is perhaps not achievable, nor wholly desirable. On the other hand, subjective scorings, albeit assigned by trained and experienced eyes, are hard to reproduce, and are reliant upon many contextual factors that might skew the judgment of the assessor in question. From another perspective, the test surfaces often used in cleaning studies are not usually homogenous (due to their complex nature), and this might introduce further bias when attempting to weigh cleaning test outcomes.

Cleaning studies in conservation often use analogous mock-ups to mimic a material surface under investigation [38,39,40,41,42,43]. In this context, mock-ups are a critical tool for conservators to explore the behaviour of a heritage object with respect to a trialled treatment, whilst respecting the integrity of the object itself. The properties of the mock-ups can be selectively tuned for the treatment (or object) under study, and the mock-ups can be consumed/reproduced according to need, or even reused for future studies on the evaluation of a treatment over time [44,45].

Surface cleaning evaluations typically select a number of salient criteria upon which to assess the treatments under consideration. Generally, these may include cleaning assessment factors such as: soiling removal efficacy, homogeneity, selectivity (i.e., lack of pigment pick-up), surface disruption (e.g., by swelling, or foaming), residue deposition, and surface appearance (with respect to colour and gloss) amongst others [10,17,25,29,30]. An example of user-dependent criteria is given in Table 1.

The above criteria are assessed visually as a function of change compared to a pre-determined surface condition by a conservator’s trained eye, but this introduces user-bias into any evaluation, regardless of the expertise of the assessor. For this reason, the conservator’s intuitive observations are often corroborated, with added empirical confidence, by measured analytical data. For example, optical and digital microscopy are usually employed to pick-up differences resulting from cleaning selectivity and surface disruption; colorimetry and glossimetry are used to measure changes in surface appearance, whereas SEM-EDX and FTIR imaging are used to monitor chemical changes in the surface, and residue deposition by treatments. Cleaning homogeneity and efficacy however remain based on estimates registered by the user.

Diagnostic imaging technologies (such as spectral imaging [46]) can provide opportunities to incorporate a high amount of empirical-based data that can act as a complementary aid for conservators in making more reliable, repeatable, and accurate assessments when assigning scores for novel treatment evaluations. Research efforts have already begun addressing this issue [45,47].

To address and overcome user subjectivity, the methodology presented in this article offers the choice of non-contact and/or non-invasive image, appearance and spectral measurements, which conservators can avail of (according to access to instrumentation). The methodology takes advantage of the image-based outputs generated by various diagnostic imaging techniques, which can then be processed in freeware, such as FIJI (ImageJ) [48], to provide semi-quantitative percentage scores, thus offering a common platform for comparison. By juxtaposing these techniques together, this work thus aims to contribute towards offering a robust, rigorous, and comparable evaluation framework for documenting surfaces before and after soiling removal tests.

The evaluation metrics developed in this study were based on optical photography and microscopy, appearance-based measurements (colorimetry and glossimetry), and imaging spectroscopy (hyperspectral imaging (HSI), micro-Fourier-transform infrared (µFTIR) mapping, and energy dispersive X-ray spectroscopy (EDX) mapping). Notably, this paper applies established post-processing techniques in imaging spectroscopy to the evaluation of surface cleaning. Spectral unmixing algorithms were implemented to identify and map soiled and cleaned surfaces. These algorithms classify pure or mixed materials based on spectral similarity [49,50,51], either using user-defined, known references or libraries (e.g., semi-supervised unmixing by PoissonNMF [52]), or blindly classifying them into groups (e.g., unsupervised unmixing by LUMoS [53]). A chemometric method for normalised difference image (NDI) mapping [54,55,56] of soiled and cleaned surfaces is also reported. Using principal component analysis (PCA) [57], characteristic wavelengths in the shortwave-infrared (SWIR) were selected to represent and then map soiling on the surfaces under the study.

2. Materials and Methods

Mock-ups. The case study presented in this article centres around mock-ups of Edvard Munch’s (1865–1944) monumental unvarnished painting, Kjemi (1914–1916, Woll no. 1227), situated at the University of Oslo Aula (1911) in central Oslo. This painting forms part of a frieze of eleven monumental artworks (oil on canvas) under study within the CHANGE-ITN and Munch Aula Paintings (MAP) projects [57,58,59]. The mock-ups have been designed to study a new cleaning technique – agar gel spray – for unvarnished, water-sensitive painted surfaces [60] as a response to the heavy soiling history of the paintings within the Aula [44,61,62,63,64,65,66], as well as to test the applicability and suitability of the imaging techniques mentioned above for the documentation of the cleaning trials.

Figure 1 illustrates one protein-bound and one oil-bound exposed ground, and one unvarnished oil paint mock-up (5 cm × 5 cm), formulated based on material analyses of Kjemi [62,67]. An in-depth description of the mock-up surfaces is provided in Table 2. They were naturally and artificially aged to emulate the mechanical and water-sensitive properties of the Aula paintings. Artificial soiling was applied following a modified methodology adapted for the Aula context [68,69]. The artificial soiling recipe used is given in Table 3. The soiling comprised of insoluble macroscopic particulates (iron oxide, silica, kaolin), microscopic soot particles (carbon black), water-reactive solids (Portland cement, type I), hydrophilic organics (gelatine powder, soluble starch), and hydrophobic organics (olive oil, mineral oil), all dispersed in Shellsol D40 mineral spirits. The mock-ups were soiled across three cycles of accelerated ageing (Table 2). Excess soiling was applied to facilitate soiling detection. The mock-ups before treatment (BT) were therefore slightly more heavily soiled than is currently observed on the Aula paintings.

Agar spray and cleaning solutions. Agar spray is a novel form of agar gel cleaning, whereby the agar in sol state is subjected to atomisation within a heated paint applicator, allowing it to be applied as a homogenous and ultra-fine layer onto a surface within a very short time over large areas. The agar spray technique tested was developed by Giordano and Cremonesi [72,73], and an extensive characterisation of the new gel has been carried out by Giordano et al. [18]. Its properties make it an attractive candidate for the cleaning of soiled, textured painted surfaces, particularly those of a monumental scale such as those like Munch’s Aula paintings.

Four aqueous cleaning solutions at various pH and conductivity values were selected as shown in Table 4. The selection was based on free-liquid trials (evaluated by naked eye and under optical microscope – assessments are available online as an externally hosted supplementary file).

Deionised water served as a control to investigate the effect of the gel application per se. Conductivity- and pH-adjusted waters were prepared for each type of mock-up class, following methodologies in the literature [68,74,75,76], and are referred to simply as adjusted waters throughout the remainder of the text. Monobasic and dibasic citrate chelating solutions (pH-adjusted with (i) NaOH; and (ii) NH₄OH, for comparative purposes, respectively) were selected to investigate chelating action, after being marked as promising for soiling removal in the Aula [17,77]. When using chelating solutions, a clearance solution (an appropriate adjusted water) was thereafter applied, as recommended by Stavroudis [78]. The pH and conductivity of used solutions were measured before use (deviating solutions were discarded and prepared freshly).

Agar gels were prepared with the different cleaning solutions, and applied in two ways: (i) as a direct spray [72,73], and (ii) as a pre-formed rigid film (prepared by spraying, as a milder modified form of cleaning). Clearance solutions were always applied as a pre-formed rigid gel (so that a double spray application did not interfere with interpretation of results). The gel was used at a concentration of 3% w/v, and sprayed evenly from a distance of 40 cm. The agar sol temperature prior to spraying was dependant on the room’s and mock-up surfaces’ temperature, and relative humidity. The experimental parameters measured during agar spray testing are given in Table S1. The gel was removed after two minutes, following recommendations by Giordano & Cremonesi [72,73].

Treatment evaluation metrics. A detailed description of the methodologies, including acquisition parameters, for each analytical technique is given in Appendix A. Table 5 below summarises the cleaning homogeneity and efficacy metrics, according to the range of the electromagnetic spectrum that they targeted, the multidimensional data types they employed, the concept through which they evaluated soiling removal, the equipment required to capture the data, and the post-processing steps needed to achieve results.

For each cleaning efficacy metric, a normalised difference was calculated according to Equation 1, where x is a measured property before or after treatment (BT, or AT, respectively), so as to show the percentage return to an unsoiled control’s surface property as a relative marker of cleaning success; the specific values used for each technique’s metric are described in Table 6, further below.

$$\mathrm{c}\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{{\mathit{x}}_{BT}-{\mathit{x}}_{AT}}{{\mathit{x}}_{BT}}}\times 100$$

(1)

All cleaning efficacy metrics were calculated in ImageJ, and then statistically treated, and plotted using the Microsoft Excel spreadsheet package. The results from the metrics were then fed into the scoring criteria listed below.

Because each painted surface has different cleaning requirements based on its inherent properties and condition [17], the evaluation process was adapted by weighting the scoring criteria to then recalculate a weighted average score out of 1.00. Based on discussions with the conservators who have treated and monitored the paintings over the past two decades, higher importance was given accordingly to cleaning efficacy and homogeneity (both 30% weighting), followed by selectivity (20% weighting) and then colour (15% weighting) and gloss integrity (5% weighting – gloss differences on the Aula paintings are only truly perceptible when standing directly underneath the paintings, which is not how the public typically views them). Calculation tables for the weighted scores are available online as an externally hosted supplementary file. Following what has become standard practice in cleaning studies in conservation [10,17,30,77], star diagrams were also used to illustrate the ideal cleaning candidate. The combination of weighted average scores with star diagrams permitted a thorough appraisal of the cleaning results.

3. Results and discussion

3.1. Metrics in Practice

The metrics aimed to semi-quantitatively compare treatment results to aid in decision-making in a reproducible and shareable manner, and to complement professional observations made by eye. The image- and spectral-based metrics enhanced this process through their visual inputs and outputs, examples of which are illustrated in Figure 2 below. The appearance-based metrics were calculated from point measurements (Table 5), and therefore did not have any visual inputs nor outputs in terms of images or maps. It must be kept in mind that each metric measured a different property or component of the surfaces or soiling, and thus efficacy scores needed to be interpreted accordingly. Data plots for cleaning efficacy per metric are presented in Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10. Processed images from photography, microscopy, FTIR, and SEM, and data tables for all the efficacy plots discussed are available online as an externally hosted supplementary file.

3.1.1. Image-Based Metrics: L*a*b*

The images serving as input data for the image-based metrics were captured using the visible (VIS) range of the electromagnetic spectrum. The image-based metrics may be considered non-invasive, non-contact, as well as more accessible and low-cost, in that the only required equipment is a DSLR camera and/or a microscope. Having been converted from RGB to CIELAB colour space for this metric, the L*a*b* images offer insights into the perceptual lightness (L*), and colours (as perceived by the human eye, a* and b*) of the mock-up surfaces. The L*a*b* images thus serve as an extension to the conservator’s eye. The metric in turn represents soiling distribution based on the pixel intensity values of the L*a*b* images. It should be kept in mind that the accentuation of soiling through pixel intensity depends upon the colours of the soiling and surface being investigated.

Cleaning efficacy results for images taken at microscale coincided better with visual observations than those taken at macroscale, which – although comparable in trend – tended to overestimate cleaning. This could have been possible due to the resolution and scale at which the images were captured, whereby macro-imaging might have not given a relatable representation of the distribution of soiling, based on the image processing method used. Furthermore, results for macro-photography were based on one mock-up (and not a set of triplicates, due to time restrictions), and thus did not cover the heterogeneity of soiling removal results. The remaining discussion therefore focuses on the microscale results.

Out of the possible L*, a* and b* images generated for the mock-up surfaces imaged in this study, it was noticed that b* images gave better indications of the presence of soiling. This is since this channel did not pick up surface texture, which was prominent in L* images (this could be taken advantage of, nonetheless, to reveal crack propagation, for example). Resultantly, the L* channel was found to be better suited for smoother surfaces, and the b* channel for rougher, textured surfaces.

Cleaning efficacy plots using L*a*b* images are given in Figure 3. The spray application superseded the pre-formed gel in terms of cleaning efficacy. However, this was only marginally so for the oil paint mock-ups, indicating that application method of the gel did not influence soiling removal for this mock-ups class, perhaps as result of the physicochemical means of soiling-to-surface binding.

For the pre-formed gel application, there was a notable increase in performance of soiling removal for the citrate/NH₄OH solution, compared to other cleaning solutions. This trend was less apparent with respect to the spray application, but the metric indicated that the same chelating solution removed the most soiling, nonetheless. Overall, the chalk-glue ground was cleaned almost twice as much as the other mock-up types, potentially reflecting differences in soiling-surface interaction.

Based on these observations, this metric offered reasonable representations of soiling removal efficacy, although it must be noted that the thresholding of the L*a*b* images can be time-consuming and user-dependant.

3.1.2. Image-Based Metrics: Skewness

The use of statistical moments (describing the shape of a histogram distribution) as indicators of cleaning efficacy is based on the concept that, as darker contaminants are removed from lighter substrates, the pixel values of the respective BT and AT images reflect this change accordingly, and thus an image’s histogram distribution can reflect the change taking place during cleaning. When considering a distribution’s skewness in a cleaning context, a larger fraction of darker pixels (i.e., darker contaminants from soiling) will skew an image’s histogram leftwards (negative skewness), whereas a larger fraction of lighter pixels (i.e., cleaner, lighter surfaces) skews the histogram rightwards (positive skewness).

The skewness metric has been successfully applied to stone cleaning [79] and paper cleaning [45], and hence was investigated for its promising application to the exposed grounds and unvarnished oil paint under study, as a fast and reliable measure. This metric was applied to the RGB micrographs, but could equally be applied to RGB macroscale photographs, as in the work by Charola et al. [79].

All three methods for calculating skewness were investigated during trials: they differed in how and which statistical descriptors were inputted for calculation [80]. Consequently, calculations indicated that Pearson’s moment coefficient best described the skewness of histogram distributions in this case, in contrast to the cited studies that used Pearson’s first and second coefficients. The selection was based on the coefficient that described the datasets best. Whereas only skewness was measured in this study, other moments such as the mean and kurtosis (flatness) of the distributions could have been additionally used to monitor changes after cleaning [79], and require further investigation for application onto unvarnished painted surfaces.

Histogram distributions of the micrographs are available online as an externally hosted supplementary file. The unsoiled mock-ups had a negative skewness (probably due to stray, darker pixels relating to canvas details, surface texture and irregularities), with a narrow distribution spread across lighter pixel values (Figure 2). The soiled ground mock-ups were characterised by less skewed and broader distributions across the centre of the plot, which narrowed and shifted towards lighter pixel values with an increased negative skewness (closer to the control’s) after cleaning (Figure 2). The soiled paint mock-ups contrarily featured positively skewed distributions centred over darker pixel values (attributed to the soiling, and chromatic changes during ageing when compared to the control), which became more positively skewed as a higher percentage of lighter pixel values were present after removal of darker soiling (the distributions remained, however, centred over darker pixels due to the inherently darker substrate).

Cleaning efficacy plots based on skewness are given in Figure 4. The efficacy plots generated through the skewness metric indicated that the agar spray application resulted in better soiling removal than the pre-formed gel, which matched visual observations. The poorer performance of both gel applications when applied to the oil paint (as opposed to the grounds) was notable, and might be explained by the fact that considerable chemically imbibed soiling was observed microscopically after cleaning.

When comparing the cleaning of both grounds, the metric seems to overestimate the efficacy for the half-chalk ground, which looked effectively more soiled than its chalk-glue counterpart on the basis of visual observations. Whilst the chelating solutions removed most soiling on the half-chalk ground and the oil paint, the adjusted water gave better results for the chalk-glue ground. In the latter case, the uncovering (or formation) of micro-cracks and other surface details by the chelating solutions may have resulted in an increase in darker pixels, which according to this metric will translate into less efficacious cleaning.

3.1.3. Appearance-Based Metrics: Colour and Gloss

The management of changes to the colour and gloss of an object’s surface is typically one of the fundamental priorities of the conservator during treatment [81,82,83,84], so as not to misrepresent the aesthetic value of the piece and detract from its creator’s intent [85]. These metrics are not novel in themselves, but rather complemented the scoring criteria selected to evaluate the agar spray treatment.

Hyperspectral data was used for the colorimetry metric. Thus, one single data acquisition simultaneously allowed for both the measurement of colour change as a result of soiling removal, and the application of the spectral-based metrics (introduced after this section). This saved on time, and rendered measuring colorimetric data a completely non-contact process (colorimetric measurements are usually taken by placing a colorimeter upon the surface of an object).

The cleaning efficacy plots based on the measured changes in colour (ΔE₂₀₀₀) are given in Figure 5. The plots indicated that the cleaning solutions applied by spray guaranteed a better return to the colour of the unsoiled control than the pre-sprayed counterpart. The chelating solutions once again afforded the most efficacious soiling removal with respect to colour change. However, the half-chalk ground fared most poorly of all mock-up classes, irrespective of cleaning solution. This was likely due to soiling-surface interactions, whereby remaining imbibed soiling detracted from the full appreciation of the cream- coloured ground. The pre-formed gel, however, did offer an effective option for promoting return to the surfaces’ original colours, especially with the citrate/NH₄OH solution.

For glossimetry, the metric returns the difference with respect to the control’s gloss, or if the change registered goes beyond that gloss measurement, the percentage of excess deviation beyond the control’s gloss. Cleaning efficacy plots based on glossimetry are given in Figure 6.

For the chalk-glue ground, both the spray and pre-formed gel resulted in glossier surfaces after treatment. Given the large extent of matte areas of exposed chalk-glue ground in the Aula, this is a consideration worth addressing. Keeping in mind the variability of the gloss measurements, the most viable options for the cleaning of this ground were the adjusted water, followed by the citrate/NH₄OH solution (both applied via spray).

With respect to the half-chalk ground and the chromium oxide oil paint, spraying the agar appeared to promote an improved return to the original gloss, with the chelating solutions delivering the best results.

3.1.4. Spectral-Based Metrics: HSI

The HSI metrics illustrated the degree of soiling removal across the VNIR and SWIR ranges, thus assessing soiling removal efficacy based on the physical properties and molecular composition of the cleaned surfaces, or soiling. An example of soiling mapping is shown in Figure 2 (bottom left). All spectral maps generated from HSI post-processing are provided in Figures S1–S16. One metric used spectral unmixing algorithms to map soiling removal, whereas the other one was based on the chemometric technique of normalised difference image (NDI) mapping. As a result of HSI setup and mock-up dimensions, the major advantage of these metrics was that they permitted the simultaneous mapping of soiling across the entire surface of many mock-ups using a single acquisition. The HSI method can therefore be considered as a non-contact, and relatively fast means of visual assessment.

Between the two spectral unmixing algorithms tested, the semi-supervised one gave the most reliable results, whereas the unsupervised algorithm suffered from some issues. These issues are presented and discussed in Appendix B. The remainder of the discussion here focuses on the semi-supervised algorithm.

The VNIR and SWIR spectral unmixing maps using the spectrum of the unsoiled reference as an endmember demonstrated that (i) all the chalk-glue ground mock-ups were cleaned to a high degree with the spray, (ii) all half-chalk ground mock-ups were cleaned to a lesser degree, and (iii) that the oil paint mock-ups gradually became cleaner on going from deionised water to adjusted water to chelating solutions. The milder cleaning effect of the pre-formed gel was also mapped.

The converse relationship between spectral maps for cleaned areas (based on the unsoiled control’s spectrum as an endmember) and for soiled areas (based on the soiled control’s spectrum as an endmember) illustrated that soiling removal can be mapped using either reference spectrum. In the VNIR range, maps using the soiling’s spectrum as an endmember reflected the contrary of the cleaned maps, i.e., they showed in a complementary fashion the presence of soiled areas, instead of cleaned ones.

In the SWIR range, indirect molecular markers of the soiling were detected [57]. These markers lend great use since soiling may not always be visible, e.g., where surface contrast is poor (dark soiling on dark surfaces), or where soiling/degradation products are not necessarily visible to the naked eye. However, in the case of the SWIR maps, false positives were detected in all cases, with soiling apparently present on unsoiled controls. In the case of the chalk-glue ground, the cleaner mock-ups were shown as the most soiled. This may have been due to the unmixing algorithm confounding spectral bands common to the soiling and the mock-up surfaces (related to the binder, for example).

For this reason, the semi-supervised HSI metric was best assessed using VNIR map pixel values (when using maps based on the unsoiled reference spectrum, however, SWIR map pixel values gave results that closely matched their VNIR counterpart). Cleaning efficacy plots based on the semi-supervised algorithm are given in Figure 7. The VNIR plots showed that using the agar spray with the chelating solutions, (i) over 90% of soiling was removed from the chalk-glue ground mock-ups, (ii) roughly 70% of soiling was removed from the half-chalk ground mock-ups, and (iii) about 80% of soiling removal was achieved for the oil paint mock-ups. The figures for the pre-formed gel were comparably lower, with the chelating solutions performing marginally better for all mock-ups.

In contrast to the spectral unmixing metric, the soiling removal maps generated through the NDI method provided more reliability with respect to detecting the presence or absence of soiling upon surfaces. This is due to the use of specific marker bands, which were significantly indicative of the materials being mapped (cf. Figure A1). However, this method required more expertise and time to achieve reliable results, as it implied carrying out multivariate statistical analysis on the datacubes collected. The semi-supervised unmixing algorithm is thus preferrable where advanced statistical expertise (or time) is lacking.

Cleaning efficacy plots based on SWIR-NDI maps are given in Figure 8. The NDI soiling removal efficacy plots matched with the observations made by eye, for both spray and pre-formed gel applications. The latter was marked as being less effective compared to its spray counterpart, irrespective of cleaning solution (except in the case of the chalk-glue ground). When the agar was sprayed, the maps and plots indicated the chelating solutions removed most soiling, particularly for the oil paint (in the case of the half-chalk ground, performance was also matched by the adjusted water).

The processed SWIR data from the NDI maps was further used to calculate a cleaning homogeneity score. As a complement to the cleaning efficacy metrics, this score was developed with the intention of empirically evaluating the degree of similar cleaning, i.e., homogeneity, achieved across the mock-ups’ surfaces. In digital image analysis, a parameter for measuring image homogeneity was described by Haralick et al. [86,87], for which a plug-in in ImageJ was developed by Cabrera [88]. This parameter is calculated from what is known as the grey level co-occurrence matrix (GLCM) of an image [89,90]. Since the pixels in the images of the soiling maps can represent soiling spread, the homogeneity parameter (also known as the inverse difference moment) from the GLCM was deemed an appropriate measure of calculation.

3.1.5. Spectral-Based Metrics: FTIR

The FTIR metric enabled an in-depth look at the cleaned surfaces at the microscale, providing imaging and spectral information from the mid-infrared (MIR) range. Cleaning efficacy plots based on µFTIR maps are given in Figure 9. The FTIR soiling maps illustrate the distribution of points where the soiling spectrum was detected. As a result, only those points where no soiling bands were detected were considered as clean, and this binary separation between soiled and cleaned could explain the lower percentages reported across the board when compared to HSI.

For the spray application, soiling removal was more effectively carried out on the exposed grounds than on the oil paint. This might reflect the nature of the soiling, which may have been more physically imbedded than chemically imbibed for the grounds. In terms of cleaning solution, the citrate/NH₄OH solution outperformed the other options, both for spray and gel. Whereas this improved efficacy gradually increased across different solutions with the pre-formed gel, less certain trends were reported for the spray.

3.1.6. Spectral-Based Metrics: SEM-EDX

A cleaning efficacy metric was developed using element mapping (SEM-EDX). One indirect, empirical observation that indicated the cleanliness of surfaces was the increased surface charging on AT surfaces during measurements, as there was less (carbon-containing) particulate matter to dissipate the electrons. Although the acquisition of data was slower than for HSI, the post-processing was relatively straightforward. This means that if chemical markers for soiling can be found, as with Al and Si in this case (due to their relative abundance within the artificial soiling), then the method could be a viable technique for monitoring cleaning. However, compared to all the other techniques discussed prior, the higher cost and significant operational expertise required to run such analyses, need to be borne in mind. For this very reason, only mock-up surfaces cleaned with citric acid in NH₄OH were mapped.

The element maps collected indicated that soiling removal was most efficacious from the chalk-glue ground surfaces, whereas about half as much soiling was removed from the half-chalk and oil paint. Cleaning efficacy plots based on SEM-EDX element maps for Al and Si are given in Figure 10. As a result of the nature of data collected for this metric, it was also possible to calculate that the agar spray removed on average three times more soiling than its pre-formed gel counterpart.

3.2. Evaluating Soiling Removal Scores Using the Metrics

The constellation of cleaning efficacy metrics, together with the cleaning homogeneity score, permitted a well-rounded assessment of the agar spray cleaning tests on the Aula mock-ups. Table 7 below shows an example of how the metrics compared to each other, and contributed towards the scoring criteria for the spray tests. The collection of scoring tables for all mock-ups are available online as an externally hosted supplementary file. The resulting star diagrams are presented in Figure 11 (overleaf).

For the chalk-glue ground and the chromium oxide green oil paint, the star diagrams indicate that the citrate/NH₄OH solution (having the largest pentagon) satisfied most cleaning criteria. The solution’s improved cleaning action was more evident when applied via spray than as a pre-formed gel, an observation that was already noticeable when evaluating the metrics and weighing in the results by eye. When applied as a pre-formed gel, the difference in base for the chelating agent was shown to be minimal, in fact for the oil paint, the citrate/NaOH solution was just about better.

The citrate/NaOH solution also seems to perform better for spray application on the half-chalk ground – however, it is interesting to note that at first glance the largest pentagons are those associated with deionised water, and therefore the action of the gel itself. Two implications follow from this: the first is methodological, in that combining the star diagrams with weighted scores is beneficial for dispelling vagueness that might result from having many polygons on one star diagram. For example, in the case of the star diagram for the pre-formed gel on the chalk-glue ground, the pentagons for deionised water and for the citrate/NH₄OH solution may almost seem similar in size – the weighted score however indicates the chelating agent’s improved performance, reflected also in the more central spread of its pentagram towards the prioritised criteria. The second consequence is more application-based. The cleaning solutions tested for the half-chalk ground were not as effective as expected, and the fact that the control solution of deionised water performed very similarly, or even better, than other tailored solutions implies that more experimentation is required to find a better candidate for cleaning.

The star diagrams also illustrate that cleaning homogeneity was best achieved when spraying the agar directly onto the surface, and the highest homogeneity was delivered by the chelator solutions.

In the end, the suggestions for more empirical-based evaluations in this paper, combined with the tools that have developed within the profession, can further improve decision-making for conservators when planning a treatment, including at monumental scale, such as in the Aula. The presentation of results in this format makes it easier to document cleaning trials, and can aid in discussion with other heritage or museum colleagues, such as curators, private owners, art historians, archaeologists, and so forth.

4. Conclusions

The multidisciplinary approaches presented expand on the means currently available to conservators for empirically evaluating cleaning tests within a treatment context. By measuring changes in the appearance and spectrally characteristic properties of the surfaces being cleaned, the metrics introduced aim to minimise on user bias. They offer a toolkit of methods, which are relatively straightforward to execute, and can promote the quality of documentation, as well as the cross-dissemination of results in the literature.

Two major forms of cleaning efficacy have been defined, namely, spectral- and image-based efficacies. The spectral metrics are rooted in data collectable from a broad section of the electromagnetic spectrum, providing insights on different element, colorimetric, and molecular features. Where access to specialised equipment for measuring spectral-based efficacy is limited (due to budget, time restraints, etc.), the image-based efficacy metrics offer a low-cost, reliable and effective alternative to surface evaluation. These scores, calculated using image processing methods, tend to be lower than spectral-based scores, and thus mitigate against over-estimating cleaning.

An empirical means of calculating cleaning homogeneity based on digital image analysis was also used in evaluating the cleaning tests, together with a simplified, non-contact means of measuring colour change using hyperspectral imaging. Moreover, the representativeness of each metric’s outputs was assessed by comparison to visual observations made by conservators. The newly proposed metrics were integrated as weighted scores with the star diagram evaluation system already established in conservation practice. Such an integration permitted a nuanced interpretation of cleaning tests by agar spray cleaning, comparing it to its pre-formed gel counterpart.

For exposed chalk-glue grounds and chromium oxide oil paint – two water-sensitive materials used by Edvard Munch in the Aula – spraying agar prepared with citric acid in ammonium hydroxide at a surface-tailored pH was evaluated as potentially the best candidate to date for efficacious and homogenous soiling removal. Notably, the pre-formed agar can be used to deliver milder, albeit slightly less effective, cleaning of fragile areas. It remains to be seen which cleaning solution is more appropriate for the half-chalk ground, and progress can be made once the nature of soiling-surface interactions on this type of ground is further elucidated.

It must also be considered that the Aula paintings themselves might be more fragile, and more water sensitive, than the mock-ups used in this study, and therefore further testing upon the artworks themselves is certainly required. This applies especially to areas that have poor surface adhesion, or have been previously consolidated [17]. For all the water-sensitive surfaces considered in this study, the effect of agar gel cleaning on other chemical forms of degradation, such as metal soap formation, has not been studied and should be considered. This is especially since developments in reflectance imaging spectroscopy for the detection of degradation products are picking up momentum [91,92,93,94,95].

Finally, the evaluation metrics themselves will benefit from further testing in other case studies to ensure their usefulness and applicability in the conservation profession. Further developments in the spectral imaging technologies and post-processing methods reported are certain to ameliorate the representativity of the metrics, which can be further tailored according to need in collaboration with conservators, conservation scientists and imaging scientists.