Differentiation of Species and Provenance of Palm-Leaf Manuscripts Using Fourier Transform Infrared Spectroscopy and Chemometrics

Abstract

For authentication and interpretation of palm-leaf manuscripts, material analyses are required that enable identification of specific characteristics of written artefacts. In this study, we apply infrared spectroscopy (DRIFTS) in combination with principal component analysis (PCA) as a fingerprinting technique for the analysis of eleven palm-leaf manuscripts. We demonstrate that manuscript-specific information is obtained and that a differentiation regarding the taxonomic species of palm leaves used for production and of their geographical origin in South and Southeast Asia is possible. The results show the potential of infrared spectroscopy for fingerprinting and authentication of written artefacts.

1. Introduction

Palm-leaf manuscripts are among the oldest and most important written artefacts of South, Southeast, and Central Asia, with dated evidence of their use from at least the 2nd/3rd centuries CE until today. They were used for dissemination and preservation of all kinds of texts (religious, belletristic, technical, legal, etc.), as well as for accounting and correspondence. Although palm-leaf manuscripts were largely replaced by handwritten paper, and later printed paper, by the late 19th century at the latest, they are still sporadically produced in the traditional way, mainly for sale to tourists in Bali or for religious purposes in a few Buddhist circles in Sri Lanka and mainland Southeast Asia [1].

Historically, leaves of the palm species Borassus flabellifer (palmyra palm), Corypha umbraculifera (talipot palm), and Corypha utan were mainly used to produce manuscripts. Other species, such as Corypha taliera, are sometimes also mentioned, but it is not entirely clear whether these were actually used [2,3]. For production of manuscripts, young palm leaves were harvested and cut to appropriate size. The leaves were then subjected to processing steps, which could vary greatly depending on regional customs, but often included boiling in water or other liquids, drying, e.g., in sunlight, grinding the surface and treating them with oils. They were then placed between covers, which were typically made of wood (sometimes richly decorated) and possibly wrapped in cloth for storage. In South India and Southeast Asia writing on palm leaves consisted of notching the writing into the leaf surface with a metallic stylus. Lampblack or charcoal powder mixed with oil was then rubbed into the indentations to make the writing more visible [2].

As palm-leaf manuscripts are made of organic materials, they are subject to natural decay processes. Fluctuations in relative humidity and temperature, exposure to light, insect and/or rodent infestation, unfavourable storage, and constant handling can lead to their rapid decay. For this reason, various conservation measures have been applied to increase the shelf life of palm-leaf manuscripts. These measures include treating leaves with spices such as turmeric and calamus to keep insects away, applying organic oils such as lemongrass oil to preserve flexibility of the leaves and repel pests, and fumigating with thymol vapours to limit fungal growth [4].

For scholars in the humanities to be able to analyse and interpret the manuscripts, it is necessary to include all available information. This concerns not only content of the manuscripts but also material properties and, of course, their origin. It is currently quite difficult to say anything about the latter, unless it is explicitly stated in the manuscript itself (which is extremely rare). Therefore, simple methods are needed to analyse these manuscript properties.

When studying objects of cultural heritage, it is of absolute importance to prevent or minimise the contact and risk of damage to artefacts. Approaches based on Fourier Transform Infrared Spectroscopy (FTIR), especially Attenuated Total Reflectance (ATR) and Diffuse Reflectance Infrared Fourier Transform Spectroscopy (DRIFTS), are powerful analytical techniques used for the analysis of various materials. These non-destructive methods are particularly valuable for analysing historical and archaeological objects, as they require minimal sample preparation and preserve integrity of the artefacts. Unlike ATR, DRIFTS analyses a comparatively large sample area, making it particularly well-suited to fingerprinting. This is particularly useful when examining materials with an inhomogeneous macrostructure, such as palm-leaf manuscripts. Although a stable and direct contact with the analysed object is necessary, this non-invasive method has proven to be a valuable tool in archaeometry [5] and the analysis of organic material e.g., herbarium specimens [6], cellulosic and textile fibres [7,8], painting canvases [9], colourants [10,11], and palm-leaf manuscripts [4,12]. Most research on palm-leaf manuscripts, however, focused primarily on processing, writing methods or ink analysis [12], while a comprehensive analysis of FTIR spectra of these manuscripts as analytical fingerprints has not yet been performed.

In order to utilise analytical fingerprints of biological samples, i.e., to uncover differences and similarities, analytical data are usually analysed using multivariate chemometric methods. Principal Component Analysis (PCA) is a quite common approach that utilises an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of linearly uncorrelated variables called principal components. This unsupervised technique is widely used in machine learning to simplify complex datasets by reducing the number of variables, while retaining the essential information, i.e., the main differences, contained in the dataset. This method is usually applied to identify patterns and facilitates data visualisation by reducing dimensionality. In addition to PCA, also supervised methods like Linear Discriminant Analysis (LDA), Partial Least Squares-Discriminate Analysis (PLS-DA), Support-Vector Machine (SVM), and Soft Independent Modelling of Class Analogy (SIMCA) are applied in chemometrics; however, by their nature, they require large datasets [13,14,15,16].

In this study, we analyse for the first time whether DRIFTS in combination with chemometrics is suitable for obtaining characteristic fingerprints of palm-leaf manuscripts, which can be used to classify them regarding provenance and taxonomic identity. This approach was chosen for its ability to provide mobile, non-destructive analysis, making it ideal for fingerprinting cultural heritage objects.

2. Materials and Methods

2.1. Manuscript Samples

Eleven palm-leaf manuscripts, all belonging to the collection of the Centre for the Study of Manuscript Cultures (CSMC, University of Hamburg), were used in this study as it was possible to determine their origin. They differ regarding their place of origin, material, number of pages, and age (probably all 19th/20th century). All manuscripts were assigned by experts through direct inspection to the species Borassus flabellifer or Corypha umbraculifera and origin of the countries India, Indonesia, Myanmar, and Sri Lanka (e.g., by description, annotation, or language). The complete list of manuscripts used in this study is presented in

Table 1. Overview of manuscripts analysed in this study. Assigned species and country/region of origin are presented as well as the number of DRIFTS measurements and the number of sampled pages per manuscript. Manuscript MS V was measured twice at different dates to enable technical replication.

Manuscript	Object No.	Species	Country	Region	Meas. (No. Pages)
MS I	CSMC, 1/2014	Borassus flabellifer	Indonesia	Bali	16 (4)
MS II	CSMC, 1/2017	Corypha umbraculifera	Sri Lanka	unknown	18 (5)
MS III	CSMC, 1/2018	Corypha umbraculifera	India	Kerala	22 (5)
MS IV	Private Collection, MS Tamil 1	Borassus flabellifer	India	Tamil Nadu	19 (5)
MS V	CSMC, T 12/2021	Borassus flabellifer	India	Odisha	10 (1) ×2
MS VI	CSMC, T 29/2021	Borassus flabellifer	Indonesia	Bali	15 (3)
MS VII	CSMC, T 32/2021	Borassus flabellifer	Indonesia	Bali	25 (5)
MS VIII	CSMC, T 33/2021	Corypha umbraculifera	Sri Lanka	unknown	25 (5)
MS IX	CSMC, T 34-1/2021	Corypha umbraculifera	Myanmar	unknown	25 (5)
MS X	CSMC, T 34-2/2021	Corypha umbraculifera	Myanmar	unknown	15 (3)
MS XI	CSMC, T add. 6/2021	Borassus flabellifer	Indonesia	Bali	25 (5)

with corresponding meta data. Figure 1 shows an image of a manuscript page together with approximate measuring points at which DRIFTS spectra were recorded.

2.2. Preparation of Oil-Treated Palm-Leaf Samples

To measure the effects of freshly applied oil, which is a typical treatment of palm-leaf conservation and production, lemongrass oil was acquired from PhytoChemia (LOT No: LG190422BLK5040). The certificate of analysis from the supplier states that nearly 70% of the oils content are the two isomers geranial and neral (3,7-dimethylocta-2,6-dienal). All other volatile constituents are listed with a maximum of 3%. The  oil was measured directly on aluminium foil as well as applied to authentic palm-leaf material prepared as writing support. The samples were dried for 5 days before measurement.

2.3. DRIFTS Analysis

To obtain DRIFTS spectra, a portable 4100 Exoscan FTIR spectrometer (Agilent Technologies (Santa Clara, CA, USA)) was used in Diffuse Reflectance mode. The spectrometer is equipped with a ZnSe beam splitter, a Michelson interferometer and a thermoelectrically cooled dTGS detector. It has a spectral range of 4000–650 cm⁻¹ and a spectral resolution of 4 cm⁻¹. A gold reference cap was used for background calibration. 256 scans were collected in one measurement. Measurements were taken on unwritten and written parts of the manuscripts.

2.4. Chemometric Analysis

Chemometric analysis was performed using R Statistical Software 4.2.2 [17] with extensive use of the package hyperSpec 0.200.9000 [18]. The preprocessing of spectra included reducing the spectral range to 800 to 1900 and 2400 to 3840 cm⁻¹. Hence, the area from 1900 to 2400 cm⁻¹, which can show peaks from carbon dioxide of ambient air, was removed for chemometric analysis. This reduced the number of analysed data points per spectrum to 1356. Furthermore, preprocessing included smoothing with a Savitzky–Golay filter (window size: 11; polynomial order: 2) and forming the first derivative using the prospectr 0.2.6 package [19]. This package was also applied for normalisation and scatter correction using Standard Normal Variate (SNV) transformation. Principal component analysis (PCA) was carried out with the prcomp() function (non-scaled, centred). For statistical testing, two-tailed, two-sample t-tests (function: t.test()with equal variances) were applied to the scores of the respective component of a PCA. Although we assessed all resulting principal components from PCA analysis, those with less than five percent of variance were not reported (which usually concerned the fifth and higher principal components).

3. Results and Discussion

3.1. DRIFTS Spectra

The spectra of eleven manuscripts were preprocessed as described and the mean spectrum of each manuscript is shown in Figure 2 for comparison. The spectra of the different manuscripts show variable absorption intensities, especially in the “fingerprint” area between 800–1900 cm⁻¹. Nevertheless, they show a typical and recurring pattern that is expected from biological samples and

Table 2. Infrared bands of the measured palm-leaf manuscripts with tentative assignments to vibrations of biomolecules. The assignment is made on the basis of the references given, which use one of the following methods: FTIR—FTIR in Transmission, ATR—Attenuated Total Reflectance Spectroscopy, DRIFTS—Diffuse Reflectance Infrared Fourier Transform Spectroscopy.

Wavenumber (cm−1)	Tentative Assignment; Wavenumber, Reference (cm−1)	Method
∼3400	$\nu$(OH) cellulosic fibres; 3335 [7]	ATR
	$\nu$(OH) cellulose; 3348 [25]	DRIFTS
	$\nu$(OH) acid/methanol; 3600–3000 [21]	FTIR
∼2930–2850	$\nu$(CH2), $\nu$(CHx) cellulosic fibres; 2850, 2900 [7]	ATR
	$\nu$(CHn) alkyl, aliphatic, aromatic; 2860–2970 [21]	FTIR
	$\nu$(CH2) asymmetric, symmetric methylene groups in fats and waxes; 2927, 2855 [24]	ATR
	$\nu$(CH) cellulose, 2902 [25]	DRIFTS
∼1736	$\nu$(C=O) ester cellulosic fibres; 1735 [7]	ATR
	pectin/hemicellulose; 1735 [24]	ATR
	$\nu$(C=O) hemicellulose/lignin/ferulate; 1738 [23]	ATR
∼1660–1592	$\delta$(O-H) adsorbed water/hydrogen bond hemicellulose-lignin; 1635–1621 [23]	ATR
	$\nu$(C=C) aromatic/cellulosic fibres; 1595 [7]	ATR
	adsorbed water; 1635 [7]	ATR
	conjugated carbonyl group of lignin; 1650 [24]	ATR
∼1515	$\nu$(C=C) aromatic/cellulosic fibres; 1505 [7]	ATR
	aromatic skeletal vibration/lignin; 1509 [26]	FTIR
∼1464, 1427	$\delta$(CH) cellulose; 1430 [25]	DRIFTS
	$\nu$(O-CH3) Methoxyl-O-CH3; 1470–1430 [21]	FTIR
	$\delta$(C-H) cellulosic fibres; 1420, 1455 [7]	ATR
	$\delta$(C-OH) alcohol/cellulosic fibres; 1455 [7]	ATR
	$\delta$(C-H2) scissoring/cellulosic fibres; 1475 [7]	ATR
∼1379–1365	$\delta$(CH) cellulose; 1372 [25]	DRIFTS
	$\delta$(C-H) cellulosic fibres; 1365 [7]	ATR
∼1175–1160	$\nu$(C-O-C) cellulose/hemicellulose; 1161 [26]	FTIR
	$\nu$(C-O-C) cellulose; 1163 [25]	DRIFTS
	$\nu$(C-O-C) pyranose ring; 1170, 1082 [21]	FTIR
	$\nu$(C-O-C) glycosidic/cellulosic fibres; 1105 [7]	ATR
	$\nu$(C-C) ring breathing/cellulosic fibres; 1155 [7]	ATR
	$\nu$(C-O) triglycerides; 1237 [24]	ATR
∼1077–1035	$\nu$(C-OH) alcohol/cellulosic fibres; 1050, 1025 [7]	ATR
	$\nu$(C-O) cellulose; 1059, 1033 [25]	DRIFTS
	$\nu$,$\delta$(C-O) C-OH; 1060 [21]	FTIR
	$\nu$(C-O) glycosidic bond of polysaccharides; 1031 [4]	ATR
	$\nu$(C-O) triglycerides; 1086 [24]	ATR
∼899	$\delta$(CH), $\nu$(glucose ring) cellulose; 897 [25]	DRIFTS

shows tentative assignments of main bands in the obtained spectra. They can mainly be assigned to polysaccharides like cellulose, hemicellulose, and lignin [4,20,21,22]. The observed differences in band intensities in the spectra could be due to different ageing of cellulose-containing manuscripts, as evidenced by a decrease in cellulose bands and an oxidation-related increase in bands assigned to lignin, carboxyl, or ester groups [23,24].

3.2. Evaluation of DRIFTS as Fingerprinting Method for Palm-Leaf Manuscripts

In order to assess the suitability of DRIFTS for classifying different manuscripts, it is first necessary to determine whether a reproducible measurement of manuscripts is possible, to what extent obtained fingerprints differ in positions with and without text, and whether the oil applied for conservation affects them.

3.2.1. Reproducibility of Manuscript Analysis

To evaluate DRIFTS as a fingerprinting technique for palm-leaf manuscripts, various spectra were measured on different pages of each manuscript. To determine whether manuscript-specific information is obtained, spectra of technical replicates of the same manuscript obtained on different days were analysed and compared with spectra from different manuscripts. Figure 3 shows scores of PCAs conducted on spectra of the same manuscript (A) and two different manuscripts (B). While spectra of different manuscripts are clearly grouped, this is not the case for the technical replicates.

In order to quantify this objective assessment, the obtained scores were also analysed using t-tests. In both cases, the first four components of the PCA explain around 80% of the total variance. For scores of the first four components of the PCA applied to the spectra of the technical replicates and different manuscripts, the lowest p-value obtained was >0.5 and <0.0001, respectively (see Table S1). This means that there are statistically significant differences for different manuscripts, while the technical replicates show no significant differences. This confirms the conclusions made from Figure 3 showing that reproducible analytical fingerprints of palm-leaf manuscripts are obtained.

3.2.2. Analysis of Different Locations on the Manuscript

An obvious difference within the manuscripts lies in distinct locations where text is or is not present. Comparing manuscripts based on their DRIFTS fingerprints would be much easier if there were no significant differences between these locations, meaning that spectra are mainly influenced by the palm material rather than the engraving and presence of carbon-based ink, especially since some of the manuscripts are almost completely inscribed. Hence, several measurements were carried out on written and blank areas of the manuscripts MS VI, VII, X, and XI and the obtained spectra were analysed with PCA. The results are shown in Figures S1–S4 in the Supplementary Materials. No clear grouping can be observed. However, the results of the t-tests (see Table S1), which was applied on the scores of the respective first four principal components, showed p-values below 0.05 for the first principal component of MS VII and the second principal component of MS XI. Hence, there could be an influence of engraving and presence of carbon-based ink to the spectra. However, the fact that significant p-values are obtained for only two of sixteen analysed principal components, which could also be explained by multiple testing, suggests that this influence is negligible.

3.2.3. Influence of Oil

The application of oils is a typical repeated treatment for the preservation of palm-leaf manuscripts. To investigate whether and to which extent oiling affects the obtained fingerprint spectra, lemongrass oil was applied to three dried leaf samples and spectra were obtained both before and after application of the oil. In addition, the pure oil was analysed. The results are shown in Figure 4 for pure oil and one of the palm leaves and in Figure S5 for the other two leaf samples. The spectrum of pure lemongrass oil shows prominent bands at 2969, 2924, 2863, 1674, 1447, 1380, 1192, and 1130 cm⁻¹, which can be assigned to citral [27] (mixture of isomers geranial and neral), the main constituents of the oil used. These bands are not visible when looking at the spectrum of the palm leaf after oil treatment (Figure 4) but can be seen to some extent when a differential spectrum is built from spectra before and after oil treatment. From this very simplified imitation of the conservation process, it can be concluded that treatment with oil does not significantly alter the spectrum of palm-leaf manuscripts, which greatly simplifies the application of DRIFTS as a fingerprinting approach for these manuscripts in practice.

3.3. Analysis of Species and Provenance

In order to analyse whether DRIFTS in combination with PCA can be utilised to group manuscripts according to specific properties, the differentiation regarding taxonomic species and geographical origin was investigated. Figure 5A shows scores of the first and second principal component coloured according to species of the palm used for manuscript production and using different shapes for different manuscripts. The first observation that can be made is that spectra of different manuscripts cluster closely together. This shows the reproducibility of the approach and the applicability of DRIFTS as fingerprinting technique for individual palm-leaf manuscripts confirming the conclusions of the previous section, especially with regard to the rather small influence of different measurement positions.

It is clear that the species of palm leaves used has a major influence on the variance of spectra, which is particularly evident in the scores of the first principal component, as spectra of Corypha umbraculifera show mostly positive values, whereas spectra of Borassus flabellifer show rather negative values. The p-values obtained for the first two components are <0.0001 (see Table S1), indicating that there are significant differences for the spectra of the two species. To identify the parts of the spectra which are important for species differences, the loadings were evaluated. They are shown in Figure 5B. It is recognisable that bands at around 2930–2850 cm⁻¹, 1736 cm⁻¹, 1515 cm⁻¹, 1175–1160 cm⁻¹, and 1077–1035 cm⁻¹, which could be assigned to CH₂ and CH, C=0, aromatic C=C, and C-O-C, C-C, and C-O vibrations of cellulose compounds are important for both principal components (see Table 2). This could indicate a different composition of the palm species with regard to cellulose components, which is also relevant for differentiation of soft and hard wood [25] and different fibre types [25].

The differentiation between the spectra of manuscripts from different species is even more apparent when scaling is applied in the PCA (Figure S6 in the Supplementary Materials). However, this normalisation makes it difficult to interpret the loadings.

In order to investigate the spectral variances within the two species in more detail and analyse whether manuscripts can be separated according to geographical origin, the respective spectra were analysed individually. The relevant scores of both PCAs are depicted in Figure 6. For the spectra of Borassus flabellifer (Figure 6A), the separation of manuscripts from India and Indonesia is partially possible using the scores of principal components three (p-value < 0.001) and especially four (p-value < 0.0001, see Table S1). However, there is some overlap between these groups, which could be attributed to similarity of individual manuscripts regarding other, unknown characteristics such as storage conditions, handling, and age of the manuscripts. Notably, the first two components, which account for the greatest variance, are also influenced by unknown characteristics of the manuscripts unrelated to their geographical origin. For the spectra of Corypha umbraculifera (Figure 6B), the manuscripts from Sri Lanka and Myanmar can be separated by the scores of the first principal component, as the spectra of the Sri Lankan manuscripts show negative values, while the spectra of the manuscripts from Myanmar show positive values (p-value < 0.0001, see Table S1). The spectra of Indian manuscripts show positive values for the second principal component, which distinguishes them from the spectra of manuscripts of other origins (p-values < 0.0001, see Table S1). However, the spectra of Indian and Sri Lankan manuscripts overlap, suggesting that they may share similar characteristics in terms of the aforementioned unknown properties.

In Figure 7, the loadings of the respective principal components are shown. Also, for the PCAs of spectra of individual species, relevant principal components, namely the fourth for the PCA of Borassus flabellifer and the first two for the PCA of the Corypha umbraculifera spectra, are mainly influenced by differences in bands at 2930–2850 cm⁻¹, 1736 cm⁻¹, 1660–1592 cm⁻¹ (only Borassus), 1175–1160 cm⁻¹, and 1077–1035 cm⁻¹ which can be attributed to differences in cellulose composition as described above. The reason for these differences could be due to different ageing, but also different treatment or degradation states of manuscripts.

The results of this study demonstrate that manuscript-specific information from palm-leaf manuscripts can be obtained with DRIFTS. Furthermore, the chemometric analysis of the manuscript spectra proved that inherent spectral differences could be observed for the taxonomic species of the used palm and the geographical origin, although a complete separation of the analysed groups is not always possible. This clear separation could be achieved by focusing the evaluation on specific differences using supervised chemometric methods, such as PLS-DA [13], which were not applied here due to the small number of analysed manuscripts. However, meta information of the samples could be inaccurate or even incorrect, as there is no record of the exact course of the manuscript creation and whereabouts. The inclusion of modern samples, where this information is known for certain, would be promising to understand the obtained spectra of palm-leaf manuscripts.

It should go without saying that the provenance is to be considered as a derived property, e.g., of environmental conditions or production methods. As these influences on this parameter are likely to be complex and diverse, it is not unexpected that the separation is not as straightforward for the provenance as for the taxonomic species. In general, we cannot rule out the possibility of seeing a separation in spectra that is due to some other, as yet unknown, influence that correlates with geographical origin.

Although the variance found in the PCA can be attributed to species and geographical origin of the palm-leaf manuscript, there is additional variance in the spectra that cannot be explained by these characteristics and some effects may not be anticipated as there is little scientific knowledge in the field of palm-leaf material analysis. However, further variance is expected as there are various differences in the composition of manuscripts, for example due to different environmental/storage conditions, production processes, or the age of the manuscripts, which affect the spectra obtained.

For other biological samples, it has been shown that different materials can be distinguished on the basis of peak ratios in DRIFTS spectra [7]. Due to the complexity of the spectra, this is not possible for palm-leaf manuscripts and an analysis of the entire spectral fingerprint is necessary for classification. Future research is needed to validate this conclusion and confirm the results obtained here, as well as comparing or combining DRIFTS with other analytical methods for classification.

4. Conclusions and Outlook

In this study, we have shown that infrared spectroscopy (DRIFTS) combined with principal component analysis is indeed a promising approach for profiling palm-leaf manuscripts, as the obtained spectra contain manuscript-specific information. This makes it possible to distinguish the manuscripts according to the palm species used in their production and their geographical origin. It is likely that supervised chemometric approaches can make this distinction even clearer, which should be tested in a next step with spectra from a larger number of manuscripts. Thus, the work presented here lays the foundation for a new approach in codicology to classify palm-leaf manuscripts according to specific material properties, which could also be applied in a similar way to manuscripts made of other materials.