Field spectroscopy in geoarchaeology – A comparative investigation of quartz and quartzite quarries and stone implements

Introduction

The study of raw material use in archaeology provides a fundamental basis for interpreting past technological strategies, mobility patterns, and systems of resource organisation. Patterns of raw-material selection and tool production represents cultural preferences, functional requirements, and the spatial distribution of accessible geological resources.^1,2 Identifying and characterising the geological origins of lithic materials enables reconstruction of procurement strategies, evaluation of territorial behaviour and assessment of the scale and organisation of exchange networks. ³ Raw-material studies therefore contribute not only to technological analyses but also to broader anthropological questions concerning knowledge transmission, technological organisation, cultural identity and the choices these represent. ⁴

Recent analytical developments in petrography, X-ray fluorescence (XRF), and near infrared (NIR) spectroscopy, have substantially improved the precision with which lithic materials can be classified and traced.^5–8 The application of non-invasive spectroscopy in archaeology is becoming increasingly common.^9,10 However, previous research has primarily addressed the geochemistry of stone tool assemblages in relation to their spatial and temporal distribution.^5,11 Although spectroscopic methods have been applied to prehistoric stone materials in provenance studies, these have concentrated predominantly on chert and obsidian materials and their geological sources.^6,7,12–16

In the context of prehistoric northern Sweden, quartz and quartzite were of considerable importance as raw materials, as indicated by their intensive use in tool production. This article focuses on the use of quartz and quartzite as raw materials in the inland of Västerbotten, Sweden (Figure 1), from the Mesolithic (7000–4200 BCE) through the Neolithic (4200–2000 BCE) and into the Epi-Neolithic (2000 BCE–800 CE).^17,18OPEN IN VIEWER

Spectroscopic analyses of quarries and artefact materials have been conducted, as a proxy for past human movement patterns and/or interactions. The ability to assign prehistoric stone tools to specific geological sources and/or classify the raw material used for toolmaking, based on chemical composition, represents a significant advancement in reconstructing mobility, procurement strategies, and intra- and inter-group interactions. For such analyses to be effective in field settings, portable and robust instrumentation is essential, as potential quarry sites must be analyzed in the field. A necessary preliminary step is therefore to evaluate whether spectroscopic methods can reliably classify quartz and quartzite when comparing field and laboratory results. Once the methodology is validated, broader questions regarding prehistoric survival strategies can be addressed. In this way, spectroscopic analysis offers the potential to employ geo-chemistry as a proxy for multiple aspects of prehistoric human behaviour.

The study area (Figure 1) is located in the pre-mountainous woodland area in Västerbotten, Sweden. The geology of the Scandinavian Mountain range was formed through a collision between the plates of Svecofennian orogeny and North American craton, roughly 500 million years ago,^19,20 during which the North American craton pushed part of the sea bed on to the Svecofennian orogeny, resulting in a “folding” action of the North American craton that would become the Scandinavian Caledonides, covering a large part of what we now know as Norway and parts of Western Sweden. ²⁰ This folding resulted in variations of the bedrock, in the form of striations in a North-Northeastern to South-Southwestern orientation (Figure 2). Later glaciation periods would erode the mountain range to their current rounded form and redistributed rock-materials including crushed materials (moraine deposits), blocks and boulder erratics, south-eastwards along the river valleys towards the lowlands.^21–24 Sourcing dispersed raw materials obviously will be challenging.OPEN IN VIEWER

Materials & methods

Materials used in this study consists of quartz and quartzite objects from various pre-historic quarries and archaeological sites documented at the Swedish National Heritage Board. The 9 quarry sites (Table 1) that have been included in this study were analyzed in the field using XRF and MicroNIR. Readings collected directly from the exposed bedrock and boulder erratics surfaces, are hereafter referred to as quarry in situ samples. In addition, adjacent geological materials interpreted as fragments from the exposed bedrock and boulder erratics, were also analyzed and are hereafter referred to as quarry sub-samples. Quarry selection from previously known quarry sites (Table 1), was based on topographical position, focusing on the Malgomaj and Vojm lakes’ area, including the lakes’ outlets forming the Ångerman river.

OPEN IN VIEWER

Table 1.

A complete listing of samples used. Sites with the same L-number (as in database, Swedish Open Cultural Heritage) are in some instances considered as one site with multiple erratic boulders, the boulders have been divided into subcategories (as is the case for L1937:3913-2), in order to differentiate different erratic boulders.

Site	Quartz	Quartzite	Insitu sample	Subsamples	Archaeological samples	Note
L1937:4871	—	X	8	10	—	Exposed bedrock
L1937:9129	X	—	7	5	—	Exposed bedrock
L2020:10270	—	X	12	10	—	Exposed bedrock
L1937:4526	—	X	8	2	—	Boulder erratic
L1937:4473	—	X	2	8 a	—	Boulder erratic
L1937:4472	—	X	2	8 a	—	Boulder erratic
L1937:4546	—	X	3	8 a	—	Boulder erratic
L1937:3913	—	X	3	8 a	—	Boulder erratic
L1937:3913-2 b	—	X	3	8 a	—	Boulder erratic
L1938:8018	X	X	—	—	53	Stone implements
L1937:3085	X	X	—	—	229	Stone implements
L1937:3295	X	X	—	—	50	Stone implements
L1937:3269	X	X	—	—	103	Stone implements
L1937:3652	X	X	—	—	119	Stone implements
L1937:3765	X	X	—	—	176	Stone implements

Added to the data set above are 730 archaeological artefact objects retrieved during earlier excavations (referred to as the Archaeological Samples), stored at the Västerbotten Museum, Umeå, Sweden (Table 1). From these, five readings per object were acquired using MicroNIR (3650 readings in total).

Results

The 108 XRF measurements (Figure 3) showed a largely homogeneous dataset, with some outliers that were subsequently excluded from further PC-modelling. In addition, one reading (L1937:4526:1-2) displayed extreme elemental concentrations and was identified as an outlier and was removed from the dataset (Figure 3) to prevent distortion of model outcomes.OPEN IN VIEWER

The extreme elemental values exhibited by the boulder erratic sample L1937:4526:1-2 likely resulted from the instrument measuring a sulfidic vein within the analysed specimen, explaining its divergence from the other readings derived from this quarry. Deviations related to the L1937:4526 quarry, shown in the XRF-derived PCA (Figure 4), may indicate that this “quarry” consists of five separate boulder erratics of potentially different geological origins. The boulder erratic samples L1937:3913.8, L1937:3913-2.11 and L1937:4526, which group with samples from the L1937:9129 quarry (exposed bedrock), were likely measured on quartz-rich veins within these erratics. This is consistent with L1937:9129 indeed being a quartz quarry.OPEN IN VIEWER

A PCA model was developed on the remaining XRF data. The PCA score plot (Figure 4), PC 1 (32,6%) versus PC2 (18,4%), shows a relationship characterized by elevated Si and Al, both bed rock forming elements. The loadings show negative p1 and p2 values for Si, whereas Zr, with strong positive p2, pulls quartz in this direction as opposed to quartzite.

Several quartz samples from quarry site L1937:9129, and to a lesser extent the main group of sites L1937:3919, L1937:4472 and L1937:4546, exhibited high Zr and Rb content. Notably, L1937:3913-2 does not group with the other objects from the same general area (i.e. L1937:4472, L1937:4546 and L1937:3913). Overall, most objects group according to quarry affiliation as seen in the score plot (Figure 4). Object L1937:4472-1-2 with strong positive p1 values, did not conform to any of the groupings in the PCA score plot. Additionally, readings from quarry L1937:3913 ranged closely with the main group from quarry L1937:4546 exhibiting slight negative p1 values and positive p2 values. However, readings L1937:3913.8, L1937:3913-2.11 and L1937:4526.1.2-1, showed positive p1 and p2 values aligning closer to readings from quartz quarry L1937:9129 in the score plot.

The PCA models for both XRF (Figure 5) and NIR datasets (Figure 6) show less than 50% variance explained by the first component. For chemically homogeneous materials such as quartz and quartzite (Figure 3), this is expected.^34,36 This should be viewed as an indication of the trends in the data being nonlinear, providing a relatively high amount of information in the lower dimensions. In the case of XRF the low explanatory power in the PCA could also be explained by the variance being spread out over multiple overlapping element peaks. ³⁴ Despite low variance in PC1 and PC2, higher components (8 or less in the case of the NIR dataset) capture >90% of the cumulative variance in both datasets (Figures 5 and 6), confirming that the models maintain enough analytical data to show trends within the dataset.^40,41OPEN IN VIEWER

OPEN IN VIEWER

Figure 6.

Cumulative variance in the PCA model based on NIR data from the archaeological, in situ and sub- samples.

Based on visual inspection of patterns in the PCA score plots derived from the NIR data, the objects were tentatively assigned to different classes. This assessment was based on the three-dimensional PCA representation using PC1, PC2, and PC3 (Figure 7). The PCA scores revealed a pattern in which five distinct groupings radiated outward from the centre of the plot, providing an initial basis for class differentiation. Subsequently, these groups were labelled A-F.OPEN IN VIEWER

To further explore the underlying structure of the dataset, a t-SNE model was generated to enhance separation among the preliminary PCA-derived groupings and to support a more robust definition of sample classes (Figure 8). The t-SNE analysis performed using three factors; however, a sufficient two-dimensional projection is shown (Figure 8). This configuration allowed the relationships observed in the first and second dimensions to be compared with those present in the third, helping to evaluate potential overlap in higher-dimensional space.OPEN IN VIEWER

The proposed groupings and resulting classifications were examined by overlaying the samples site affiliations onto the t-SNE score plot, enabling comparison between the data-driven classes and their corresponding topographic distributions.

The PCA score plot (Figure 4) showed general agreement with the class assignments derived from the t-SNE analysis. When site information was superimposed onto the t-SNE representation (Figure 8), a strong relationship between site location and grouping distribution was observed. Classes C, C2, and F3 occurred exclusively at site L1937:3765, with the exception of two objects from L1937:3085 that fell within class C2. Class D was dominated by objects from L1937:3269, with only two samples from L1938:8018 sharing this classification. Class A included objects from both L1937:3652 and L1937:3269, while class C3 comprised only of samples from L1937:3085. Notably, in situ readings from the L1937:4871 quarry and related sub-samples exhibited a strong association with class B.

Although class C, C2 and C3 were initially interpreted as a single group in the PCA model, the t-SNE analysis (Figure 8) revealed clear separation among them. Similarly, classes F, F2 and F3 were first grouped together in the PCA results but resolved into three separate groupings in the t-SNE analysis. Further examination of the mean spectra for each grouping (Figure 9), displays distinct features in each class, clearly showing material differences.OPEN IN VIEWER

In Figure 9, Savitzky-Golay second derivative spectra are shown for the ten t-SNE classes, where large negative peaks indicate wavelengths of great diagnostic importance. Classes A, B C, D, E and F each display distinct spectral responses. Contrary to this, the main group C does not share spectral resemblance with either C2 and C3. Group F, F2 and F3 have common spectral characteristics with main features at around 1460 nm.

Inspection of the class-mean NIR spectra (Figure 9) reveals several distinct features that permit discrimination between the defined classes. Prominent bands at 1329 and 1336 nm occur within the ∼1300–1400 nm second-overtone region, consistent with vibrational overtones of O–H functional groups. Additional features observed at 1459, 1466, 1509, 1534, 1571, 1602, and 1608 nm are attributed to O–H and molecular H₂O-related overtones and combination bands, indicating varying degrees of hydration among the classes. ⁴²

In the class A, feature bands at 1088 and 1137 nm are evident and are attributed to Fe²⁺ electronic transitions, commonly interpreted as spin-allowed crystal field absorptions in Fe²⁺-bearing silicates. These bands have been widely reported in the NIR spectra of olivine and pyroxene and are characteristic of iron-rich igneous and metamorphic mineral assemblages, although minor overlap with weak overtone features cannot be excluded. Furthermore, absorptions at 1137 and 1199 nm observed in selected classes provide additional evidence for variable Fe²⁺ content across the class-mean spectra.^42,43 Class F with a peak at 1466 nm likely represents the same mineralogical family as Class F2 with a peak at 1459 nm, possibly with minor compositional or grain-size variations affecting band position. Comparing classes F and F2 to F3, suggests that class F3 represents a local variation of Class F or F2, possibly affected by a more progressed metamorphosis. Class C3 could be interpreted as an advanced metamorphose variant of Class C, possibly related to an increased alteration or Fe/Mg substitution due to local conditions. Because of a lack of recognizable bands in the mean spectra of class C2, this may represent a poorly defined form of quartzite dominated by unaltered or spectrally inactive lithologies like quartz. ⁴⁴ It is possible that with a broader spectral range and higher spectral resolution, will improve diagnostic information could be acquired from class C2.

The noise primarily affecting the C2 and D groups in Figure 9, and to a lesser degree F2 and F3, may contribute to the low explained variance in the higher order principal components. Notably, the noise in the F2 and F3 classes are below the zero line and as such does not affect the relevant diagnostic peaks relevant for identification of elements in the objects.

Discussion

The results of the PCA of the XRF geochemical data demonstrate clear patterns of quarry level separation among the sourced objects. In particular, quarries L1937:3912 and L1937:3913-2 could be distinguished from each other (Figure 4), despite their immediate spatial proximity in the landscape. Furthermore, samples from L1937:3913-2 exhibit a strong geochemical affiliation with those from quarries L1339:4526 and L1939:4871 in the PCA score plot (Figure 4). The XRF results also reveal that quarry L2020:10270 does not align with any of the other identified quarry groupings in the score plot. This distinct geochemical signature may prove valuable for future provenance studies of stone implements in the region. Consequently, any artifacts manufactured from this raw material can be directly linked to this specific source, which can now be reliably identified using XRF analysis.

The PCA of the NIR data (Figure 7) showed a complex, but useful grouping as it offered valuable insights into the material characteristics of the objects. The interpretation of the PCA score plot is complicated because of considerable overlap in the object groupings that only becomes apparent when presented in 3-dimensions. The clear patterns radiating outward from the centre of the plot shows the overall homogeneity of the dataset. However, slight variations in chemical composition and the presence of inclusions within the rock matrix result in subtle trends. These trends are expressed as radiating groupings extending along multiple PC dimensions. This observation is consistent with the relatively low variance expressed in the higher principal components, which likely indicates the presence of several overlapping yet meaningful compositional patterns within the dataset.

The overlay of site-association metadata onto the t-SNE results provides strong support for the interpretation that the visually identified groupings show meaningful structure in the data. In addition, constructing the t-SNE model using three factors demonstrates that these groupings remain coherent along the third dimension, indicating that the observed clusters are not anomalies of two-dimensional projection but represent stable patterns in the higher-dimensional space. Examination of the material classes within the t-SNE score plot (Figure 8) shows that site L1973:3765 displays a distinct pattern of raw material use compared to the other sites. The occurrence of raw material classes C, C2 and F3 is unique to this site. Similarly, site L1973:3269 exhibits a strong association with raw material class D. These site-specific patterns may be related to exploitation of specific localized raw material sources for these sites. Another possible interpretation is that the raw materials observed at sites L1973:3765 and L1973:3269 indicates their use as recurring seasonal basecamps over an extended period. The large variation in raw materials compared to the other sites, may show shifts in procurement strategies or changes catchment-areas over time. Despite these site-specific differences, a shared characteristic is evident across all archaeological sites included in the study, namely the consistent presence of raw material class E within the assemblages. This recurring pattern may indicate a common procurement tradition or shared mobility practices, possibly suggesting the presence of a single group of people (or closely related groups) occupying the area periodically, potentially on a seasonal basis. This distribution pattern may also show the presence of widely dispersed moraine materials across the landscape, which would result in a large and diffuse potential catchment area rather than a single localized source. Additionally, Class B material is present at all analyzed archaeological sites. This class shows a strong association with quarry L1937:4871 (Figure 8), an exposed bedrock quarry, suggesting a potential geographic linkage between the archaeological assemblages and this quarry location. However, this association should be interpreted cautiously, as the primary distribution of this material may extend across a broad region, potentially spanning 50 km or more from Southwest to Northeast (Figure 2). Consequently, Class B may represent a widely distributed lithology rather than a uniquely local source, and the inferred quarry connection remains provisional, pending further verification.

The Savitzky–Golay second derivative NIR mean spectra (Figure 9) emphasize diagnostically important wavelengths and reveal clear spectral differentiation among most t-SNE classes. Distinct spectral responses are observed for classes A, B, C, D, E, and F. The main group C shows limited spectral similarity to C2 and C3, indicating compositional heterogeneity between these groups. Classes F, F2, and F3 share prominent absorptions near ∼1460 nm, suggesting a common mineralogical affinity. The mean spectra display absorption bands at ∼1329–1336 nm, corresponding to second-overtone O–H stretching vibrations, and a series of features between ∼1459 nm and ∼1608 nm attributed to O–H and molecular H₂O overtones and combination bands. Variations in the intensity and position of these bands relates to differing hydration states and degrees of alteration among the classes. Class A and D is further distinguished by bands at ∼1088 nm and 1137 nm, characteristic of Fe²⁺ crystal field transitions in silicate minerals, indicating a relatively Fe-rich material. Additional features at ∼1137 nm and ∼1199 nm in classes A and C, suggests Fe²⁺ present in the minerals of the analyzed geological material.

The close correspondence between classes F (∼1466 nm) and F2 (∼1459 nm) is likely to show minor compositional or grain-size differences, while F3 may represent a locally modified or more metamorphosed variant of this group. Similarly, class C3 appears to be related to a more altered or metamorphosed expression of Class C. In contrast, the absence of specific diagnostic peaks in Class C2 suggests a quartz-rich, spectrally inactive lithology. The F-F3 and C-C3 classes underline the importance of understanding the geology of the study area, considering that these materials could be local variations of materials from the same origin. The Southwest to Northeast striations in the bedrock of the Scandinavian Caledonides (Figure 2) suggest that these variations of raw material may have a fairly large catchment area along this direction.

The results from the Savitzky-Golay transformed spectra for each t-SNE score group plot (Figure 9) show that there are spectral and chemical differences in material composition between each group. Given that we can identify differences in chemical composition in stone implements, then the spatial spread and variation in this material will be significant and can bring a different and detailed level compared to ocular classifications alone. When the t-SNE clusters (Figure 9) are projected onto the topographical map (Figure 10), a clear spatial pattern emerges. The sites L1938:8018, L1937:3085, L1937:3295, and L1937:3652 exhibit similar material signatures. In contrast, sites L1937:3765 and L1937:3269 display distinct material compositions that differ, both from the aforementioned group, and from each other. The quarry site L1937:4871 was included in the spatial projection to illustrate its strong correspondence with raw material class B. As shown in Figure 10, class B material is present across all sampled sites. Note that sites L1937:3765 and L1937:3652 are located further north along the Ångerman River. The Vojm Lake still exhibit a portion of the raw material from class B, associated with the L1937:4871 quarry further west along the catchment area of the Lake Malgomaj. This indicates either a broad spatial distribution of this raw material class or a naturally widespread occurrence of the source material.OPEN IN VIEWER

Conclusion

X-ray fluorescence provides reliable results for determining the provenance of quartz and quartzite and has shown to be useful for field-based comparison of stone materials during archaeological excavations. Its primary limitations are the requirement for trained operators and relatively long acquisition times. In contrast, NIR spectroscopy is highly user-friendly, completely safe to operate, and offers rapid, non-destructive measurements suited to large-scale archaeological projects. Although NIR is less chemically specific than XRF, particularly when considering the analytical techniques required for interpreting its spectra, it remains a powerful tool for high-throughput field analysis.

Sampling under laboratory settings should be conducted under controlled conditions. Field sampling, however, poses greater challenges due to variable environmental conditions. Furthermore, field adapted equipment usually has limited wavelength range and resolution but still produces useful information. In this study, field sampling proved to be challenging as moist, growth of lichens add to spectral variation. While such effects can be mitigated, they cannot be fully eliminated. Nonetheless, the non-invasive nature of both NIR and XRF makes them valuable for in situ sampling, and many cases these techniques may represent the only solution. Bringing the lab into the field, documentation and analysis of stone implements could easily be conducted and thus enabling an increased understanding of the archaeological context already in field.

The strong correspondence between the t-SNE clustering results and the associated site affiliation metadata (Figure 8), demonstrates the robustness and validity of the proposed approach. These findings indicate that the combined approach is capable of classifying geological materials at a substantially higher resolution than conventional visual assessment alone. Furthermore, the approach shows significant potential for supporting the interpretation of spatial and temporal patterns within archaeological contexts, thereby providing a quantitative foundation for future site-based analyses.

Supplemental material