Computer-Assisted Reading Tools Marks: First Step of a Promising Protocol

1.1. Context

The study of tool marks and rock-cut crafting practices across historical periods—including Antiquity, the Medieval period, and the early modern pre-industrial era—developed alongside the fields of building and technical archaeology. These archaeological specialisations originated in France, England, and Germany before spreading to Spain and Italy (Dessales 2017; Sapin et al. 2022). In France, the method of interpreting tool marks on stone blocks or rock surfaces, also known as traceology, was primarily advanced by Jean-Claude Bessac. Following the publication of a typology of quarry tools and their marks (Bessac 1986), Bessac shifted his focus in the mid-2000s towards rock-cut monuments (Bessac 2007). Building on traceology, he analysed the organisation of rock-cut worksites. In the context of rock-cut structure, he also used traceology to define the chaîne opératoire implemented for the shaping phase and for reuse of the structures as well as to study the man holding the tool. Concurrently, in England, Peter Rockwell produced an inventory of tools used by stone sculptors (Rockwell 1993). This foundational work was later augmented by various studies in collaboration with Will Wotton and Ben Russell (Wootton, Russell & Rockwell 2013). During the 2010s, Rockwell also shifted his focus to rock-cut monuments, publishing a study on Buddhist temples in India. Like Bessac, Rockwell examined excavation sites, concentrating particularly on unfinished works to gain insights into excavation processes (Dehejia & Rockwell 2016). Nevertheless, building archaeology—a branch of archaeology ideally suited to investigating rock-cut structures with the same approach used for buildings—has yet to be fully applied in this context (Bessac, Lamesa & Sciuto 2021; Sciuto 2023). Although buildings from historical periods were treated as stratified archaeological objects from the early 2000s, it was only in the early 2010s that research teams began considering the possibility of establishing phases of occupation at rock-cut sites. Notable examples include research on the site of Lalibela by a French team, which emphasised the changes made within and around the monuments over time (Fauvelle-Aymar et al. 2010; Fauvelle & Mensan 2019), and Marion Liboutet (2022).

In studies of Byzantine Cappadocia, rock-cut sites are predominantly dated using chrono-typology methods. Guillaume de Jerphanion was the first to implement this methodology in his seminal work, Les églises rupestres de Cappadoce. Une nouvelle province de l’art byzantin (1925–1942). This four-volume series mainly focuses on frescoed churches, using the style and iconography of each fresco to determine the relative chronology of the monuments. Since then, this approach has been extensively discussed and refined, primarily by French and German scholars (Thierry 2002; Warland 2013; Jolivet-Lévy and Lemaigre Demesnil 2015; Ousterhout 2017). Additionally, inscriptions, which are rare in Cappadocia, have also been employed to date the paintings within these churches (Xenaki 2011; Jolivet-Lévy & Kourtzian 2013). In 2010, forms and the presence of architectural elements began to be analysed to date the monuments themselves (Lemaigre Demesnil 2010). An essential aspect of chrono-typology involves establishing a fixed perception of the monument. This was the case for rock-cut structures in Cappadocia, which were formerly viewed as static entities. It is only recently that scholars have begun to investigate the changes within the structures themselves, employing building archaeology methods (Ousterhout 2017; Lamesa 2017).

Unlike the extensive studies on prehistoric rock art by scholars using 3D models to highlight overlay and work phases (see Robin 2015 for a thorough synthesis; Botica, Luís & Bernardes 2023), a fully developed methodology employing new technologies to interpret tool marks visible in building and rock-cut structures for historical periods is still lacking. It was with this in mind that we began our work with the Archaeological Unit of the University of Minho (UAUM).

1.2. Objective

In this paper, we propose the hypothesis that employing methods with computer-assisted information systems alongside approaches from building archaeology, which intersect with chrono-typologies based on the style of paintings and architectural features, will significantly advance the dating of rock-cut churches. Although it may not be possible to establish an absolute date for the monument, defining phases of carving will illuminate the phases of occupation, thereby enhancing our understanding of the monument’s use and subsequent alterations over time. By utilising a relational database, we can focus on qualitative data for a specific site or quantitative data for a particular area, and even extend this to a regional territory.

The experiment lasted one and a half months and involved an interdisciplinary team: Anaïs Lamesa, who specialises in rock-cut monuments and Geographic Information Systems (GIS); Natália Botica, an expert in information systems; and Paulo Bernardes, a computer graphics expert. Our work was structured around three main axes:

1. The first axis pertains to the study of the structure itself. To delineate the history of the monument, we need to understand its spatial structure and consider its integration into the surrounding landscape.

2. The second axis focuses on integrating these rock-cut areas into the historical sciences to place them within their socio-economic and political contexts. This involves understanding how and why these areas were excavated. Often overlooked by traditional excavation campaigns, rock-cut structures serve as primary witnesses to the historical narrative of the areas in which they stand.

3. The third axis concerns heritage and conservation, as the researcher analyses the monument at a specific time. It is essential to assess the condition of a structure while also proposing predictive solutions for its conservation and preservation.

However, as this paper aims to define a protocol, we will concentrate solely on one wall of the church for this case study and therefore focuses on the first axis.

1.3. Limits of the study

Given the constraints of limited time and the labour-intensive nature of data processing in the laboratory, we have opted to focus on developing a protocol rather than examining the church as a whole. This objective is included in our project RACCTURK that we are conducting at the University of Edinburgh (UKRI – Horizon Europe Guarantee, EP/Y028120/1). Consequently, we are concentrating on a specific wall, which allows us to manage the scope of data collection effectively. We are prioritizing a manageable and reproducible model of study, which can serve as a foundational framework for future expansive research.

3.1. 3D Model

For this experience, the photographed wall is in plain sight, lit by daylight; therefore, we did not have to take pictures for the photogrammetry in constrained spaces, like underground structures, or dark spaces. Two types of cameras were used for the photogrammetry capture: a Digital Single-Lens Reflex (DSLR) camera (Canon EOS 650D, 18 megapixels, equipped with an APS-C sensor 26.8 mm in diagonal) with a 18–55 IS II lens, using a 33 mm focal length, as well as a smartphone (Samsung S20FE). Targets were placed around the perimeter of the wall and distances between them were measured. A photo scale was also placed at the bottom of the wall. The ground sampling distance varied throughout the shooting due to partial destruction of the ground in the west, the presence of a tomb carved into its centre, and a bench in the east. As a result, it was impossible to maintain a constant distance between both camera and the surface of interest.

This first series was shot by Anaïs Lamesa. It consisted of 299 photographs taken with the DSLR and without a tripod. An 80% coverage area was chosen to ensure significant overlap between consecutive photographs, providing the necessary data redundancy and detail for accurate 3D modelling. In doing so, we followed the method explained in the report published online by Adrien Arles et al. (2013) A similar protocol is used by John McCarthy and also in the Archaeology Unit of the University of Minho (McCarthy 2014; Botica, Luís & Bernardes 2023). The time allowed for this first session was approximately 1 hour and 30 minutes.

The second series of shots consisted of 77 photographs and was also taken by Anaïs Lamesa using the smartphone and a selfie stick. An overlap of between 60% and 80%, depending on the accessibility of the parts photographed, was chosen (Arles et al. 2011). This range of overlap ensures sufficient data redundancy and detail for accurate 3D modelling, even in less accessible areas. The shooting session lasted 30 minutes.

The choice to use different cameras (a DSLR for 299 photos and a smartphone for 77 photos) was made in order to investigate the resolution of the 3D models in relation to the shooting time.

3.2. Relational Database

The second step was the creation of a relational database to record and cross-reference data collected in the field and through observation of the models. The structure of the database integrates data relating to sites, buildings, and the characterisation of spaces and architectural elements. This facilitates the organisation and comprehension of the data, while enabling faster and more efficient access to information.

The relational database, designed with integrity rules, ensures the prevalence of records and the integration and contextualisation of all elements, along with data consistency. All elements, attributes and descriptors are defined according to commonly used scientific terminology, and data structures were created to encompass all relationships between different entities. Information can be input, modified or deleted in a structured and consistent manner through data entry forms, with automatic attribute validation procedures and associated value lists, ensuring data reliability, standardised recording, and a reduction in typing errors. Using a relational database also makes establishing relationships between different entities possible, for example by associating buildings with their historical and geographical context, as well as with the spatial units that make them up. This relational structure provides a holistic view of architectural projects and their connections.

In fact, as in the archaeology of buildings, it is important to use the typological approach and the stratigraphic approach to best define the chronology, even if this is relative. For example, certain types of covering or paint play a role in establishing a time range (Lemaigre Demesnil 2010; Jolivet-Lévy and Lemaigre Demesnil 2015).

The relational database is also an essential tool for long-term quantitative analysis. Collecting data related to stone working, such as the identification of tool marks and the types of cutting used and located in the monuments will enable us to perceive recurrences or breaks in practices. Cross-referencing all these data will also allow us to detect dynamics on a regional scale rather than one restricted to a monument or a site.

3.3. GIS

The third step was the creation of a GIS. As the goal of this method is to understand the history of the sites where the structures are located through their successive occupations and the evolution of their context, the spatial dimension is an integral part of the historical reflection. It is important to emphasise here that rock-cut structures can easily be transformed, changing their internal spatial organisation. Similarly, the sites themselves have several buildings that need to be understood as a whole.

Generally, data collected for the 3D model is geo-referenced thanks to GIS (Katsianis 2024) selfie stick and targets, even when the data is collected underground. Within the excavated space, it is necessary to recreate a system of local coordinates linked to external topographic points (Arles et al. 2011). The geomorphological analysis of an area, the position of the structure in a valley and the articulation of the structure with other surrounding structures offer a better understanding of the occupations of the landscape, their density and their distribution in a given area (Crow et al. 2011).

By combining it with a relational database, GIS can reveal points of analysis and patterns that are not apparent, such as the use of an unusual tool that marks a particular practice or a morphology that signifies structures both in time and space.

Finally, GIS software, applied at the scale of individual walls with tool marks, is a powerful vectorisation tool that can be used to automate the definition of stratigraphic units.

4.1. Creation of the 3D model and related data

The photographs captured with the DSLR and the smartphone were processed using Agisoft Metashape 2.0 Professional Edition. The processing was performed on a workstation equipped with an Intel® Core™ i7-5820K CPU @ 3.30 GHz, 32 GB DDR4 RAM, and an NVIDIA Quadro K4200 GPU with 4 GB GDDR5 memory.

The standard model creation workflow was applied: the images were processed to generate a tie point cloud, a depth map, and a mesh, which was then textured. The resulting DSLR model (hereafter referred to as Model 1) consists of 75,596,109 faces (at 4193.08 points/cm²), whereas the smartphone model (Model 2) consists of 8,123,271 faces (at 1379.37 points/cm²). The difference in resolution is therefore substantial. Since the models were created using a batch process, it was not possible to record the processing time for each model individually. However, given the hardware constraints, the subsequent processing clearly demonstrated that Model 1 was extremely heavy and difficult to handle, unlike Model 2.

Following the generation and texturing of the 3D mesh, the next step in this computer-assisted tool for mark reading was the production of a Digital Elevation Model (DEM), created for both models with frontal XY orientation. It should be noted that no geolocation data were collected, whether from the DSLR or the smartphone, and therefore the output files are expressed in local coordinates. No external ground control points (GCPs) or check points (CPs) were collected. The models were therefore processed in local coordinates and scaled using reference markers placed in the scene. Distances between the markers were measured directly in situ and compared with the corresponding distances in the 3D models. This procedure ensured internal consistency of scale.

From these two models, orthophotographs were generated as the next step in the computer-assisted reading of tool marks. To achieve this, once the two models were aligned, they were carefully cleaned to remove any geometric noise that might interfere with orthophotograph generation. Although the outputs differ greatly in resolution—11,457 × 15,116 pixels at 0.17 mm/pixel for Model 1 and 5,742 × 7,558 pixels at 0.324 mm/pixel for Model 2—the two orthophotographs initially appear quite similar. However, closer inspection shows that shallower tool marks, such as those made with a polka, are not clearly visible in the orthophotograph of Model 2. While the reduced size of Model 2 makes it easier to handle in the laboratory, this also compromises the visibility of finer tool marks (see Figure 2). For the remainder of the experiment, the orthophotograph derived from Model 1 was therefore preferred and designated as the reference orthophotograph. To further enhance the analysis of surface morphology and ensure that subtle features were captured, a dense point cloud was subsequently generated for both models.

Figure 2

The next step was the creation of an ultra-high-quality dense point cloud. This stage is essential to capture fine-scale surface details, particularly the shallow tool marks that might otherwise be lost in lower-resolution datasets. In addition to improving geometric accuracy through higher redundancy of 3D data, the dense point cloud also provides a versatile basis for subsequent analyses, supporting the derivation of multiple products such as DEMs, orthophotographs, and cross-sectional profiles. The point clouds were first cleaned automatically—using the Projection Accuracy and Reconstruction Uncertainty filters, both applied with thresholds corresponding to approximately 10% of the total number of pixels—and then refined manually to ensure integrity and usability. The dense point cloud of Model 1 has 4193.08 points/cm², compared with 1379.37 points/cm² for Model 2.

Because the models were initially processed in batch mode, it was not possible to record the time spent on every intermediate step individually. However, during the creation of the dense point clouds, the processing times were monitored separately: Model 2 was completed in less than two hours, whereas Model 1 required more than eight hours (see Figure 3).

Figure 3

4.2. Creation of the relational database

Our relational database was based on the 2ArchIS information system, developed by Natália Botica at the Archaeology Unit of the University of Minho. 2ArchIS is a modular information system that integrates data from archaeological sites, contexts, and excavations, designed to be reproducible and interoperable. The relational database was implemented in MySQL, and a back-office application was developed using HTML and PHP to run on a web platform, serving as the user interface for data management, listings, and parameterised queries. This web interface allows all users to access data on their personal computer or mobile device using a browser, ensuring available and up-to-date information for all users based on predefined usage profiles.

The 2ArchIS database has 3 main entries: a territory entry, a site entry subdivided into two entries: excavations and rock art, and a documentation entry (Botica, Luís & Silva 2022).

The rock-cut structures remain in their original context, so the information system, in addition to characterizing the site, architectural, and archaeological entities, should include data that integrates them into landscape studies to deepen the understanding of these structures. Thus, in the territory module, the characterisation of landscape stratigraphy is included, specifically from the perspective of natural layers (geographical and geological context, hydrography, and other natural features).

For the site entry, we had to adapt it to the specific needs of analysing rock-cut and built structures by creating a third subdivision. The data collected had to be processed by associating spatial and architectural entity data tables with the “building” table. To do this, we adopted two methodological approaches to the archaeology of buildings, those proposed by Daniel Arroyo-Bishop and Lantada Zarzosa (1993), and that proposed by Ivan Lafarge at the archaeology unit of the Département de Seine-Saint-Denis (unpublished).

A built or rock-cut wall must be included in larger data sets, but it also bears the marks of smaller data sets. The first step is to define the stratigraphic unit (SU), the smallest unit on which the stratigraphic reading must be based (called, in the case of building archaeology, the constructive unit, considered as positive SU). We prefer to use Stratigraphical Units (SUs) instead of context which has several meanings in English.

In the case of buildings, a positive SU is a coherent whole (a set of blocks of the same dimensions and/or of the same geological nature, or even just one block); for rocky walls made by human hand, it seems impossible to single out each tool mark, as there are too many on the same wall (Figure 4). We have therefore preferred Ivan Lafarge’s choice of archaeological unit in the Seine-Saint-Denis department, taking the homogeneous unit as positive SU.

Figure 4

In our case, SUs were determined thanks to the technique of traceology, that is, the method used to read tool marks: one stratigraphical unit is a set of tool marks made by the same tool and having the same direction. This technique, which involves interpreting marks on rock surfaces, assists in identifying and cataloguing tool marks. In order to define this set of marks, it is important to focus on the morphology of the impacts, their depth, their spacing, etc. Another type of SU is the natural hazards visible on the wall. In the case of the church G4b, big inclusions (larger than 10 cm) had impacted in the process of carving the structure, therefore, their marks on the rock had to be considered as SUs.

Once this stratigraphic unit had been defined, we adopted the organisation proposed by Arroyo-Bishop and Lantada Zarzosa (1993), i.e., the architectural unit (e.g. a wall or opening), the spatial unit (e.g. a room or corridor) and the structure (building or pseudo-building, which may be mixed, rock-cut or built). However, in the context of specific terrain such as Cappadocia, the sites where we are working must be included as a step between the structure and the site; therefore, we have called it “envelope”, i.e., the whole in which the structure is contained. Cappadocian rock-cut structures can be carved into cliffs (as at Ihlara), fairy chimneys (as at Göreme) or blocks detached from the cliff (as at Mavrucan) (Figure 5). In some cases, the site may include structures excavated in both fairy chimneys and cliffs. Identifying the envelope and characterising it is necessary to integrate the structure into its landscape. By default, the envelope for the built structure is the ground. This choice is explained by the modularity of the database. Excavations can be carried out before, at the same time as or after the study of the built structure. In this case, the soil is the envelope of the excavation within the database. Finally, the last level is the site itself.

Figure 5

The third entry in the database concerns documentation collected in the form of publications, maps, scanned documents, drawings, photographs, 3D models and videos.

The structure of the database constitutes its architecture. However, in order to populate it with data, it is necessary to define the attributes that characterise these entities, as well as the descriptors associated with them, in order to use controlled vocabularies, a fundamental task for data standardisation and interoperability (Botica, Luís & Silva 2022).

As illustrated on the SU form (Figure 4 above), most of the attributes used to characterise the SU have a list of descriptors in the form of terms to choose from (the down arrow opens a box with choices). The use of a controlled vocabulary mainly avoids clerical errors; the standardisation of descriptors also prevents the use of different terminologies for the same concept. When searching within the database, all the data relating to the same concept can therefore appear without omission.

Stratigraphic relationships are also recorded in the database, allowing them to be visualised in a table (Figure 6) or even in a graph by exporting these relationships to an application such as “Le Stratifiant” (Figure 7). The same process can be used to establish relationships between buildings or spatial units.

Figure 6

Figure 7

4.3. Computer-aided tracing

The open-source software QGIS 3.28 Firenze was chosen to create the GIS.

Initial analytical processing was carried out using CloudCompare open-source software v2.12.4 (Kyiv) and Dense Point Clouds, in order to use the Morphological Residual Model (MRM) method (Pires et al. 2014), with the PoissonRecon plugin created by Misha Kazhdan. The Dense Point Cloud for model 2 was created easily, which was not the case for model 1, which took CloudCompare several minutes to process. Despite this issue, details of model 1’s Dense Point Cloud explained that we only used it for further processing.

Using the MRM gives a clear image of the tool marks when they are deep. Shallow impacts such as those made by axe do not appear in this modified image. Keeping this aspect in mind, this first process with MRM came with good results as it enabled to observe a cross that had not been identified in previous publications.

Images processed with CloudCompare, orthophotos and the DEM created with Agisoft Metashape (now referred to as rasters) were proceeded in the GIS using the GDAL tool to assign the WGS 84 UMT 36N (EPSG 32636) projection to all images. We also defined the SCR of our QGIS project based on the projection chosen for the rasters.

In the rock-cut structure, in order to observe the tool marks properly, it is recommended to use raking light in a circumscribed area. Applying this method to our DEM, we used the shading symbology to highlight the relief of the wall: after several attempts, we decided to set the altitude at 90° and the azimuth at 315°, with a Z at 4, the layer renderings by default (Figure 8).

Figure 8

We carried out similar processing on the MRM image and applied a grey band symbology with the following calibration: Brightness 178, saturation 0, contrast 78 and gamma 1. We also applied a colour inversion. The choice of colour inversion allows the tool marks to appear more clearly (Figure 9).

Figure 9

Once the rasters have been processed in the GIS, it is possible to vectorise each SU, by drawing it with the ad hoc QGIS tool (Figure 10). We used all implemented rasters to vectorise SU. The orthophoto, for example, provides a clear view of the marks of flat-edged tools such as axes, while the DEM and the modified MRM raster are useful for the marks of pick-type tools and clearly see their direction.

Figure 10

Once the vectorisation process was completed, the results of the tool marks reading were entered into the database. It was necessary to create the site, rock-cut and building entities and then the stratigraphic units. For the north-west wall of the G4B church, we defined 26 SUs, giving them their own identity, a brief description and, where necessary, a physical relationship with the other SUs (Figure 11).

Figure 11

A final action was the connection between the relational database and the GIS. To do so, we used XAMPP v3.3.0, a package with the open-source servers to run Apache and MySQL database and QGIS. The first step was to launch XAMPP and then connect the database to QGIS by means of a join between the vector layer created for the SU and the SU table.

4.4. Production of stratigraphic trees

As explained above, the database makes it possible to relate the various SUs to each other, chronologically or physically. Le Stratifiant is an application that enables archaeologists to draw automatically stratigraphic trees (Desachy and Djindjian 1991; Desachy 2016). It is then possible to create stratigraphic trees for phase excavations and/or recovery work. To begin with, you need to create a view table using the MySQL administrator command box. This view is used to obtain all physical relationships between the SUs (Figure 12).

Figure 12

Once data has been extracted into a table, the relationships need to be translated into French and some relationships like “on” have to be removed. Once the table is cleaned, it can be copied and pasted directly into Le Stratifiant’s “data” tab, where each SU is also identified (Figure 13).

Figure 13