Ancient Hermione revealed: the contribution of high‐performance computing and digital methods to the analysis of a hidden cityscape

This article explores the potential of combining high‐performance computing techniques and a set of integrated digital methods to investigate the cityscape of ancient Hermione, Greece. Unmanned aerial vehicles (UAVs), terrestrial laser scanning, image‐based modelling techniques and high‐performance computing have been combined to provide a fully‐three‐dimensional (3D) representation of the city landscape, which encompasses both the topography and those still visible archaeological features, which are nowadays annexed into the modern buildings. As a consequence, the resulting geo‐located digital platform is now opening up interesting opportunities for research, such as the possibility to analyse spatial interconnections between sacred buildings, to formulate hypotheses about their location and to put them in comparison with the accounts made by historical sources. By taking advantage both of an entirely‐3D reconstruction and the analytic tools provided by geographical information systems (GISs), more sophisticated analyses can now be performed and specific issues such as visual perception and movement to and from prominent buildings/spaces can now be investigated.


| INTRODUCTION
This contribution presents the results obtained in the frame of a project related to the multi-disciplinary investigation of Ancient Hermion in Greece. The first part of the article provides a general overview of the combined use of digital methods in the study of ancient urban contexts, pointing out main achievements and critical aspects in connection to the different methods applied.
In section 2 the study context is described. In section 3, the methodological framework described is divided in sub-paragraphs presenting the different methods of data collection and analysis, which include unmanned aerial vehicles (UAVs) and highperformance computing (HPC) platforms. In the last section (section 4), the obtained results are presented and discussed. In this context, more long-term research goals are also introduced.

| Digital methods in urban contexts
Conducting archaeological survey in urban contexts has always been a quite difficult task even when advanced acquisition methods were employed. Among the non-invasive techniques involved, aerial photography and satellite remote sensing have been widely used and discussed (Campana & Francovich, 2003;Kaimaris, Karadedos, Georgiadis, & Patias, 2018;Kaimaris, Patias, & Georgoula, 2017;Mozzi et al., 2016). Geophysics has proven to be one of the more successful and often the only applicable one, due to the presence of physical obstructions and barriers that characterize the urban environment and which prevent the adoption of more traditional approaches (Basile et al., 2000;Papadopoulos, Sarris, Yi, & Kim, 2009). Datasets derived from geophysical prospection have been sometimes employed as a geometrical reference to make a virtual three-dimensional (3D) reconstruction of the urban landscape (Klein, Vermeulen, & Corsi, 2012). Additionally, geophysics when properly integrated with aerial photography and historical cartography can provide important information to detect new archaeology (Verdonck, Vermeulen, Corsi, & Docter, 2012). The recent advances made in digital archaeology now allow archaeologists to combine different multi-resolution techniques with the purpose of investigating complex urban environments. This raises a broad discussion about the definition and implementation of best-practices to be employed in fieldwork (Vermeulen & Corsi, 2015). The successful integration of ground-based prospecting techniques and remotely sensed satellite data also allowed for a broader and deeper overview of the landscape and its palaeoenvironmental evolution in a diachronic perspective (Keay, Parcak, & Strutt, 2014). Another field of application is the monitoring of threatened archaeological heritage, where ground-based prospecting techniques and advanced data post-processing operations have been combined to support archaeologists and planners to identify and prioritize those areas more likely to be damaged by new development (Neubauer et al., 2012).

| A 3D-based approach
More recently, significant technical advances made 3D technology and 3D-based multi-resolution sensors a valuable option to perform high-detailed on-site documentation, making it possible to acquire even extremely complex urban contexts such as the Mayan city of Copan, which could be analysed and communicated through an interdisciplinary, collaborative approach von Schwerin et al., 2013). Analogously, laser-scanning and imagebased modelling techniques have been employed to digitally acquire portions of still preserved Roman cities (Barsanti et al., 2012(Barsanti et al., , 2013Fiorillo et al., 2013;Guidi et al., 2009) so as to produce high-resolution 3D models to be used in some cases as a geometrical reference for reconstructing the architectural space as it was discernable in ancient time (Dell'Unto et al., 2013). An additional refinement in the field of 3D acquisition methods is constituted by the massive increase of air-based computer vision techniques, which was made possible by the widespread of lowbudget aerial platforms, such as UAVs. These allow for a faster and more accurate documentation of large archaeological landscapes and urban complexes (Ostrowski & Hanus, 2016;Smith, Passone, Al-Said, Al-Farhan, & Levy, 2014;Stal, Lonneville, Nuttens, De Maeyer, & De Wulf, 2014;Verhoeven, 2011). There is undoubtedly great potential in the use of 3D models and they are deemed to play a major role in the archaeological documentation process although more experiments need to be conducted. Indeed, as some recent scholarship has proven (Earl, 2007;Opitz, 2017;Paliou, 2011;Paliou, 2013;Paliou, Wheatley, & Earl, 2011), threedimensionality has a clear 'heuristic' value in the way it can dramatically improve the quality of data interpretation through the adoption of statistically-oriented spatial analysis techniques, combined with a highly accurate representation of the space in all of its dimensions. Research conducted at Lund University has recently proven the potential of 3D technology in combination with a geographical information system (GIS) to analyse and interpret archaeological data Landeschi, 2018;Landeschi et al., 2018). The results obtained both on a single site and on a landscape scale provided archaeologists with interesting solutions to address different types of questions, concerning: (a) the perception of the ancient space ; (b) the structural degradation of ancient architectural structures (Campanaro, Landeschi, Dell'Unto, & Leander Touati, 2015); (c) the damage evaluation in the context of a site threatened by human action (Landeschi, Nilsson, & Dell'Unto, 2016). On a broader scale, heuristic 3D models have been employed for example in the analysis of the Mayan city of Copan, and the virtual reconstruction provided the geometrical reference to measure visibility and accessibility as a means for better understanding how the spatial settings were defined by different socio-political groups in the city (Richards-Rissetto, 2017).

| STUDY CONTEXT AND PREVIOUS RESEARCH
The location of ancient Hermione (today's Ermioni in the Argolid) has long been known through the lengthy testimony of the ancient travel-  Frickenhaus & Müller, 1911;Jameson & Jameson, 1950;Pharaklas, 1973) and archaeological work, notably in the extensive necropolis of Hermione, was carried out in the 20 th century (Philadelpheus, 1909;Papadimitriou, 1994;Spathari et al., 1991). 1 Furthermore, the remains of a large temple were investigated by A. Philadelpheus and M.H. McAllister (Philadelpheus, 1909;McAllister 1969) and a significant corpus of inscriptions was collected and published in Inscriptiones Graecae IV (cf. Jameson, 1953Jameson, , 1959Stamires,1960  In addition, during the last field campaign more GCPs were acquired throughout the city as we had the possibility to employ a more sensitive RTK device (Leica GS18 T) that is capable of getting the correction signal even in locations with poor sky visibility and is completely immune to any electromagnetic interference. density. This approach allowed us to generate a single 3D mesh that could be geo-located based on the points located close to the church area, and which subsequently provided the right position for all the connected archaeological features (Figure 4). In connection to this laser scanner acquisition, the remaining part of the streets which included archaeological objects to be documented was acquired with an SLR digital camera (Canon EOS 550d) and the photographs were processed through image-based 3D modelling (IBM) techniques in order to derive a polygon mesh to be properly re-scaled and connected to the previously-described mesh. As a result, a set of 3D meshes was imported and properly georeferenced in 3D GIS. To deal with both textured and meshed 3D models and to add new layers as 3D vectors, ESRI ArcGIS was employed.

| 3D data capturing
In 2016 field season, two additional goals were set: the first one consisted of the digital acquisition of the topography of the entire peninsula of Ermioni and the subsequent creation of a high-resolution DTM; the second goal was the testing of an advanced supercomputing system for processing a large dataset of information. To completely acquire the whole peninsula area a UAV DJI Inspire 1 RAW with a Zenmuse ×5 digital camera was employed and different sets of photographs were acquired in order to cover the entire area and obtain a point cloud to be georeferenced and imported in GIS.
More laser scanner acquisitions were also performed in order to collect additional features to be integrated into the geodatabase.
F I G U R E 4 A portion of street connections, linking different archaeological features still visible in situ but hard to be georeferenced through GPS, has been digitally captured through the use of laser scanner and SLR cameras (a) and then visualized as a set of meshed 3D models in a GIS environment (b). Such operations allowed adjusting the position of the 3D models of the archaeological objects for which it was not possible to put down any marker (c). As reference GPS station, the area around the church of Aghioi Taxiarches (P) was considered [Colour figure can be viewed at wileyonlinelibrary.com]

| UAV-based data acquisition
As for the drone acquisition, a couple of dedicated applications (DJI and Pix4D) were installed on the tablet personal computer and connected to the remote controller in order to define the desired camera settings and flight parameters to be used. Ideally, the flight plan should be set up so as to cover the whole area to be modelled in the form of a regular grid, taking pictures with a sufficient level of overlap to make them usable for image-based 3D modelling. The total surface to be covered being 0.974 km 2 and with a drone's battery autonomy of ca 20 minutes, the best-suited option was to acquire the whole area in eight separate flights, which resulted in eight chunks of approximately 300 photographs to be processed ( Figure 5). Each flight height was defined trying to find the right balance between a reasonable distance between the camera and the ground surface and the need of keeping the drone within a safety distance from the more proximal buildings and/or any other physical barrier. As an average value, 62.9 m above the taking off surface was determined to be sufficient to capture details that could enable the reconstruction of a spatially accurate point cloud to be used as a basis for the DTM.

| Data processing: image-based 3D modelling reconstruction
Considering the huge amount of data collected during the UAVbased data acquisition campaign, consisting of 2265 raw images, it was essential to define novel methods for performing image-based 3D modelling in an efficient way. By taking advantage of highperformance computational resources available through the Lund University Center for Scientific and Technical Computing (LUNARC), the entire dataset has been processed and a high-resolution 3D model produced ( Figure 6). Agisoft Photoscan Pro software was installed to run a parallel processing on multiple compute nodes by F I G U R E 7 Different stages of the 3D model reconstruction: (a) the sparse point cloud was obtained from the camera alignment through the use of a structure-from-motion (SfM) algorithm and after the gradual selection and this consisted of 669 810 xyz points; as a following step (b) the dense point cloud, resulting from the multi-view stereo (MVS) matching that allowed to increase the number of points up to 233 639 780, has been georeferenced (c) with 38 GCPs previously acquired through the RTK GPS. Finally (d) a special filtering tool has been used to classify terrain surface points and separate them form the points belonging to buildings and vegetation cover [Colour figure can be viewed at wileyonlinelibrary.com] the possibility to set the numbers of nodes and the walltime for both the point alignment and the dense cloud reconstruction, which are typically the more demanding stages in terms of hardware performance.
After this process a dense point cloud was produced and the entire dataset was georeferenced based on 38 GCPs, acquired during the previous field campaigns (Figure 7). Ten additional check points (CPs) were then added in order to evaluate the overall accuracy of the model and such an operation gave us a total root mean square error (RMSE) of 26.18 cm, similar to the one observed and described by Jebur, Abed and Mohammed (2018), where an RMSE of cm 31, 27 and 29 resulted respectively from x, y and z coordinates (- Figure 8). As a result, the point cloud was filtered in order to keep only the points belonging to the ground surface and use them to build up the raster DTM (Figure 9). Since the priority was to reconstruct a high-resolution terrain model of the city area, special filtering tools available in Agisoft Photoscan were applied to remove redundant point data information, such as buildings, walls, roofs and vegetation elements. Some additional 'background noise' due to points not correctly triangulated in Photoscan were manually cleaned up at a later stage. As a final result, a DTM with a spatial resolution of 5.75 cm/pixel and a point density of 302/m 2 was obtained.

| GIS data implementation
To better manage all the spatial information collected during the field campaigns and to effectively use these datasets to analyse the archae- More datasets were then imported into the GIS. All the 3D models, irrespectively of technique employed for their production, are usually managed by the 3D Analyst module as 'multipatch' feature classes, which represent a file format that describes surface boundary F I G U R E 9 A dense cloud classification tool has been employed to automatically divide all the points into two classes so as to separate those belonging to the terrain from the rest of the features (above, the RGB point cloud, below the classified dataset). In this two-step process, points are assigned or removed from the 'ground' class based on a certain threshold angle that is considered the maximum slope value from the calculated terrain model for the examined area. If the angle value of an ideal line connecting a certain point to the reference terrain is higher that the value of the threshold angle then the point is not So far, the project on ancient Hermione has proven the potential of employing 3D data information to recreate lost topographic relations among archaeological and natural features that characterized the ancient urban space: this is an essential prerequisite to define and geo-locate areas of the city linked to specific functions, such as sanctuaries or public squares. As an example, the topographic location of the Roman aqueduct, in which the verticality plays a significant role for defining its original path through the landscape, can be more effectively addressed through a fully-3D data representation (Cambray, 1993), while a deeper understanding of the settings of artificial terraces as attested in other Greek cities (Bintliff & Evelpidou, 2002) can be more easily obtained through the setup of a spatially high-resolution DTM of the study area.
In a similar way, additional questions relate to the way those temples originally described by Pausanias were perceived from different areas of the city and which among the possible interpretative scenarios would better fit the ancient authors' account. Digitizing and reprojecting the supposed trajectory of the ancient city walls on the high-resolution DTM enables a better understanding of which points were more likely to be crossed by the wall and which areas indeed were included within their perimeter. This can also be useful to deduct those portions of land that were more exposed to external attacks and therefore more demanding of defensive structures.
Based on such premises we need to carefully re-discuss the role of three-dimensionality and its centrality in addressing questions on the ancient urban space and the way it was experienced by its original inhabitants. Still, as in any simulation process, some degree of uncertainty must be taken into consideration. In this sense, recent studies have tested procedural 3D modelling as a valuable option to perform an accurate and critical reconstruction of ancient urban areas, by introducing procedural rules that contribute to add transparency to the overall workflow and to map uncertainty in a more solid way (Piccoli, 2016). Additionally, an important source of information lies in the so-called 'legacy data' (Allison, 2008), consisting of archival sources that can often provide significant clues about transformations and changes occurred in the landscape and can therefore contribute to the development of a diachronic representation of the city evolution drawing upon important sources such as historical photographs, unpublished excavation drawings or field reports. By integrating such data in GIS it will be possible to build up an interpretative framework that can allow us to gain new knowledge about the way space was structured and used in antiquity. As recent scholarship has proven, by combining vertical and horizontal map representations together with 3D models it is possible to virtually reconstruct the original context in a way that would not be possible through traditional approaches of documentation . In an urban context it would be interesting also to test volumetric analysis for engaging at a deeper level with the study of visual connectivity, following theories on space syntax analysis (Hillier & Hanson, 1989), to cope with that empty space often disregarded from archaeological enquiry (Campana, 2017) and possibly to add new meaning to the use of the landscape, trying to explore concealment and hiddenness as additional properties (Gillings, 2015).
As this contribution has sought to demonstrate, there are new openings for a more comprehensive study of the urban space in antiquity thanks to advances occurred in digital technology and the use of high-performance computing. Using 3D technology as a heuristic tool for data analysis is still at a very basic stage, and its contribution can be crucial to allow archaeologists to define the original topography of an ancient urban space. More importantly, the analysis of an ancient space as a lived, dynamic environment in which people moved and interacted remains an objective for future research. Data modelling in archaeology has already proven the importance of providing data representations that 'simplify' complex phenomena in order to make them more easily understandable (Lock, 2003): it is about time now to make any simulation process more transparent so that it can afford the possibility to build up different scenarios of reconstruction to test several hypotheses on the use of space. It is therefore essential to generate representations that tend to provide a more accurate picture of the reality, so that they can be effectively used as an integral part of the interpretative process. The massive diffusion of 3D technology within archaeological practice must be encompassed within a theoretical framework where new questions are raised and addressed through the use of dynamic solutions where three-dimensionality