A Case Study on the Reuse of Digital Archaeological Data: An Analysis of the Atlas of Hillforts of Britain and Ireland

Introduction

The data gathered for the Atlas of Hillforts of Britain and Ireland (AHBI) (Lock & Ralston 2022a; 2022b) contains 4147 records, with each record having 244 fields of associated information. This paper summarises an analysis and restructuring of the AHBI data using the programming language Python (Python Software Foundation 2023). This analysis (Appendix 1) provides an open data research tool that presents the online data to facilitate interpretation and reuse. It also restructures the data to make it more easily used in any computational data-driven analyses, including machine learning applications and predictive modelling.

The AHBI entries were compiled by a team of five. Ian Brown did the majority of England and Wales with contributions from Paula Levick, Stratford Haliday compiled Scotland, while James O’Driscoll and Alan Hawkes did Ireland. Professors Ian Ralston and Gary Lock provided an editorial role with William O’Brien providing a similar role in Ireland (S Halliday 2023, personal communication). The AHBI includes core administrative information, location, landscape context, boundary characteristics, dating evidence, previous investigations, and fields such as the character of the interior and entrances. To be included as a ‘confirmed’ fort a site had to meet two of three criteria, namely a locally dominant topographic position, the presence of outer works, and a minimum enclosing area of 0.2 hectares. An ‘unconfirmed’ fort could be included if it met at least one of these criteria and was close to one or other of the criteria (Lock 2019: 6; Lock & Ralston 2022b: 30).

The analysis and restructuring of the Hillforts data presented here comprises seven, live code, Jupyter Notebooks (Appendix 1; Jupyter Team 2023; Middleton 2024), referred to hereafter as the ‘Hillforts Primer’. Each notebook contains Python code within code cells and explanatory text in markdown cells. Together the Notebooks contain 500+ figures which the user can choose to display over a topographic or outline map and output 15 reprocessed data packages. For those interested only in the outputs, the combined Hillforts Primer Notebooks are available as a single PDF (Appendix 1) and the reprocessed data packages in a single Excel file (Appendix 2). All data is made available under a Creative Commons Attribution-ShareAlike 4.0 International licence — CC BY-SA 4.0.

The aim of the Hillforts Primer is to analyse all the online AHBI data (Lock & Ralston 2022a) downloaded from the Atlas of Hillforts Rest Service (Lock & Ralston 2022c), supplementing the published atlas which primarily discusses sites with a ‘confirmed interpretation’ (Lock & Ralston 2022b: 32). If the user wishes to analyse the same filter of hillforts, with a ‘confirmed interpretation’, a toggle is provided in the Hillforts Primer, under User Settings, to enable this. Four additional data sets (dating, entrances, investigations, and references) available from the Rest Service and the AHBI ADS archive are not included in this analysis (Lock & Ralston 2022c; 2024). Every field of the dataset has been analysed to explore its structure, completeness, bias and flaws, and how confident the user can be when using it.

Data engineering, that is analysing and restructuring data for use by automated processes, is a common practice in many disciplines such as medicine and finance but has not been commonly deployed in archaeology. Batist and Roe (2024) identified 493 archaeological data repositories of which 65 contained scripts capable of reuse. None of the repositories specifically mention data engineering. Data engineering or data cleaning (Géron 2019: 62–72) is an important step in the reuse of data. It requires each field of the data to be reviewed individually, to understand what this field is, what the data is telling us and how the data is structured. This process provides insights into the amount and spread of the data, how much data is missing and if there are factors affecting the data that might impact future use; and these factors may develop into drivers for future research and data collection. This data engineering approach restructures the data into a format that is ready for computational analyses, including machine learning. Where there are gaps in data, mixed data types, duplication or empty fields, the data has been cleaned and reprocessed. Duplication is removed, empty fields (e.g. with null values) are filled and mixed data types are split into multiple fields. Where the data can be plotted, distributions, spreads and clustering have been examined visually to draw out insights. In some contexts, the data is transformed to generate distributions that highlight patterning. Similarly, the categorical data (data that can be divided into groups such as word lists) is restructured so that it can easily be converted into numeric data. A key element of this phase is to provide analysis that can be used to validate the observations made by the AHBI.

The context for data reuse

Data reuse by individuals other than those who created it should be at the heart of the scientific method (Huggett 2018: 96). Borgman (2012: 17–24) identifies four key reasons for reusing archaeological data, namely, to validate research (see also The Turing Way Community 2021), to make publicly funded research accessible, to make new insights, and to drive research and foster innovative analysis (see also Huvila 2016: 12). Archiving data is underpinned by an assumption that others will use it and for data preservation to remain justifiable there is a need to demonstrate data reuse. This also upholds the principles of reviewing research outcomes (Huggett 2018: 93) and is vital for maintaining data integrity and identifying errors (Archaeology Data Service 2023). On the challenges of data reuse, Wilkinson et al. (2016) assert the urgent need for improved infrastructure supporting scholarly data reuse and published the FAIR Data Principles that data should be findable, accessible, interoperable, and reusable. Nevertheless, Huvila (2016: 12) notes the relative rarity of data reuse in archaeology and presents data reuse as the poor relative of new data creation. Garstki (2022: 177) highlights the limited opportunities for students to learn methods of data reuse, also identifying the broader issue of poor data literacy in archaeology. Marwick (2017: 440) explores the benefits of digital literacy, acknowledging the initial investment in learning these skills is significant but pointing out that the advantages in facilitating research and making analysis more open and professional outweigh this initial investment. Schmidt & Marwick (2020: 24–25) expand on this further, providing evidence that research published using open-source programming languages and open data lead to increased citations.

Results

The Primer notebooks aim to explore two main questions. Firstly, what pattern of spatial distribution can be found for each field in the AHBI, and how do these patterns vary across different regions? And, secondly, how does data quality and bias in the AHBI affect the patterns of spatial distribution for each field and what data can be confidently used in classification and automated learning?

Another key aim of the Hillforts Primer is to restructure the data so that it can be easily used in computational analyses, of which machine learning is used as an example. A computer will use all the fields of data provided in the same computation, but for a human being there are too many fields of data to visualise all at once. It is impossible for humans to understand machine learning outputs unless we can explore the clusters and bias ourselves. For this reason, the Hillforts Primer reviews each data field individually and plots it using the location data. The published Atlas is a synthesis of the data on the website, whereas the Hillforts Primer presents all the data often complementing figures in the Hillforts Atlas, but also providing a visualisation for every field and, where applicable, adds a density value to aid in highlighting clusters. This allows users to review each field to identify what data will be useful (e.g. for machine learning) but also what data contains a bias that may skew results.

It is not possible to present all the results in this paper. Instead, the paper examines three key themes that have been important in hillfort studies, location, area and classification, and explores how bias in the data can impact interpretation. These examples demonstrate how the Hillforts Primer can be used to confirm, challenge and push forward previous research. The complete analysis and results can be found in the project archive, deposited at the Archaeology Data Service, and made available via appendices 1 and 2.

Distribution of Hillforts

Distribution and density are used throughout the Hillforts Primer. Two examples explored here are the basic spatial distribution of sites and the internal area of forts. These have been chosen as location and size have long been identified as key to classifying hillforts and identifying regional variability (Rivet 1961; Forde-Johnson 1976; Hogg 1979; Harding 2012; Lock & Ralston 2019 & 2022b).

Location

A published analysis of spatial distribution using percolation analysis identifies several regional and local clusters (Maddison 2019: Chapter 8, Figures 8.4–8.6; 2022: Chapter 8, Figure 8.1). The Hillforts Atlas also includes visualisations of density based on historic counties in chapter 9 (Lock & Ralston 2022b: 399–408, Figures 9.1 to 9.4 & Appendix 3; cf. Forde-Johnston 1976). In contrast, the Hillforts Primer uses kernel density (the density in relation to the entire dataset) (SciPy Community 2024) and variance (the distribution and spread of the data) (Bruce et al. 2020: 15) as a guide to identifying groupings within the data. A basic density plot of location data shows a strong cluster in the Northeast and a secondary, less intense, cluster over S Wales and into Wessex (Figure 1a). Because the northeastern cluster is very intense it may mask weaker clusters. Using a boxplot, which shows the distribution of 50% of the data, the box is stretched between these two clusters, and it is not focussed on either (Appendix 1: Part 1 Density Map showing Extent of Boxplot). A boxplot that is not focussed may indicate that there is more than one cluster or that there is data bias. Using a histogram of the north-south spread of the data, the data was split at its lowest point, between the two clusters, to create separate northern and southern data packages. This process was repeated in the northern data package where it is possible to identify a northwestern cluster, focussed over Dunadd. The Irish Sea was used to split off the Irish data and a boxplot used to identify the northwestern and southwestern clusters in Ireland. In total, five regional clusters are identified (Figure 1b) in which density is displayed relative to each data package, enabling the smaller clusters to be more easily identified. This draws out subtle differences in the distributions.

Figure 1

This analysis and means of presentation confirm known aspects of the distribution, namely those areas which lack sites, such as higher ground in northern England and Scotland and some low-lying areas on the eastern seaboard of Britain (Lock & Ralston 2022b: 8). However, in contrast to earlier studies of regional variability, which tended to focus primarily on hillfort area (Rivet 1961; Forde-Johnson 1976; Hogg 1979; Harding 2012), or the Hillforts Atlas, which uses national boundaries or historic counties to create regional statistics (Lock & Ralston 2022b: Chapter 9 & Appendix 3), the Hillforts Primer uses an approach more in line with that of Simon Maddison (Maddison 2019; 2022), by using the data itself to define five regional groupings, which are then used throughout the study to examine variability between these regional groups.

Internal Area

Implicit in the Primer approach is the value of combining multiple characteristics to explore the dimensionality of the data. For example, the five clusters identified in the spatial distribution can be refined further in combination with the enclosed fort area, one of the enduring ways to categorise hillforts data. Rivet’s classes defined for the Ordnance Survey Map of Southern Britain in the Iron Age have had the greatest impact (OS 1962; Rivet 1961: 32–33), comprising three ranges: <3 acres (<1.21 ha); 3 to 15 acres (1.21–6.07 ha); and >15 acres (>6.08 ha). These classes have been most recently used by Harding in his Iron Age Hillforts in Britain and beyond (2012: Figures 1.1–1.3). In 1975 Hogg refined and extended the OS classification to create five groups: <0.24 ha; 0.24–1.2 ha; 1.2–6 ha; 6–30 ha; and >30 ha (Hogg 1975: Figure 3). The Hillforts Atlas reviews several ways in which to split the data statistically based on internal area, ultimately presenting distributions based on a split of the data into five quintiles or groups: <0.2 ha; 0.2 to 0.49 ha; 0.5 to 0.99 ha; 1 to 4.99 ha; and ≥5 ha and over. For forts over 5 ha, further subgroups are created at ≥10 ha, ≥20 ha and ≥30 ha. These are qualified by a recognition that these groupings are selective and that there are many ways in which the data can be cut (Lock & Ralston 2022b: 118).

Statistical analysis of the distribution of the area values in the Primer, where the data is ordered from smallest to largest and split proportionally (as opposed to spatial distribution), split the data into four quartiles, each containing 25% of the data. The central two quartiles were then grouped together to create the central 50% of the data (the Inter Quartile Range). The result is three main clusters of forts according to size: tiny forts (≤0.13 ha); small to medium size forts (0.13–1 ha); and large hillforts (>1–10 ha) (Figure 2). These three groups account for 95.6% of the data. Any fort outside these groups is an outlier and is classified as a very large fort (>10 ha).

Figure 2

Ralston describes how Rivet’s size categories were defined through trial-and-error (Lock & Ralston 2019: 16; 2022b: 12; Rivet 1961: 32–33), while the AHBI categories are acknowledged to be subjective (Lock & Ralston 2022b: 118). In contrast, the Primer splits the data into groups based on the data itself. These splits may not be definitive, as they are influenced by what data was included in the Atlas (selection bias will be examined in more detail below), but they classify the data into four groups based exclusively on the data itself.

While most of the area groupings correspond to the regional ones identified in Figure 1b, the southern region contains distinct clusters of medium sized forts in Wales and large hillforts across south central England (Figure 2). This highlights that classification based on a single criterion, specifically area, is too simplistic and that a range of fields, such as used by Forde-Johnston (1976: 259–61) may be better at identifying smaller sub-groups. By supporting the analysis of every field of data, the Primer provides the foundation for this approach, though before attempting any form of classification data quality and bias need to be assessed.

Data quality and bias

The Primer visualises every field of data in the Atlas so that users can review each map, or plot of the data, and assess what bias, data quality issues or revelations each field might contain. Where these are obvious a comment has been added, but it is assumed that readers with deeper understanding of the data will see more.

Bias of many kinds is endemic in archaeological data but is often ignored in analyses of distributions (cf. Halliday 2011: 19–23). Sources of bias may include topography, land use, survey strategies, individuals, monument recognition, classifications, and terminology (Cowley 2016a: 147–167). Bias does not mean that the data is necessarily wrong or of poor quality, but identifies issues that should be explored and understood to inform subsequent interpretations. An example might be a regional recording bias, where data has been collected intensively in one area and not in another. But recognising bias is important as it helps identify potential research opportunities, such as identifying where harmonisation of terminology is needed.

The term ‘bias’ is used once in the AHBI in relation to the underlying data, to explain that much of the commentary in the Atlas is focused on southern England and Wessex because of the relative frequency of excavation in this region (Lock & Ralston 2022b: 414). Here the Atlas focusses on what the data can tell us about the various hillfort characteristics, which is not to say it shies away from discussing bias. Rather distributions are defined as ‘skewed’ because of the south-central England excavation focus (Ibid: 190), with 40% of all forts in England excavated to some extent (Ibid: 43, 45), accounting for 53.2% of excavations on fort interiors. Since most dating derives from excavation, it highlights the excavation bias in the data when it states that most sites dated between 800BC–AD50 are found in England (Ibid: 324).

When considering classification bias, the Atlas identifies how hillforts have no unique structural features that are exclusive to them and how this is a problem (Ibid: 245); and how combined with the complex nature of a site’s topographic position, this can introduce issues of objectivity when assigning a fort to any specific type (Ibid: 96).

The distributions of tiny forts provide a good illustration of selection bias. The AHBI identifies how the inclusion of small promontory forts in Ireland has deflated the area figures for Ireland (Ibid: 118) – presumably something that must also be true of Scotland and Wales. However, the selection criteria for the Atlas excluded other small sites in Ireland, and similar sites are rare across England, meaning that the spatial distribution reflects selection bias rather than a meaningful distribution of similarly sized sites. The Atlas identifies that the density of forts in England is low, as many smaller sites in England are not traditionally classed as hillforts (Ibid: 105) in contrast to Scotland where small hillforts are identified. Similarly, in the Republic of Ireland, the use of a one-hectare minimum threshold when classifying hillforts excluded many sites (Lock & Ralston 2022b: 144). Halliday (2019: 68–9) discusses the problem of tiny forts that became apparent during the triage of sites for inclusion into the Atlas in Scotland. This resulted in the decision to include sites that met two out of the three Atlas selection criteria, enabling tiny sites, that have traditionally been called forts in Scotland, to be included. However, the same criteria were not applied in Ireland, where an estimated 60,000 examples of early medieval small circular enclosures (O’Driscoll et al. 2019: 83) were excluded, something Ralston (2019: 10) concedes was just too large a corpus to include (Lock & Ralston 2022b: 102). Thus, there is without question, a profound selection bias in the Atlas data.

In relation to survey bias, the impact of aerial reconnaissance provides a good example of where looking beyond self-evident bias in a distribution can provide insights about whether the clusters are skewed by survey methods or just amplified (Figure 3a) (Lock & Ralston 2022b: 7; Halliday 2019: 69, 73). Intensive aerial reconnaissance in Scotland has identified a concentration of cropmark forts in the low-lying agricultural parts of Southeast Scotland, in an area where land use and soils are conducive to archaeological cropmarking (Cowley, 2016b: 65–66; Lock & Ralston, 2019, Figure 2.2). While aerial reconnaissance is known to suffer from a wide potential range of biasing factors (cf. Cowley 2002: 262–264; Cowley & Dickson 2007; Lock & Ralston 2022b: 56), Halliday (2019: 69, Figure 4.10) points out that the cluster in the cropmark sites in Southeast Scotland does not extend north into areas with a similar high density of cropmarks. Thus, while the outputs of aerial reconnaissance may be biased, in this instance the differential densities of forts recorded as cropmarking in lowland eastern Scotland demonstrates different densities in past distributions of forts, suggesting that the concentration is meaningful.

Figure 3

Conversely, the long history of excavation and survey of hillforts in southern England has generated a large amount of dating evidence (Lock & Ralston 2022b: 43, 45 & 414). This is evident in the Atlas in discussion of the ‘zenith’ of hillfort construction in southern Britain during the middle Iron Age while offering no similar statement for the north (Ibid: 314). The clustering of dated sites (Figure 3b) emphasises that the broad dating of hillfort activity is heavily reliant on a specific region that may only be partially relevant to other areas (e.g. the proliferation of Early Medieval sites in Scotland (Noble 2016: 26–35; 2023: 433–443)).

To reduce the problem of regional variation being masked by bias, the data can be reviewed by the regional data packages identified in Figure 1b. This significantly reduces the impact of Atlas wide bias and enables regional variations in the data to be visualised. Figure 3c shows how reviewing the proportion of dated hillforts by region highlights significant regional variations, such as a peak in Irish forts dated to the late Bronze Age and the continued importance of forts in Northwest Scotland and Ireland in the Early Medieval period. It is important to still question these results as, although interesting, away from the southern data package there are very few dates informing Figure 3c, and within the southern data package there is a bias toward Wessex which may mask differences between the large forts of Wessex and the smaller forts of Wales and the Southwest (Figure 2 top).

Classification

The use of analogy in creating groups of archaeological evidence and classifications is a routine but very often implicit process (Nyrup 2020). This is discussed in the Atlas, where hillforts are described as falling on a ‘continuum’ of enclosed sites that do not have clearly defined breaks between different forms (Lock & Ralston 2022b: 101). The three criteria for inclusion (topographic position, scale of enclosing works and size of enclosing area (Lock 2019: 6; Lock and Ralston 2022b: 30–31)) define a ‘confirmed’ hillfort as one that meets two of the three criteria. However, the Atlas data does not contain yes/no fields to identify where a site has met these criteria, so it is not clear on what criteria ‘unconfirmed’ and ‘unreconciled’ hillforts have been included. In the section on bias above, we have already seen how the problem of resource led to the exclusion of 60,000 examples of early medieval, small circular enclosures in Ireland. Selection criteria and different perspectives on what constitutes a hillfort are evident in Wales, where Louise Barker and Toby Driver identified 106 promontory forts (Baker and Driver 2011: 66; Driver 2023, 219) against the 73 listed in the AHBI on tighter criteria.

Christison was one of the first to use classification to show regional differences. He used five classes based on the length of forts to identify a difference between the forts in Argyll and those in Southeast Scotland, observing that 50% of the forts in Argyll were very small, but were absent from some counties in the south east or present in very few in the others (Christison 1898: 385). Forde-Johnston also defined subgroups, using 11 classes based on form, size, vallation, scale of defences, entrance form and occasionally topographic position and regional tradition (Forde-Johnston 1976: 259–61; Lock & Ralston 2022b: 18).

Nevertheless, for all this work on classification there continues to be disagreement about what is and is not a hillfort. Cunliffe (2005: Fig 4.3), for instance, limits his ‘Hillfort Dominated Zone’ to Southeast Scotland, southern England, and the Welsh marches. Hogg (1975: 37) pointed out that classification by a single field is problematic as it ignores the likelihood that grouped forts were likely to have been built over centuries. Brown (2021: 7) suggests that because hillforts are ultimately a product of their topographic setting, any attempt to classify them will produce ‘spurious’ results. Driver (2023: 160) argues that each hillfort is unique, and their morphology should be based on how the builders wanted the fort to be seen by those approaching it along the ‘correct path’. Harding (2012: 14) notes that classification encourages a perception of the hillfort being pre-planned and designed, built to a template, in a way that was likely never imagined by the hillfort builders. The Atlas has not settled this debate, but rather has presented its own selection of sites, area classes (Lock & Ralston: 2022b, 112) and type classifications (Lock & Ralston 2022b: 79, Fig. 3.5).

For all this ambiguity, classification continues to be an important tool (e.g. Adams & Adams 1991). In ‘Understanding the Archaeology of Landscapes’ (Historic England 2017) the importance of analogy in classifying monument types and placing them into periods is discussed. Within archaeological archives such classes remain a key tool for the storage and retrieval of information and these types and classifications have, over time, been developed into wordlists and thesauri (FISH, 2023) with an example from Scotland being the Monument’s Thesaurus definition of “fort”:

“FORT (Hillfort (Non-Preferred)), An enclosure, often located on a hilltop, bounded by one or more banks, ditches, ramparts or walls. Use for prehistoric and early historic sites.” (HES 2023; FISH 2024)

The Hillforts Primer analysis of the AHBI builds on an empirical approach, combining it with a statistical analysis of variance and clustering to support the creation of robust, evidence-based regional classifications. By reviewing the Atlas fields, the typical fort within an area can be identified. The example in Figure 4 shows a classification for the northwestern region, which is isolated in Figure 1b. The Primer is used to filter the data within the extract to identify the dominant fields within the selected area. Fields are selected if they apply to 50% of the forts in that region or are the dominant type within their class. Having regional classifications for the typical fort makes it possible to focus on the origins of that regional class, also enabling examination of those that fall outside the norm to see if further subdivision is possible or if there are reasons these forts diverge from the norm. The Hillforts Atlas does not attempt to offer classification using a suite of fields in this way.

Figure 4

Conclusions

The Atlas of Hillforts of Britain and Ireland is a valuable resource that has created a large, open dataset available for reuse and reinterpretation. The Hillforts Primer provides a companion to the Atlas, analysing, mapping and plotting all 244 fields of data. It provides context and insight into the spread, distribution, density, and bias of the data that can be used to explore issues such as regional groupings and to create classifications for the typical fort within extracts of the data. It moves classification from a subjective process to a statistical repeatable one based on the data. As a comprehensive reusable analysis, it can be used by researchers for verification and as the foundation for further research as well as offering the potential for the code to be reused in similar systematic analysis of other monument types.

Analysing, restructuring and cleaning the data has benefits for future human and computer-aided research. It promotes reuse by improving the machine readability of the data and by providing a full review of all the fields of data. This enables the results of automated processes to be assessed for the impact of bias in their outputs. By providing the code used in the analysis, any future research will benefit by being able to use and repurpose this code to ask more questions of the Atlas data. It also means that as the corpus of hillforts develops and grows this analysis can be run again to review the impacts of new information.

One of the big questions in hillfort studies remains what is and what is not a hillfort. The discussion above notes that the criteria for inclusion in the Atlas have not been the same across the nations. One area where automated learning might contribute could be in using the current data as a training set that can be used to identify further sites that match the criteria of the Atlas.

There are many archaeological observations in the Hillforts Primer and not all can be listed here. In terms of the examples discussed, creating data driven classifications for the average fort will be one way to further explore the clusters identified using Euclidean distance (Maddison 2019: Chapter 8; 2022: Chapter 8). Similarly, the use of kernel density and the division of the data into five regional clusters is novel in hillfort studies. Although the clusters in Ireland may be compromised by selection bias, and the Tweed Basin, Wessex and Wales clusters are well known, the author is not aware of any publications highlighting a cluster over the western seaboard of Scotland, focussed on Dunadd. Dunadd is thought to have been a place of inauguration in the Early Medieval period but the evidence for the Iron Age occupation is heavily truncated (Lane & Campbell 2000: xiii & 44–58). Sitting at the centre of this western seaboard cluster may indicate that the Kilmartin Glen, and possibly Dunadd, were equally important in the Iron Age.

Data Accessibility Statement

Please see appendices 1 and 2.

Ethics and Consent

The data reused in this study was made available under a Creative Commons Attribution-ShareAlike 4.0 International licence — CC BY-SA 4.0. All outputs of this study (see appendices 1 and 2) are made available under the same licence. Other than the data made available under CC BY-SA 4.0, no other 3rd party data, or restricted content, has been used in the writing of this paper.

Funding Information

This research was carried out independently in the author’s own time. The archiving and publication of this work has been supported by funding from Historic Environment Scotland (HES).

Competing Interests

The author has no competing interests to declare.

Author Contributions

This paper is the work of Mike Middleton.