Original Research

published: 11 August 2017

doi: 10.3389/fdigh.2017.00017

Visibility Network Patterns and

Methods for Studying Visual
Relational Phenomena in Archeology
Tom Brughmans* and Ulrik Brandes

Department of Computer and Information Science, University of Konstanz, Konstanz, Germany

A review of the archeological and non-archeological use of visibility networks reveals

the use of a limited range of formal techniques, in particular for representing visibility
theories. This paper aims to contribute to the study of complex visual relational phe-
nomena in landscape archeology by proposing a range of visibility network patterns and
methods. We propose first- and second-order visibility graph representations of total
and cumulative viewsheds, and two-mode representations of cumulative viewsheds. We
present network patterns that can be used to represent aspects of visibility theories and
that can be used in statistical simulation models to compare theorized networks with
Edited by:
Colin D. Wren, observed networks. We argue for the need to incorporate observed visibility network
University of Colorado Colorado density in these simulation models, by illustrating strong differences in visibility network
Springs, United States
density in three example landscapes. The approach is illustrated through a brief case
Reviewed by:
study of visibility networks of long barrows in Cranborne Chase.
Irmela Herzog,
Landschaftsverband Rheinland,
Keywords: visibility networks, signaling networks, landscape archeology, viewsheds, Bernoulli random graphs,
Germany exponential random graph model
César Parcero-Oubiña,
Consejo Superior de Investigaciones
Científicas (CSIC), Spain
Jessica Munson, INTRODUCTION
Lycoming College, United States

*Correspondence: A large diversity of visual relational phenomena is studied in landscape archeology using a wide
Tom Brughmans range of methodological tools (Llobera, 2003). This diversity derives from the complexity of
[email protected] human behavior afforded by landscapes (Gillings, 2009). Visibility networks are one of many
methodological tools that can help landscape archeologists to limit the complexity of the range
Specialty section: of affordances of the landscape, when studying particular ways in which landscapes affect and
This article was submitted structure human behavior. This paper aims to contribute to the study of complex visual relational
to Digital Archaeology, phenomena in landscape archeology by proposing new visibility network patterns and methods.
a section of the journal The representation and study of visual properties as network patterns is largely restricted to stud-
Frontiers in Digital Humanities
ies in cognitive science, architecture, geography, and archeology. Different network science methods
Received: 01 February 2017 are used in these disciplines to study and represent a very diverse range of phenomena. In spatial
Accepted: 20 July 2017
cognition and artificial intelligence, networks are used as models of the mental representation of
Published: 11 August 2017
environments and applied, for example, in autonomous navigating robots (e.g., Mallot et al., 1997).
Citation: Schölkopf and Mallot (1995) proposed the view graph as an approach to study visual navigation and
Brughmans T and Brandes U (2017)
path planning in cognitive science and artificial intelligence. In a view graph, nodes represent views
Visibility Network Patterns and
Methods for Studying Visual
(snapshots) and edges represent movements from one view to another. In architecture, they are used
Relational Phenomena in Archeology. to study the visible area from observation points in a building or urban environment, to study the
Front. Digit. Humanit. 4:17. movement through and use of space (e.g., Turner et al., 2001). For example, an axial map created
doi: 10.3389/fdigh.2017.00017 by drawing a set of axial lines through space on a two-dimensional plan of a built environment

Frontiers in Digital Humanities  |  www.frontiersin.org 1 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

is a prominent tool in space syntax and is commonly used for network creation. These methods will be illustrated using the
the study of pedestrian movement patterns in architectural space locations of late pre-colonial sites on the island of La Désirade
(Hillier and Hanson, 1984; Turner et  al., 2005). In archeology (Guadeloupe, French West Indies) and through a brief case study
visibility networks are used to represent the intervisibility of sites of intervisibility networks of long barrows in Cranborne Chase
or features, most commonly for the study of past communica- (Tilley, 1994).
tion networks through visual signaling (e.g., Swanson, 2003).
Visibility networks are also used to explore the properties of time VISIBILITY NETWORK STUDIES OUTSIDE
series data (Lacasa et al., 2008), to study topics as diverse as hur- ARCHEOLOGY
ricanes (Elsner et al., 2009) and Alzheimer’s disease (Ahmadlou
et al., 2010). Visibility networks are here defined as all uses of network data,
Despite this strong diversity of topics, only a very small consisting of a set of nodes and a set of edges, for the representa-
range of network science techniques has been used to study the tion and/or analysis of visual properties (visual characteristics of
properties of visibility networks and scholars’ theories about entities within physical or abstract space). Different approaches
how lines-of-sight affect human behavior. One area that has to visibility network creation exist in a wide range of disciplines
seen particularly little development is the study of landscape as mentioned above [see Franz et  al. (2005) for a review of
visibility analysis through visibility networks (O’Sullivan and graph-based spatial models in general]. We will discuss the use
Turner, 2001), although this field of study has great potential for of visibility graphs in architecture and landscape studies in more
landscape archeology. detail because we believe they could inspire interesting new
In this paper, we argue that a number of the theories land- applications in archeology, as illustrated in the remainder of this
scape archeologists formulate about how lines-of-sight could paper (see Table 1 for an abstract summary).
have affected past human behavior can be appropriately studied
through visibility networks, and we present network science tech- Visibility Graph (Architecture)
niques for representing aspects of such theories. After a discus- In a visibility graph in the context of architecture, nodes rep-
sion of selected multi-disciplinary visibility network studies, we resent isovists/viewsheds and edges represent intervisibility of
present new visibility network patterns, we explore the variability isovists/viewsheds (Turner et al., 2001). A viewshed or “isovist
of visibility network density through a range of examples, as well is the set of all points visible from a given vantage point in space
as propose methods that can be used to formally represent and and with respect to an environment” (Benedikt, 1979, p. 47),
study aspects of theories proposed by landscape archeologists. and it can therefore be considered a visual property of the
A range of network science methods will be proposed to study vantage point. The intervisibility of viewsheds can be defined
the properties of these networks, to identify the frequency of in two ways: a first-order visibility relationship exists when
network patterns in observed networks, and to compare observed the viewsheds of a pair of observation locations intersect and
networks with models representing different theories of visibility if these observation locations are themselves intervisible and a

TABLE 1 | Summary of multi-disciplinary overview of network data representation for visual relational phenomena.

Graph model Nodes Edges Directionality Example Network data representation

First-order Observation Intervisibility of Directed or undirected

visibility graph location locations
(architecture)

Second-order Observation Overlap of Directed or undirected

visibility graph location locations’
(architecture) viewsheds

Visibility graph Observation Intervisibility of Directed or undirected

(landscape) location locations

Frontiers in Digital Humanities  |  www.frontiersin.org 2 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

second-order visibility relationship exists when the viewsheds VISIBILITY NETWORK STUDIES
of a pair of observation locations intersect but the locations IN ARCHEOLOGY
are themselves not intervisible (Turner et al., 2001). Although
it is argued such networks can subsequently be studied with a Although formal visibility studies in archeology are well estab-
wide range of graph theoretical measures (Turner et al., 2001, lished (Lake and Woodman, 2003; Llobera, 2003), visibility
p. 109), the published use of such measures is largely restricted network studies in archeology are rare and employ a similarly
to node clustering coefficient, node neighborhood size, and the limited range of network science techniques, often restricting
mean shortest path length (Batty, 2001; Turner et al., 2001). The the study to a network diagram representation that is explored
neighborhood size of a node is the set of nodes it is directly con- visually. An early example is David Fraser’s (Fraser, 1980, 1983,
nected to (its degree in network science terms) and in visibility pp. 379–387) study of past power relationships of dominance
graphs it represents the number of locations directly visible and subservience, which Fraser argues might be expressed in
from an observation location. A node’s clustering coefficient the dominant and intervisible positions of chambered cairns in
is the number of edges between nodes in its neighborhood Orkney (United Kingdom):
divided by the maximum number of edges in a neighborhood
of this size and is interpreted as indicating “how much of an If one cairn can be seen from many other cairns then it
observer’s visual field will be retained or lost as he or she moves is always present and can never be ignored by the users
away from that point” (Turner et al., 2001, p. 110; this interpreta- of those other cairns. If messages were passed between
tion is contested by Llobera, 2003, pp. 27–28). A node’s mean cairns by visual means (such as flags or beacons) then
shortest path length is defined as the mean of all shortest paths a highly intervisible cairn becomes an important link
from this node to all other nodes (what Turner and colleagues in the dissemination of information (Fraser, 1983, 380).
refer to here is known in network science literature as a node’s
closeness centrality, whereas the mean shortest path length is Intervisibility of cairns was determined through observation
more commonly used to refer to the mean of all shortest paths in the landscape, reciprocity (i.e., mutual visibility) between
between all nodes as a measure for the network as a whole). cairns was assumed, and a visibility network was created where
Batty (2001) further explores the distance attribute of lines-of- nodes represent chambered cairns and undirected edges repre-
sight in a visibility graph, by measuring the average, minimum sent intervisibility. Three network measures were applied: the
and maximum distance of a location’s viewshed, as well as the connectivity index refers to the proportion of observed edges over
compactness (average divided by maximum distance). the number of potential edges (i.e., the network density), nodality
refers to the degree of a node and degree distribution is explored,
Visibility Graph (Landscape) and a node’s cutpoint index refers to the number of components
Crucial in the context of this paper is that first-order visibility created by the removal of this node (Fraser, 1980). Although
graphs have been proposed for landscape visibility analysis Fraser concluded the approach did not offer strong evidence for
(De Floriani et al., 1994; Nagy, 1994; Puppo and Marzano, 1997; the existence of power relationships expressed through cairn
O’Sullivan and Turner, 2001). An overview of the applications of location, it did suggest that certain groups of cairns may have
landscape visibility analysis that can be represented by visibility been purposefully located in visually dominant places.
graphs is offered by Nagy (1994), who distinguishes between: Another example of the creation of visibility networks through
locating observation points and hiding places, determining line- observation and visual assessment is provided by Christopher
of-sight communication networks, scenic and hidden surface Tilley (1994) (pp. 156–166) in his book “A Phenomenology of
paths, identification of landforms like peaks and pits, and the Landscape”. An undirected visibility network is explored visually
use of visibility for navigation. De Floriani et  al. (1994) used and using node degree, and a process is suggested to explain the
visibility graphs to represent lines-of-sight between locations patterns identified. This work will be discussed in more detail and
in a landscape, to study the line-of-sight network problem of expanded on in a brief case study below.
finding the minimum number of nodes, which must be placed Ruestes Bitrià (2008) in her work on intervisibility of Iron Age
on a landscape to ensure communication through intervisibility hillforts in Catalunya derives a visibility network by representing
between a set of nodes. O’Sullivan and Turner (2001) used hillforts as nodes and their intervisibility as edges. The resulting
the visibility graph as a network data representation of a total network was explored visually, and it was concluded that it could
viewshed in a landscape (i.e., lines-of-sight from every location have supported a communication network. Through his study of
to every location in a landscape; Llobera, 2003), and measured hilltop features around the site of Paquimé (Mexico), Swanson
the node degree, node clustering coefficient, and the network’s (2003) also aimed to explore whether these locations could
mean shortest path length. The authors suggested visibility have functioned as a fire-signaling communication network.
graphs could be studied using a wider range of network science An undirected visibility network was created where nodes repre-
approaches, like centrality, cohesive subgroups, and structural sent sites and edges their intervisibility. Swanson also compared
equivalence. Such studies are still to appear to our knowledge, the visibility network he observed with simulated visibility net-
and the current paper significantly expands the network science works by sampling random points in the landscape, an approach
arsenal for studying visibility networks with a different range of also taken by Earley-Spadoni (2015) in her study of Early Iron
techniques closely tied to the theoretical motivations why one age and Urartian fire beacon signaling network. A further study
creates a visibility network. of visual signaling networks is provided in Shemming and

Frontiers in Digital Humanities  |  www.frontiersin.org 3 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

Briggs’ (Shemming and Briggs, 2014) analysis of Anglo-Saxon phenomena. However, there is little cross-fertilization between
place names and communication networks in Southern England. disciplines in the use of visibility network data representation.
Possible beacon locations were conjectured from Anglo-Saxon A notable exception to this is the use of first-order visibility
place names, and compared with patterns of intervisibility to graphs in both architecture and archeology. We believe that
explore whether they could indeed serve as a communication exploring more diverse ways of using network data representa-
network, therefore adding credibility to their identification as tions might well lead to useful new ways of studying visibility
beacon locations. networks in these and other disciplines. In this section, we will
De Montis and Caschili (2012) perform a formal vis- develop this argument for landscape archeology in particular,
ibility network analysis in their study of intervisibility among by offering examples of alternative network data representa-
Nuraghes (prehistoric towers) in Sardinia (Italy). The authors tions of visual properties. To illustrate these examples, we
aim to “investigate the hypothesis that the spatial patterns of will use the locations of the late pre-colonial sites of Morne
the Nuraghes obey rules of intervisibility control over the sur- Cybèle, Morne Souffleur, and Anse Petite Rivière on the island
rounding territory” (De Montis and Caschili, 2012, 315), that of La Désirade (Guadeloupe, French West Indies) (we selected
they could have functioned as a communication network, and these three sites for their prominent locations only to best
that they were not isolated entities but worked as a “networked illustrate the concept of our proposed network methods, an
settlement” linked together by lines of intervisibility. The actual visibility study of these sites is published as Brughmans
intervisibility network was derived from binary viewsheds et al., 2017).
using single observer locations on the Nuraghes. Like Ruestes In landscape archeology, visibility networks are most com-
Bitrià and Swanson, the authors also only consider mutual vis- monly created by representing archeological features or sites
ibility of importance and their network is therefore undirected, that are the focus of research interest as nodes, and lines-of-sight
missing the opportunity to explore situations where a line of between them as undirected edges (Figures  1A,B). However,
sight between Nuraghes is not reciprocated, which we believe most sites and features are not single points in space but occupy
should be included if one wishes to evaluate visual control (we areas. This can be incorporated in a visibility network in multiple
conceptualize visual control as the ability to observe many loca- ways:
tions from a vantage point whereas a node’s visual prominence
is the ability to be seen from many locations; see Brughmans (a) Representing each point location within a site area as a node,
et al., 2014). Three local node-based measures are used: degree, and each line-of-sight from each point location as an arc
betweenness centrality and clustering coefficient. The use of (a directed edge) (Figures 1C,D).
betweenness centrality (a measure that calculates paths between (b) Bundling all lines-of-sight of each point location and repre-
nodes which are considered media for the flow of material and senting them as a weighted arc (where the weight attribute of
immaterial resources) is appropriate in light of the authors’ the arc represents the number of lines-of-sight) emanating
assumption to treat this network as a communication network. from a single node per site (Figure 1E).
In her study of settlement patterns in the north-western (c) Representing the area of a site as a node attribute that can be
Dominican Republic, Samantha de Ruiter (2012) aims to evaluate used in further network analysis (Figure 1F).
what role visibility could have played in the selection of site loca- (d) A combination of the above (Figure 1G).
tions. De Ruiter emphasizes the need to experience visibility as
well as calculate it. Through a survey of the study area the author Representing point locations in site areas as shown in
became familiar with the physical landscape and was able to visu- Figures 1C,D offers the advantage of being able to address dif-
ally identify key landscape features, resource areas and other sites ferent questions: from what locations in a site area can locations
from site locations, information that was subsequently compared in another site area be seen? From what locations can other sites
with the results of formal visibility analyses. She calculated the not be seen? What locations have both incoming and outgoing
percentage of the study area that is visible from sites and the lines-of-sight?
percentage of sites visible from other sites, which allowed her to Lines-of-sight are commonly represented in landscape
create a network of intervisible sites that was explored visually. archeology as binary undirected edges without attribute infor-
Finally, exponential random graph modeling (ERGM) has mation (Figure 1B). However, lines-of-sight in this context are
recently been proposed as an approach to represent and explore fundamentally directional phenomena, connecting the eyes of
the dependence assumptions motivating an archeological visibil- an observer with an observed feature. They could therefore be
ity study (Brughmans et al., 2014, 2015). The method compares represented as directed arcs (Figures 1E,G), although in many
the frequency of configurations (small network patterns repre- research contexts it is desirable to assume intervisibility when
senting the dependence assumption) in the observed visibility an equal height is used for the observer and target, or due to
network with their frequency in stochastic models. the limitations of the input data (accuracy of site location,
resolution of topography model; see Conolly and Lake, 2006,
NETWORK DATA REPRESENTATION FOR 230, Figure 10.17). This presents an advantage of network data
VISIBILITY NETWORKS representations over raster representations of cumulative and
total viewsheds in particular, which do not typically distinguish
This overview reveals that network data have been used very dif- between incoming and outgoing lines-of-sight within a single
ferently in a diversity of disciplines to represent a wide range of raster. Instead of binary edges representing either the presence

Frontiers in Digital Humanities  |  www.frontiersin.org 4 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 1 | Alternative representations of a visibility network between a pair of sites: single node per site (A,B); nodes for each location within site area
(C,D); directional arcs representing bundled lines-of-sight (E); site area represented as node attribute (F); combination of (E,F) (G).

or absence of a line-of-sight between two sites, weighted edges subsequently be used to select and study subnetworks (e.g., all
could be used to represent the number of lines-of-sight from lines-of-sight with a probability over 50% and a length shorter
all locations in a site area (Figures  1E,G). Finally, additional than 3 km).
information about the lines-of-sight, like their length or experi- Network data can also be used to represent phenomena other
mentation settings, could be represented as attribute informa- than direct lines-of-sight between observers and sites, as we
tion of the edge or arc. In research contexts where probabilistic have learned from the multi-disciplinary overview in Section
viewsheds are generated (Fisher, 1994), the probability of “Visibility Network Studies Outside Archeology.” We argue
a line-of-sight could be included as attribute information the use of second-order visibility graphs might prove a useful
(e.g., Brughmans et  al., 2015). This attribute information can alternative representation of the results of a cumulative viewshed

Frontiers in Digital Humanities  |  www.frontiersin.org 5 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

analysis (the locations visible from a set of observation loca- The one-mode network of observation locations (Figure  3E)
tions; Wheatley, 1995), where a pair of nodes representing sites offers the advantage of being able to explore clusters of locations
is connected by an undirected edge if part of the surrounding that can be observed from the same or similar sets of sites, by
landscape is visible from both sites (Figure 2). First- and second- using network community detection algorithms (see Fortunato,
order visibility networks allow one to address different research 2010 for an overview of such methods). Moreover, note how
questions. Direct lines-of-sight between sites hypothesized to the color coding does not necessarily reflect geographically
have formed a signaling network can be explored using a first- proximate locations (Figures 3B,D), whereas clusters of nodes
order visibility network representation (Figure 2A), whereas the with the same color are more clearly visible in this one-mode
ability to observe the same parts of the surrounding landscape network representation (Figure 3E).
from sites hypothesized to have been used for visually control- Finally, we believe landscape visibility graphs could serve as
ling the landscape can be explored using second-order visibility a useful alternative representation of total viewsheds (Llobera,
networks (Figure 2B). 2003), where each location in a landscape is represented by
Moreover, the use of second-order graphs and landscape a node and lines-of-sight connect all intervisible locations.
visibility graphs can be combined in a two-mode network Although much archeological visibility analysis is site-centric
representation, where one mode represents the observation by focusing on the visual properties of a set of archeological
locations of the cumulative viewshed as nodes and the other features, recent studies emphasize the need to study how sites
mode represents each location in the landscape as nodes. are embedded within the visual properties of landscapes as a
An example of such a two-mode network is shown in Figure 3A, whole through the use of total viewsheds: the sum of viewsheds
and a geographical representation of the same network in from every single location in a landscape (e.g., Llobera, 2003;
Figure  3B. We purposefully created a visibility network with Eve and Crema, 2014; Gillings, 2015; Brughmans et al., 2017).
a very coarse geographical resolution (landscape points are In such studies, landscape graphs as introduced by O’Sullivan
spaced at 200 m intervals), and we want to highlight that this and Turner (2001) can be used as alternative representations
example and the borders of the study area are entirely artificial. of total viewshed results, in which nodes represent landscape
This visibility network merely serves to clearly illustrate the locations and pairs of nodes are connected by an edge if a
method we propose, whereas in real studies an appropriate line-of-sight connects these locations. Figure 4 illustrates this
geographical resolution will need to be selected and the issue of through a purposely coarse resolution total viewshed, to better
possible edge effects at the borders of the analysis area will need explain the approach. Both the geographical and network data
to be addressed. A one-mode projection of this two-mode net- representations of total viewsheds are useful for highlighting
work on the observation locations results in the second-order certain patterns, each has distinct advantages over the other.
graph of observation locations, where a pair of observation The network data representation does not aim to be as useful
locations is connected by a weighted edge representing the in communicating spatial patterns, for which a geographical
number of locations in the landscape they can both observe representation would be more appropriate. We argue these
(Figures  3C,D). A one-mode projection on the landscape alternative representations can be used alongside each other,
locations represents pairs of landscape locations visible from applying the most appropriate representation for the data
the same observation location (Figure  3E). The two-mode patterns one is interested in. The network data representation
network (Figures 3A,B) has the advantage of allowing one to of a total viewshed has the advantage of being able to identify
explore what locations can be observed from what sets of sites, structural properties of research interest for a landscape as a
information that is lost in a raster representation of a cumu- whole rather than just for a set of site-based observation loca-
lative viewshed with more than a few observation locations. tions. Some network measures can be used as alternatives to

FIGURE 2 | First-order visibility graph where directly intervisible nodes are connected (A). The cumulative viewshed results in Figure 3D can be represented as a
second-order visibility graph where a pair of nodes is connected if other locations are visible from both (B).

Frontiers in Digital Humanities  |  www.frontiersin.org 6 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 3 | A two-mode representation of a cumulative viewshed, where one mode (triangles) represents sites and the other (circles) represents landscape
locations (A,B). From this, two one-mode projections can be derived representing either the sites connected through the mutual visibility of landscape locations, i.e.,
the second-order visibility graph, (C,D) or the landscape locations connected through being visible from a pair of sites (E). Color coding is used to identify locations
visible only from Morne Souffleur (red), only from Morne Cybele (green), only from Anse Petite Riviere (blue), from pairs of sites (cyan, yellow, purple), and from all
three (white).

equivalent measures in a raster data representation of a total of values in the total viewshed raster. The degree distribution
viewshed. For example, the degree of a node is equivalent to in Figure 5 of the network shown in Figure 4 is skewed to the
the number of locations visible from a given raster cell, and the right, suggesting that most nodes have a limited degree whereas
degree distribution is therefore equivalent to the distribution a few have a far higher degree. In this network, the site of Morne

Frontiers in Digital Humanities  |  www.frontiersin.org 7 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 4 | Geographical (A) and network (B) data representation of total viewshed results, where each node in the landscape is represented by a node and pairs
of nodes are connected by an edge if this pair of locations is connected by a line-of-sight. This figure presents a purposely coarse resolution total viewshed
experiment to better illustrate the approach.

FIGURE 5 | Degree distribution of visibility network representing total viewshed results shown in Figure 4.

Cybèle has the highest degree: it is intervisible with 33 other like betweenness centrality, or the node clustering coefficient
locations. However, some network science measures have no as proposed for landscape visibility graphs by O’Sullivan and
equivalent in raster data representations and therefore offer new Turner (2001). Of particular interest among these techniques is
approaches to exploring total viewshed results. Examples include the ability to identify network data configurations representing
techniques based on the calculation of paths through a network theoretical dependence assumptions (see Table 3 for the count

Frontiers in Digital Humanities  |  www.frontiersin.org 8 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

of configurations in the network shown in Figure 4). In the next TABLE 2 | Configurations for theories of visibility networks in landscape
archeology.
section, we will introduce a range of theoretical dependence
assumptions and their associated network data representations Configuration Description Network pattern
that have particular potential for landscape archeology research.
Isolate Tendency for nodes to
be invisible from other
nodes
VISIBILITY THEORIES AND NETWORK
Baseline tendency for
PATTERNS Edge
pairs of nodes to be
intervisible
A particularly striking conclusion drawn from the literature
above is the limited use, especially in landscape archeology, of Triangle Tendency for the
network patterns to represent theories about the functioning, visibility network to be
emergence, or evolution of visibility networks. Such theories clustered

are referred to in network science as dependence assumptions

(Brandes et  al., 2013): theories about how edges are depend-
3-path Tendency for paths of
ent on each other’s existence. Network patterns can be used intervisible nodes. Can
as representations of what one expects to see in light of such be used to represent
a theoretical assumption. Visibility theories are often implied visual signaling
in archeological studies of visibility networks and even com-
Alternating-star Tendency for some
monly stated explicitly. For example, in their study of vis-
nodes to have far
ibility networks among Nuraghes in Sardinia, De Montis and more lines-of-sight
Caschili (2012) hypothesize the existence of a communication than others, i.e., the
network and argue this is reflected through paths in the vis- degree distribution is
ibility network. This represents a visibility theory: Nuraghes are spread or skewed.
Can be used to
positioned in such a way that they are intervisible with other
represent visually
Nuraghes and together create paths, where information can be prominent nodes
spread by visual signaling between a pair of not-intervisible or nodes visually
Nuraghes through a third. Such theories can be represented by controlling other nodes
network patterns, commonly called configurations in network in the landscape

science. For example, De Montis and Caschili (2012) consider

a path through their visibility network to represent the ability
for information to be shared through visible signaling between network. Triangles can be used to represent a tendency for the
the Nuraghes. visibility network to be clustered. Paths can be used to represent
It is crucial to highlight that the approach we suggest is not communication through intermediary nodes, in studies where
concerned with evaluating the probability and reason of the visual signaling networks are hypothesized (e.g., Fraser, 1983;
observed network structure emerging in a particular physical De Montis and Caschili, 2012; Ruestes Bitrià, 2008; Swanson,
landscape. While this would be the ultimate goal, its confound- 2003; Shemming and Briggs, 2014): we expect a non-trivial
ing factors include the underlying social processes, topography, signaling network to contain paths where information can be
and human interference with it. Instead, we focus on one shared between a pair of non-intervisible locations via one or
question only: how likely is it that this observed network is the more intermediary locations. Alternating-stars can be used to
result of hypothesized structural processes? This assumes that represent visually prominent nodes, or nodes visually control-
the landscape offers sufficiently many possibilities for the net- ling surrounding nodes, in studies where the importance of
work structure to form, so that it is not necessarily the case that hub nodes in the visibility network are hypothesized (e.g.,
configurations are under- or over-represented. Thus, the model Fraser, 1983; Tilley, 1994). A network with a high number of
is not restricted by topography and only the hypothesized struc- alternating-stars will have a highly skewed degree distribution,
tural processes are simulated. The abundance of patterns found with very few nodes having a high number of links while most
in simulated networks is then used to assess the significance of nodes have a very limited number of links (for technical details
those in the observed network. see Koskinen and Daraganova, 2013, 65–67). Moreover, where
In Table 2, we list a number of theories drawn from the study more geographical reality is desired and appropriate, the length
of visibility networks in landscape archeology, which could also of lines-of-sight can be incorporated in all these network data
possibly find useful application in other disciplines. We present patterns as an edge attribute to explore correlations between
undirected network data configurations for these theories, changes in length and probabilities of the creation of network
equivalents for directed visibility networks are published by data patterns (for geospatial models including edge length, see
Brughmans et al. (2014). Robins and Daraganova, 2013, 99–101).
Isolates can be used in studies where the tendency for nodes These network data configurations can be used in a diversity
to be invisible is explored (e.g., Gillings, 2015), while edges can of ways to express and explore theories about the structure and
be used to represent a tendency for nodes to be connected in the development of visibility networks, and how they affected past

Frontiers in Digital Humanities  |  www.frontiersin.org 9 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

human behavior. In the first instance, one can identify those they are compared with. In the next section, we will explain why
configurations that best express one’s theories. One would it is more appropriate to compare with a model with network
expect the configurations representing one’s theories to be par- density equal to the observed network.
ticularly frequent in the observed visibility network, which can
be evaluated by counting the frequency of all configurations in
the network being studied. The first column of results in Table 3 DENSITY OF VISIBILITY NETWORKS
shows the count of configurations of the total viewshed network IN PHYSICAL LANDSCAPES
presented in the previous section (Figure  4): this network has
a high number of star configurations, triangles and three-paths, The main reason why randomly generated networks are far
and few isolated nodes. denser than the observed visibility network is simply that, com-
However, a simple count of configurations does not reveal binatorially, there are many more dense networks. An instance
whether they are more frequent than would be expected sampled uniformly at random from the set of all networks is not
by chance in a network of this size. To evaluate this, we can subject to limitations imposed by physical realities. However, in
represent a random process where edges between each pair of real-world physical environments, visibility network density is
nodes are created with equal probability, through a Bernoulli affected by the maximum viewing distance, observer and target
random graph model: each edge will have a probability of being elevation, the spatial density of observation points (points-per-
created at the toss of a coin. The count of configurations of the area), the landscape topography, vegetation and atmospheric
graphs resulting from such a model can be compared with those conditions: in actual landscapes some lines-of-sight between
of the observed network to evaluate whether a random graph pairs of observation points can therefore never be created, while
model where edges emerge independent of each other is a good in randomly generated networks all edges between all pairs of
description of the observed visibility network. The result of a nodes have a probability of being created. In this section, we give
Bernoulli random graph model with 120 nodes are shown in the a few examples of how much visibility network density can vary
second column of Table 3 (the configuration counts are averages in physical landscapes, to illustrate the importance of incorporat-
over 10,000 randomly generated networks), confirming that the ing visibility network density into visibility network models. The
observed network differs strongly from the randomly generated examples are shown in Figures 6–10 and their network density
ones, which are far more dense (3,581.422 edges on average, scores are presented in Table 4.
compared to 460 edges for the observed network). Network The maximum viewing distance concerns both the effects
density is defined as the fraction of the number of edges that of the curvature of the earth and the distance at which features
are present to the maximum possible number of edges in the of interest are visually distinguishable. In a landscape with flat
network. However, comparison with such randomly generated topography, the higher the observation height and the higher
networks is problematic because they will always have about the observed feature, the further this feature will remain visible
50% of all edges present, whereas this is rarely possible for vis- above the horizon. But inevitably all features become invisible
ibility networks in physical landscapes. Comparison of configu- to a human observer as they disappear behind the curvature of
ration counts between networks with different densities as in the earth, which limits the network density of visibility networks
Table 3 is problematic, since a higher count can be expected for spanning huge areas. Moreover, landscape archeologists are often
all configurations in networks with higher network density. We particularly interested in the ability of human observers to see
merely included this example of a random network model with particular features such as archeological sites, smoke columns,
network density 0.5 because we noticed this model is very com- or other humans. This can be affected by the distance at which
monly but inappropriately used in archeological and historical the feature is physically visible to humans in general, the ability
network research due to the misleading appearance of unbiased of a particular observer to distinguish the feature (i.e., cultural
randomness, which is in reality a strong bias toward fixed 0.5 affinity with the feature, regular exposure to the feature, or the
network density. We strongly argue for random network models state of the observer’s eyesight), and how the feature contrasts
to assume the same network density as the observed networks against the background (Wheatley and Gillings, 2000). In such

TABLE 3 | Count of visibility network configurations in the total viewshed network shown in Figure 4 above, and in four different Bernoulli random graph models
(10,000 randomly generated networks with 120 nodes).

Configuration Observed network Random network Fixed density Structural zeroes Fixed density and structural zeroes

Nodes 120 120 120 120 120

Edges 460 3,581.422 460 3,568.37 460
Density 0.064 0.502 0.064 0.499 0.064
Isolates 4 0 0.0422 0 0.0415
Triangle 682 35,448.42 74.525 35,065.17 74.576
3-path 64,883 12,546,873 26,411.43 12,412,036 26,410.529
Alternating-star 1,416.301 13,845.688 1,369.571 13,793.479 1369.547

All counts are averages over all 10,000 samples. Note how comparison of counts of configurations between these networks with strongly different densities is problematic because
higher configuration counts can be expected merely due to the higher network density.

Frontiers in Digital Humanities  |  www.frontiersin.org 10 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 6 | Densities (D) of visibility networks between differently distributed observation locations in a completely flat square landscape of 9 km by 9 km. The
following experiment settings were used for Figures 6–8: maximum viewing distance = 5 km; curvature of the earth taken into account; elevation observation
points = 1.6 m; elevation observed point = 0 m.

cases, a theoretical maximum distance can be determined up they are derived from. Figure  7 shows visibility networks in a
to which features of interest are distinguishable by the observer hilly landscape in Grande-Terre and Figure  8 shows visibility
(e.g., Gillings, 2015). The effect maximum viewing distance has networks in a particularly mountainous landscape in Basse-Terre
on the network density of visibility networks in a flat landscape (both in Guadeloupe, French West Indies). Most obvious is that
is therefore to connect pairs of nodes spaced at distances shorter visibility networks in flat landscapes (Figure 6) are far denser than
than the maximum viewing distance and not to connect pairs those in hilly and mountainous landscapes (Figures  7 and 8).
of nodes at longer distances. This is illustrated most clearly in Moreover, a higher number and overall spatial density of obser-
the patterned distribution shown in Figure 6, where nodes posi- vation locations (points-per-area) does not necessarily lead to a
tioned in the four corners of a flat landscape are not intervisible higher visibility network density, since network density is relative
because they are too far away from each other. to the number of nodes in the network: the higher the number
However, the examples shown in Figures  7 and 8 caution of nodes, the more edges that need to be created to get a certain
against generalizing trends in visibility network density derived network density score. High local spatial densities of observation
from unrealistically flat landscapes: visibility network density can locations often lead to clusters in the visibility network between
be as variable as the landscapes and the observer distributions closely spaced sets of nodes, but can also lead to lower visibility

Frontiers in Digital Humanities  |  www.frontiersin.org 11 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 7 | Densities (D) of visibility networks between differently distributed observation locations in a hilly landscape (Grande-Terre in Guadeloupe, French West
Indies) of 9 km by 9 km. In Figures 7–10, the topography is only illustrated once in the top left to more clearly show differences in network structures.

network densities if these clusters are spaced at distances higher locations are positioned far from each other, observer and target
than the maximum viewing distance (see patterned distributions elevations are low, and there is much visual obstruction caused
in Figures 6–8). A factor that does clearly lead to higher visibility by topography.
network densities is higher elevations of observer and target loca- In landscape archeology research contexts, there can be enor-
tions (e.g., through the use of observation towers), as shown in mous variety in the spatial density (points-per-area) and distri-
Figures 9 and 10 where elevations of 10 m were assumed. We will bution of observation points, the theorized maximum viewing
not address the cases of vegetation and atmospheric conditions distance, and the topography of the research area. However, the
here, although they can be expected to further reduce visibility archeological literature review presented above revealed that,
network density through obstruction and limiting maximum so far, visibility networks have been mainly used in research
viewing distance. contexts where the intervisibility of large features is considered,
We did not aim to provide a complete overview of visibility like site areas, towns, long barrows, or hillforts, that are rather
network density variability, but merely a range of examples that sparsely distributed over large and often hilly to mountainous
nevertheless allow for some summarizing remarks. Dense vis- landscapes and only a single observation point per feature is
ibility networks can be expected in cases where most observation considered. In such cases, one can expect rather sparse visibility
locations are positioned close to each other (within the maximum networks if past communities did not consider intervisibility
viewing distance), observer and target elevations are high, and between features in deciding on their locations, and slightly
there is little visual obstruction caused by topography. Sparse denser networks if features are positioned on great vantage
visibility networks can be expected in cases where observation points that offer lines-of-sight to other features. Higher densities

Frontiers in Digital Humanities  |  www.frontiersin.org 12 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 8 | Densities (D) of visibility networks between differently distributed observation locations in a mountainous landscape (Basse-Terre in Guadeloupe, French
West Indies) of 9 km by 9 km.

can be expected in research contexts where high observation “structural zeroes”: edges that we know cannot possibly exist.
or observed elevations are assumed (e.g., Figures  9 and 10), Third, the previous two techniques can be combined: a model
such as lookout towers or smoke columns, where multiple where the simulated network density is equal to the observed
observation locations would be used throughout site areas network density, and no edges can be created between nodes
(e.g., Figures  1C,D), or where visibility over huge distances farther removed from each other than the maximum viewing
can be assumed, such as for the use of fire beacons in visual distance. Table  3 shows the results for these three techniques
signaling networks. applied to the total viewshed visibility network introduced in
The impact on visibility network density of a research con- the previous section: we notice that in this particular case fix-
text’s particular landscape topography and assumed maximum ing, the network density has a much stronger effect on results
viewing distance suggests to us that a convincing visibility since the research area is small and there are few node pairs
network model should account for these elements. For example, spaced at distances larger than the maximum viewing distance
we could modify the Bernoulli random graph model presented (i.e., few structural zeroes). However, in all cases, the observed
in the previous section in three ways. First, when randomly network configuration counts are still very different from those
creating networks the network density can be fixed to that of randomly generated networks. Since this is merely an abstract
of the observed network. Second, when randomly creating example to illustrate these techniques, we will not interpret
networks we can prevent edges being created between node these results any further here but rather apply them in a more
pairs that are spaced at distances higher than the maximum interesting archeological case study in Section “Intervisibility of
viewing distance. This can be done by marking these edges as Long Barrows in Cranborne Chase” below.

Frontiers in Digital Humanities  |  www.frontiersin.org 13 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 9 | Densities (D) of visibility networks between differently distributed observation locations, with 10 m observer and target elevations, in a hilly landscape
(Grande-Terre in Guadeloupe, French West Indies) of 9 km by 9 km. The following experiment settings were used for Figures 9 and 10: maximum viewing
distance = 5 km; curvature of the earth taken into account; elevation observation points = 10 m; elevation observed point = 10 m.

ERGMs FOR VISIBILITY NETWORKS of particular configurations in observed networks with their
frequency in stochastic models that include selected configura-
Counts of configurations provide formal descriptions of tions as effects. This approach applied to the study of visibility
the visibility networks, but on their own they do not reveal networks is described elsewhere (Brughmans et al., 2014), and
much about the processes giving rise to this network. We can detailed technical documentation of the ERGM method and
explore such processes by comparing the structural features all configurations is available (Lusher et  al., 2013, and see in
of the observed network with those of models that represent particular for geospatial ERGMs including edge length, Robins
our archeological theories about the processes governing the and Daraganova, 2013, 99–101). We further illustrate the use of
creation and evolution of this network. Does a model with a ERGMs in the brief case study below.
tendency to create paths to enable visual communication and There are a number of advantages to exploring our theories
visually prominent nodes but no isolated nodes offer a good surrounding visibility networks through stochastic statistical
description of the observed network? Such questions can be models. First, it provides a formal representation of our archeo-
addressed with ERGM, an approach belonging to a family of logical theories that facilitate communication as well as formal
statistical models originally developed for social networks comparison with visibility networks in other geographical or
(Wasserman and Pattison, 1996; Anderson et al., 1999). ERGMs temporal contexts. Second, our theories are often rather com-
aim to investigate the dependence assumptions underpinning plex, involving multiple effects that can influence each other and
hypotheses of network formation by comparing the frequency cannot be easily derived from merely exploring the structural

Frontiers in Digital Humanities  |  www.frontiersin.org 14 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

FIGURE 10 | Densities (D) of visibility networks between differently distributed observation locations, with 10 m observer and target elevations, in a mountainous
landscape (Basse-Terre in Guadeloupe, French West Indies) of 9 km by 9 km.

TABLE 4 | Density of example visibility networks shown in Figures 6–10.

Node distribution Landscape topography

Flat Hilly Mountainous

Observer elevation 1.6 m (Figure 6) 1.6 m (Figure 7) 10 m (Figure 9) 1.6 m (Figure 8) 10 m (Figure 10)
36 nodes Regular 0.378 0.002 0.035 0.011 0.030
Patterned 0.270 0.008 0.056 0.022 0.035
Random 0.416 0.016 0.078 0.016 0.033
100 nodes Regular 0.436 0.005 0.045 0.003 0.034
Patterned 0.244 0.012 0.070 0.020 0.036
Random 0.487 0.009 0.058 0.025 0.038

features of the observed visibility network or even counting the to discard and refine the theories they formulate by evaluating
frequency of configurations. Third, a stochastic model introduces whether they provide a plausible description of the observed net-
a degree of uncertainty resulting in a distribution of possible work which is more likely than that the network was merely the
visibility networks, which is appropriate given the uncertainty result of chance. In the remainder of this paper, we will illustrate
inevitably associated with most archeological theories about vis- this approach by applying it to an early and highly influential
ibility networks. Fourth, the approach can allow archeologists example of the archeological use of visibility networks.

Frontiers in Digital Humanities  |  www.frontiersin.org 15 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

INTERVISIBILITY OF LONG BARROWS

IN CRANBORNE CHASE
Christopher Tilley (1994) studied the intervisibility of Neolithic
long barrows at Cranborne Chase (United Kingdom) using a
network representation in which nodes represent long barrows
and edges represent their intervisibility. The network was deter-
mined from observations during multiple visits in clear weather
conditions during a period of 5 weeks, and when a clear line-of-
sight was impossible due to obstructions like buildings or dense
woodland. Tilley made observations from locations closeby or
inferred them from a map of the landscape’s elevation. Through
an evaluation of the presumed past land cover, it is argued that
the observed lines-of-sight are similar to, or underestimate, the
degree of visibility in the Neolithic, especially for the more prom-
inent barrows: “there is little evidence to suggest that barrows
which are prominently sited and intervisible today might not also
have been during the period of their initial construction and use”
(Tilley, 1994, 157). The network is subsequently explored visually
and by comparing node degree. Tilley argues that some barrows
were purposefully positioned with reference to visibility and
suggests a process as an explanation of the network patterns he
identified through exploratory analysis of the observed visibility
network. This presents a highly insightful study of a visibility net- FIGURE 11 | Subnetwork at the center of the Cranborne Chase study area
work in an archeological context, of how the network might have of the visibility network of long barrows by Tilley (1994) (Fig. 5.5).
evolved and how visibility of long barrows might have structured
past human behavior.
In this section, we will show how applying additional net- this network still has three isolated nodes 14, 38, and 27 at the
work science techniques provide alternative descriptions of this center of the study area.
network and the suggested process, which might lead to new An important structural feature for the center of the study
insights. We will illustrate how structural features of particular area is a difference in the degree of nodes and the components
interest can be identified, how we can formally represent the to which they belong: “In general, it is possible to distinguish
hypothesized process, and explore to what extent it is in fact between (i) those monuments in which visibility was a primary
an appropriate description of the observed visibility network. concern in their location, situated on local high points, false crests
This case study merely aims to illustrate these methods and it is and skylines and (ii) those in which visibility was not a dominant
outside our scope to evaluate and replicate Tilley’s visibility data concern, situated at lower points in the local terrain” … “It is
collection. We therefore use the visibility network from Figure of interest to note that those barrows with the highest degree of
5.5 in Tilley (1994, p. 156; verified through personal communica- intervisibility with others form members of pairs of larger bar-
tion with the author) for a secondary analysis. We will discuss row groups” (Tilley, 1994, pp. 158–159). This structural feature
each structural feature of the network that was considered of is revealed by the degree and connected component measures
interest by Tilley, and how they can be explored using network shown in Table 5. The network consists of two large connected
science methods. components, referred to by Tilley as the eastern and western
A first structural feature is a tendency for long barrows groups. In addition to these two, there is a smaller component con-
on the periphery of the study area to be isolated: “Part of the sisting of four nodes, and three isolated nodes forming their own
process of siting barrows in the center of Cranborne Chase components. On average, long barrows are intervisible with about
was their relationship to other barrows and the Cursus while, 3 other long barrows (see average degree in Table 5). However,
on the periphery, a relationship to topographic features of the when we plot the distribution of the number of lines-of-sight per
landscape, rather than to other monuments, appears to have long barrow then we notice a strong difference between them
been of paramount significance” (Tilley, 1994, 158). This sug- (Figure 12): the majority have a low degree (the most frequently
gests different processes are responsible for the creation of the occurring degree is 1), and a few nodes have a particularly high
network structure in the center and the periphery of the study degree (long barrow 15 is intervisible with 9 other long barrows,
area. We will therefore focus our study on understanding the long barrow 10 is intervisible with 7 other long barrows).
structural properties of the center of the study area (for this A further structural feature for the center of the study area
reason, the isolated peripheral nodes 29, 30, 31, 32, 33, 34, and is a tendency for nodes to be connected and not be isolated:
39 are excluded from our analyses presented here). This central “Being able to see other barrows from each mound was clearly
subnetwork is shown in Figure 11, and the count of configura- an important factor in the location of many of them” (Tilley,
tions and summary statistics are presented in Table 5. We notice 1994, p. 158). The network measures in Table 5 reveal that this

Frontiers in Digital Humanities  |  www.frontiersin.org 16 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

TABLE 5 | Network measures and count of visibility network configurations in the subnetwork at the center of the Cranborne Chase study area shown in Figure 11
above.

Configurations Observed network Random network Fixed density Network patterns

Nodes 32 32 32
Edges 49 247.702 49
Density 0.099 0.499 0.099
Isolates 3 0 1.146

Triangle 22 617.907 4.504

3-path 628 55,607.46 407.347

Alternating-star 102.789 862.827 94.015

FIGURE 12 | Degree distribution of subnetwork at the center of the Cranborne Chase study area shown in Figure 11 above.

network has 49 edges and 32 nodes of which 3 are isolated, with which provided the network measures shown in the second
a network density of 0.099. This network density score means column of Table  5. By comparing the structural features of the
that about 10% of all possible edges in a network with 32 vertices observed network (first column) with those of the randomly cre-
are present, which is actually quite dense when compared to the ated networks (second column), we notice they are very different:
densities of example physical landscapes in Table 4. We did after the network density in the latter is much higher, triangles, 3-path,
all purposefully focus on a smaller area at the center of Tilley’s and alternating-star are all much more frequent, and there are no
study area where the network is particularly dense. isolates. However, we argued above that this high network density
But what do these numbers mean: is this network density is to be expected and that this is not an appropriate comparison.
higher than expected by chance for a network of this many nodes? To represent a random network creation process more appropriate
To evaluate this, we created a Bernoulli random graph model, for this particular research context, we can fix the network density
simulating 10,000 randomly generated networks with 32 vertices, to that of the visibility network observed by Tilley (we will not use

Frontiers in Digital Humanities  |  www.frontiersin.org 17 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

structural zeroes since no theoretical maximum viewing distance TABLE 6 | Two estimated exponential random graph modeling models with a
good fit to the observed visibility network shown in Figure 11 above.
is used in Tilley’s original study). The results of this model are
shown in the third column in Table 5, and they are much more Effects Estimates SE Significant? GOF t-ratios
similar to the observed network than those of the previous model.
MODEL 1
However, the number of triangles, 3-paths, and alternating-stars Edge −3.761 0.960 * −0.056
are now all quite a bit less frequent than in the observed network. Triangle 1.598 0.229 * −0.013
These results suggest that a random network creation process does 3-path −0.041 0.017 * −0.043
not describe the observed visibility network very well. Indeed, Alternating-star 0.607 0.415 −0.060
we do not expect this network to have been randomly created by MODEL 2
Edge −7.567 2.480 * 0.000
individuals in the past, but it is crucial that we confirm this before
Triangle 1.610 0.227 * 0.065
testing more interesting archeological theories. It is important to Isolates −2.529 1.450 −0.063
show that a random network creation model does not offer a good 3-path −0.059 0.023 * 0.000
description of the observed network, because this represents the Alternating-star 1.913 0.921 * −0.005
simplest relational assumption: there is no assumption about the Results are significant if the absolute value of the estimate is at least two times
dependency between edges. Excluding a random process allows us the absolute value of the SE. Goodness-of-fit (GOF) t-ratios should be lower
to explore more interesting and more complex processes. However, than 0.1 for the effects included in the model; this t-ratio is derived by comparing
the observed count of a configuration with those sampled from simulations:
if a random network creation process were to offer a good descrip-
t-ratio = (observation − sample mean)/SE.
tion of the observed network then we would have to conclude that
our ability to attach any confidence to a more complex network
TABLE 7 | Goodness-of-fit t-ratio results for models 1 and 2 presented in
creation model with dependency assumptions is very limited.
Table 6 above.
In addition to Tilley’s identification and interpretation of
key structural features mentioned above, the author suggests a Effects t-ratio MODEL 1 t-ratio MODEL 2
hypothetical process that could have led to the observed network
Edge 0.016 −0.029
structure: “One explanation for this pattern might be that sites 2-star 0.198 0.073
that were particularly important in the prehistoric landscape and 3-star 0.426 0.302
highly visible ‘attracted’ other barrows through time, and sites 4-star 0.641 0.648
built later elsewhere were deliberately sited so as to be intervisible 5-star 0.885 1.024
Triangles 0.107 0.013
with one or more other barrows. In this manner the construction 4-clique −0.388 −0.514
of barrows on Cranborne Chase gradually created a series of visual 5-clique −0.29 −0.367
pathways and nodal points in the landscape” (Tilley, 1994, p. 159). 6-clique −0.193 −0.214
To represent this theory, we can create a network model where 7-clique −0.133 −0.136
the creation of edges happens respective to the presence or absence Isolates −0.405 −0.006
Triangle2 −0.215 −0.309
of other edges, rather than generating edges independent of one Bow_tie −0.068 −0.04
another as we did with the Bernoulli random graph model. This 3-path 0.084 −0.003
model should include a tendency for edges to be created to nodes 4Cycle 0.317 0.336
that already have a higher degree, which can be represented by Alternating-star 0.038 −0.026
AT(2.00) 0.438 0.326
the alternating-star configuration. We can add more dependence
A2P(2.00) −0.1 −0.24
assumptions to the model to represent the other structural features AC(2.00) −0.407 −0.537
that Tilley described as important: a tendency to create lines-of- AET(2.00) 0.065 −0.044
sight (represented by the edge configuration), visual pathways SD degree dist 0.694 0.505
(represented by the 3-path configuration), and clusters or groups Skew degree dist 1.033 0.79
Global clustering 0.181 0.093
of intervisible barrows (represented by the triangle configuration).
Mean local clustering 0.433 0.464
We estimated an ERGM with the effects shown in Table 6. The Variance local clustering 0.06 0.054
results of a goodness-of-fit test are shown in Table 7, suggesting
Effects included in the model shown in bold.
that the model is a rather good fit but not perfect since the trian-
gle parameter has a t-ratio just higher than 0.1 (here we use the
criteria by Harrigan (2007) to determine what makes a good fit: not be over-interpreted since it is not significant. The negative
estimated parameters in the model should have a t-ratio below 0.1 significant edge parameter suggests intervisibility of long barrows
and all others below 2). With the aim of obtaining a model with occurs relatively rarely, especially if they are not part of triangles
an even better fit, we decided to add to the model the tendency for and stars. Such a low network density can be expected in geo-
long barrows in this central part of the study area to be intervisible graphical networks where nodes are spread over large landscapes
with at least one other long barrow and therefore not be isolated with complex topographies: the distance and obstructions might
(represented by the isolates configuration). The estimates of this make many lines-of-sight between pairs of long barrows physi-
second model are also shown in Table 6 and the goodness-of-fit cally impossible (see section 6 above). The positive significant
results in Table 7, this time the model is a perfect fit. triangle effect suggests a tendency toward clusters of intervisible
This second model suggests a tendency against long barrows long barrows, as was argued by Tilley. The positive significant
being isolated as was argued by Tilley, although this result should alternating-star parameter suggests a tendency for some long

Frontiers in Digital Humanities  |  www.frontiersin.org 18 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

barrows to be intervisible with far more long barrows than others. The brief study of Christopher Tilley’s (Tilley, 1994) visibility
This is in line with Tilley’s theory that some long barrows might network of long barrows in Cranborne Chase illustrated many of
have “attracted” others, that for some long barrows intervisibility these methods. Its contribution to Tilley’s study was twofold: (1)
was more important than for others, and that some might be to provide a formal representation of the important structural fea-
located to ensure intervisibility with several existing long bar- tures and the hypothesis describing the evolution of the visibility
rows in the landscape. This model can therefore not disprove that network formulated by Tilley, which enables replication of results
this process might have been deliberate. However, the negative as well as formal comparisons with visibility networks in other
significant 3-path parameter suggests this process did not create research contexts and (2) to statistically evaluate the probability of
“a series of visual pathways” as argued for by Tilley, although it Tilley’s theory of the visibility network’s creation and evolution.
definitely does create “nodal points” with a higher degree than We argue that the methods illustrated here are particularly appro-
other long barrows. priate in research contexts where particular structural features are
inherent to the formulation of theories about visibility networks
CONCLUSION (like Tilley’s visibility network of long barrows) and provide
formal approaches to complement rich contextual studies of how
Visibility networks have been used to represent a wide range lines-of-sight could have structured past human behavior.
of relational phenomena in disciplines as diverse as archeol-
ogy, cognitive sciences, architecture, and urban planning. We SOFTWARE
have argued that there is room for cross-fertilization between
archeology and architecture and landscape studies in the use of Viewsheds and visibility networks were created in QGISv2.12.1
network data for representing visibility-related phenomena. We using the open source Viewshed Analysis plugin v0.5.1 by Zoran
proposed a number of new visibility network data representa- Čučković (2016a,b). Configuration counts, Bernoulli random
tions for landscape archeology: first- and second-order visibility graph models, and ERGMs were performed using PNet (Wang
graph representations of total and cumulative viewsheds and et al., 2009).
two-mode representations of cumulative viewsheds. Through
a range of examples, we further explored the variability in AUTHOR CONTRIBUTIONS
visibility network density in physical landscapes, leading us to
argue for the need to explicitly incorporate network density in TB and UB: conception and design of work. TB: drafting of work.
stochastic models of visibility networks. But most importantly,
we identified a very restricted use of formal network science ACKNOWLEDGMENTS
measures for describing, exploring, and analyzing visibility
networks, and in particular of the representation and study of We thank Christopher Tilley for his support in verifying the inter-
theories about how lines-of-sight affect each other’s existence visibility network of Cranborne Chase used here. We thank Zoran
(i.e., dependence assumptions). Čučković for the development of his Viewshed Analysis QGIS
In this paper, we have focused on the archeological study of plugin and his technical support. We thank Viviana Amati for
visibility networks to suggest methods for the formal study of advice concerning ERGMs. We thank the reviewers for their con-
archeological relational visibility theories. We suggested a range structive and detailed feedback, and specifically Irmela Herzog
of configurations to represent and study such theories as network for suggesting the color coding used in Figure 3 and helping to
patterns, and statistical simulation methods for comparing these improve the presentation of the paper. The work presented here
theories with the archeologically observed (or inferred) visibility was performed as part of the Caribbean Connections: Cultural
network. Ultimately, these topological network methods shall Encounters in a New World Setting project, financially supported
incorporate geographical constraints, and possibly processes of by the HERA Joint Research Program, and the European Union’s
topographical reshaping. These approaches should always go Seventh Framework Programme for research, technological
hand in hand with the critical formulation of relational visibility development, and demonstration under grant agreement no 1133;
theories, and results of such formal methods should always be this research is also part of the project NEXUS1492 (http://www.
interpreted in light of these theoretical frameworks, addressing nexus1492.eu/), which has received funding from the European
issues such the ability of observers to recognize observed features Research Council under the European Union’s Seventh Framework
and the temporality of feature creation. Programme (FP7/2007-2013)/ERC grant agreement no. 319209.

REFERENCES Benedikt, M.L. (1979). To take hold of space: isovists and isovist fields. Environment
and Planning B: Urban Analytics and City Science 6: 47–65. doi:10.1068/b060047
Ahmadlou, M., Adeli, H., and Adeli, A. (2010). New diagnostic EEG markers of the Brandes, U., Robins, G., McCranie, A., and Wasserman, S. (2013). What is network
Alzheimer’s disease using visibility graph. Journal of Neural Transmission 117: science? Network Science 1: 1–15. doi:10.1017/nws.2013.2
1099–109. doi:10.1007/s00702-010-0450-3 Brughmans, T., de Waal, M.S., Hofman, C.L., and Brandes, U. (2017). Exploring
Anderson, C.J., Wasserman, S., and Crouch, B. (1999). A p* primer: logit transformations in Caribbean indigenous social networks through visibility
models for social networks. Social Networks 21: 37–66. doi:10.1016/ studies: the case of late pre-colonial landscapes in East-Guadeloupe (French
S0378-8733(98)00012-4 West Indies). Journal of Archaeological Method and Theory 1–45. doi:10.1007/
Batty, M. (2001). Exploring isovist fields: space and shape in architectural and s10816-017-9344-0
urban morphology. Environment and Planning B: Planning and Design 28: Brughmans, T., Keay, S., and Earl, G. (2015). Understanding inter-settlement visi-
123–50. doi:10.1068/b2725 bility in Iron Age and Roman Southern Spain with exponential random graph

Frontiers in Digital Humanities  |  www.frontiersin.org 19 August 2017 | Volume 4 | Article 17

Brughmans and Brandes Visibility Network Patterns and Methods

models for visibility networks. Journal of Archaeological Method and Theory 22: Lusher, D., Koskinen, J., and Robins, G. (2013). Exponential Random Graph Models
58–143. doi:10.1007/s10816-014-9231-x for Social Networks. Cambridge: Cambridge University Press.
Brughmans, T., Keay, S., and Earl, G.P. (2014). Introducing exponential random Mallot, H.A., Franz, M.O., Schölkopf, B., and Bülthoff, H.H. (1997). The view-
graph models for visibility networks. Journal of Archaeological Science 49: graph approach to visual navigation and spatial memory. In Proc. 7th Int. Conf.
442–54. doi:10.1016/j.jas.2014.05.027 Artif. Neural Networks ICANN97, Lausanne, Switzerland.
Conolly, J., and Lake, M. (2006). Geographical Information Systems in Archaeology. Nagy, G. (1994). Terrain visibility. Computers & Graphics 18: 763–73.
Cambridge, NY: Cambridge University Press. Available at: http://www.loc.gov/ doi:10.1016/0097-8493(94)90002-7
catdir/enhancements/fy0665/2006296605-d.html O’Sullivan, D., and Turner, A. (2001). Visibility graphs and landscape visibility
Čučković, Z. (2016a). Advanced viewshed analysis: a quantum {GIS} plug-in analysis. International Journal of Geographical Information Science 15: 221–37.
for the analysis of visual landscapes. The Journal of Open Source Software 1. doi:10.1080/13658810010011393
doi:10.21105/joss.00032 Puppo, E., and Marzano, P. (1997). Discrete visibility problems and graph algo-
Čučković, Z. (2016b). Viewshed Analysis v0.5.1 QGIS Plugin. Available at: rithms. International Journal of Geographical Information Science 11: 139–61.
https://github.com/zoran-cuckovic/QGIS-visibility-analysis. doi:10.1080/136588197242437
De Floriani, L., Marzano, P., and Puppo, E. (1994). Line-of-sight communication Robins, G.L., and Daraganova, G. (2013). Social selection, dyadic covariates, and
on terrain models. International Journal of Geographical Information Systems 8: geospatial effects. In Exponential Random Graph Models for Social Networks,
329–42. doi:10.1080/02693799408902004 Edited by D. Lusher, J. Koskinen, and G.L. Robins, 91–101. Cambridge:
De Montis, A., and Caschili, S. (2012). Nuraghes and landscape planning: coupling Cambridge University Press.
viewshed with complex network analysis. Landscape and Urban Planning 105: Ruestes Bitrià, C. (2008). A multi-technique GIS visibility analysis for studying
315–24. doi:10.1016/j.landurbplan.2012.01.005 visual control of an Iron Age landscape. Internet Archaeology 23. doi:10.11141/
Earley-Spadoni, T. (2015). Landscapes of warfare: intervisibility analysis of early ia.23.4
iron and urartian fire beacon stations (Armenia). Journal of Archaeological Ruiter, S. de. (2012). Mapping History: An Analysis of Site Locations in the
Science: Reports 3: 22–30. doi:10.1016/j.jasrep.2015.05.008 Northwestern Dominican Republic. Unpublished RMA thesis.
Elsner, J.B., Jagger, T.H., and Fogarty, E.A. (2009). Visibility network of United States Schölkopf, B., and Mallot, H.A. (1995). View-based cognitive mapping and path
hurricanes. Geophysical Research Letters 36: 1–5. doi:10.1029/2009GL039129 planning. Adaptive Behavior 3: 311–48. doi:10.1177/105971239500300303
Eve, S.J., and Crema, E.R. (2014). A house with a view? Multi-model inference, Shemming, J., and Briggs, K. (2014). Anglo-Saxon Communication Networks.
visibility fields, and point process analysis of a Bronze Age settlement on Available at: http://keithbriggs.info/AS_networks.html
Leskernick Hill (Cornwall, UK). Journal of Archaeological Science 43: 267–77. Swanson, S. (2003). Documenting prehistoric communication networks: a case study
doi:10.1016/j.jas.2013.12.019 in the Paquimé polity. American Antiquity 68: 753–67. doi:10.2307/3557071
Fisher, P.F. (1994). Probable and fuzzy models of the viewshed operation. Tilley, C.Y. (1994). A Phenomenology of Landscape. Places, Paths and Monuments.
In Innovations in GIS: Selected Papers from the First National Conference on GIS Oxford: Berg.
Research UK, Edited by M.F. Worboys, 161–175. London: Taylor & Francis. Turner, A., Doxa, M., O’Sullivan, D., and Penn, A. (2001). From isovists to visibility
Fortunato, S. (2010). Community detection in graphs. Physics Reports 486: 75–174. graphs: a methodology for the analysis of architectural space. Environment and
doi:10.1016/j.physrep.2009.11.002 Planning B: Planning and Design 28: 103–21. doi:10.1068/b2684
Franz, G., Mallot, H.A., and Wiener, J.M. (2005). Graph-based models of space in Turner, A., Penn, A., and Hillier, B. (2005). An algorithmic definition of the
architecture and cognitive science – a comparative analysis. In Proc. 17th Int. axial map. Environment and Planning B: Planning and Design 32: 425–44.
Conf. Syst. Res. Informatics Cybern, 30–38. doi:10.1068/b31097
Fraser, D. (1980). The cutpoint index: a simple measure of point connectivity. Area Wang, P., Robins, G., and Pattison, P. (2009). PNet. Program for the Simulation
12: 301–4. and Estimation of Exponential Random Graph (p*) Models. User Manual.
Fraser, D. (1983). Land and Society in Neolithic Orkney. BAR British Series 117. Available at: http://sna.unimelb.edu.au/__data/assets/pdf_file/0006/662865/
Oxford: Archaeopress. PNetManual.pdf
Gillings, M. (2009). Visual affordance, landscape, and the megaliths of Alderney. Wasserman, S., and Pattison, P. (1996). Logit models and logistic regressions for
Oxford Journal of Archaeology 28: 335–56. doi:10.1111/j.1468-0092.2009.00332.x social networks: I. An introduction to Markov graphs and p*. Psychometrika 6:
Gillings, M. (2015). Mapping invisibility: GIS approaches to the analysis of hiding and 401–25. doi:10.1007/BF02294547
seclusion. Journal of Archaeological Science 62: 1–14. doi:10.1016/j.jas.2015.06.015 Wheatley, D.W. (1995). “Cumulative Viewshed Analysis: a GIS-based method for
Harrigan, N. (2007). PNet for Dummies: An Introduction to Estimating Exponential investigating intervisibility, and its archaeological application,” in Archaeology
Random Graph (p*) Models with PNet. Available at: http://www.mysmu.edu/ and Geographical Information Systems: a European Perspective, eds G. Lock and
faculty/nharrigan/PNetForDummies.pdf Z. Stančič (London: Taylor & Francis), 171–186.
Hillier, B., and Hanson, J. (1984). The Social Logic of Space. Cambridge: Cambridge Wheatley, D., and Gillings, M. (2000). Vision, perception and GIS : developing
University Press. enriched approaches to the study of archaeological visibility. In Beyond the Map:
Koskinen, J., and Daraganova, G. (2013). Exponential random graph model fun- Archaeology and Spatial Technologies, Edited by G.R. Lock, 1–27. Amsterdam:
damentals. In Exponential Random Graph Models for Social Networks, Edited IOS Press.
by D. Lusher, J. Koskinen, and G. Robins, 49–76. Cambridge: Cambridge
University Press. Conflict of Interest Statement: The authors declare that the research was con-
Lacasa, L., Luque, B., Ballesteros, F., Luque, J., and Nuño, J.C. (2008). From time series ducted in the absence of any commercial or financial relationships that could be
to complex networks: the visibility graph. Proceedings of the National Academy of construed as a potential conflict of interest.
Sciences of the United States of America 105: 4972–5. doi:10.1073/pnas.0709247105
Lake, M.W., and Woodman, P.E. (2003). Visibility studies in archaeology: a review Copyright © 2017 Brughmans and Brandes. This is an open-access article distributed
and case study. Environment and Planning B: Urban Analytics and City Science under the terms of the Creative Commons Attribution License (CC BY). The use,
30: 689–707. doi:10.1068/b29122 distribution or reproduction in other forums is permitted, provided the original
Llobera, M. (2003). Extending GIS-based visual analysis: the concept of visuals- author(s) or licensor are credited and that the original publication in this journal
capes. International Journal of Geographical Information Science 17: 25–48. is cited, in accordance with accepted academic practice. No use, distribution or
doi:10.1080/13658810210157732 reproduction is permitted which does not comply with these terms.

Frontiers in Digital Humanities  |  www.frontiersin.org 20 August 2017 | Volume 4 | Article 17