Archeologia e Calcolatori

20, 2009, 277-290

PREDICTIVE MODELLING OF ROMAN SETTLEMENT

IN THE MIDDLE TIBER VALLEY

1. Introduction
The aim of this research is to apply and evaluate the potential of predic-
tive modelling as a means to further understanding ancient settlement in the
middle Tiber valley. This area is one of the most intensively surveyed regions
of the Mediterranean yet even here the extent of archaeological knowledge
is uneven and some areas have yet to be subject to any systematic survey at
all.
Since 1997, the area has been the focus of the British School at Rome’s
Tiber Valley Project, under the direction of Dr. Helen Patterson. The overall
goal of the project is to study the changing landscape of the middle Tiber valley
from protohistory to the medieval period (Fig. 1). It examines the impact of
the growth, success and transformation of the city of Rome on the history of
settlement and economy in the river valley (for project aims and results, see
Patterson 2004; Patterson et al. 2004b; Coarelli, Patterson 2008).

2. The dataset
The data used in this analysis have been generated through a restudy
of the material from John Ward-Perkins’ South Etruria Survey carried out
between the 1950s and 1970s, and their integration with data from more
recent �eld work and numerous published surveys and excavations. The data
were collected and integrated by two Leverhulme-funded fellows, Helga Di
Giuseppe and Robert Witcher; the restudy of the South Etruria survey material
was undertaken by twelve ceramic specialists (Patterson et al. 2004a, and
papers in Coarelli, Patterson 2008). The data are housed in a relational
database, which includes over 5500 �ndspots. As well as the management
and archiving of data, this systemisation also opens up potential for spatial
analyses (Kay, Witcher 2005; Witcher, Kay 2008). Each record includes
spatial coordinates that enable visualisation within a GIS. The precision of
�eld recording for the bulk of sites is 100 metres, and this relatively coarse
resolution is re�ected in the predictive models developed below.

3. Background of application
The process of data collation emphasised that, whilst most areas have
produced some evidence of human activity, not all areas have been subject to

277
S.J. Kay, R.E. Witcher

Fig. 1 – Location of Tiber Valley Project study area (black line), main
�eld surveys (in grey) and predictive model case study area (boxed,
see Fig. 2).

the same level of systematic study. For example, it became apparent that there
was a marked contrast in the numbers and chronological development of sites
on either side of the river Tiber (Patterson et al. 2004c). In order to assess
whether this was a genuine pattern, or a product of uneven archaeological
attention, a small �eld survey was undertaken in the Sabina Tiberina on the
east bank (Di Giuseppe et al. 2002). The results suggest that the contrasting
patterns of ancient settlement on either bank of the Tiber are likely to be a
product of different histories of, and approaches to, landscape archaeology.
In particular, the Corese survey was able to document higher densities of
settlement than previously identi�ed in the Sabina Tiberina, including sites
from periods that have traditionally been dif�cult to recognise due to limited
quantities of material culture (for example, the late antique).

278
 Predictive modelling of Roman settlement in the middle Tiber valley

Since the South Etruria survey commenced in the 1950s, some areas have
been lost to quarrying and development, and many individual sites have been
destroyed by erosion. It seems likely that many undiscovered sites have also
been permanently lost. There are also several areas in South Etruria which
were, and which remain, inaccessible, including military training areas and
the Vatican Radio antenna farm. Predictive modelling offers the possibility to
explore the potential archaeological signi�cance of these different gaps and
inconsistencies. Indeed, such work is of critical importance if we are to evalu-
ate the signi�cance of existing settlement patterns and trends. Furthermore,
predictive modelling also has the potential to respond to a question of critical
importance – what percentage of past settlement does �eld survey actually
identify? (Witcher in press).

4. Theoretical perspectives on predictive modelling

The majority of published applications of predictive modelling con-
cern North American case studies (see papers in Wescott, Brandon 2000;
Mehrer, Wescott 2006; for brief introductions, see Lock 2003, 168-70;
Chapman 2006, 157-159). In Europe, applications of the technique have
concentrated on northern countries, particularly the Netherlands, Germany
and France (Lock, Stančič 1995; García Sanjuán, Wheatley 2002; van
Leusen, Kamermans 2005; Verhagen 2007).
Some of this work is driven by speci�c research questions, but much
concerns Cultural Resource Management where predictive modelling has come
to form part of the planning process. There exist very few published applica-
tions of predictive modelling in southern Europe (see van Leusen 2002, 146-
149); as well as Gaffney and Stančič’s (1991) seminal study of the Croatian
island of Hvar, there has also been work by Kamermans (2000) in the Agro
Pontino to the south of Rome, by Stančič and Veljanovski (2000) on the
Croatian island of Brač, and by Rua (2009) in southern Portugal. Closely
related GIS applications include work by van Hove (2004) on “taskscapes”
in Calabria and applications of Historic Landscape Characterization (HLC)
in Greece and Turkey by Crow and Turner (in press).
Of all the analytical techniques that have �ourished since the wide-
spread adoption and use of spatial technologies (for overview, see Wheatley,
Gillings 2002), predictive modelling has been the most heavily critiqued,
particularly from a European perspective. Ebert (2000), and more recently
Wheatley (2004), have outlined a number of criticisms of inductive or data-
driven predictive modelling. These criticisms can be grouped as two main
points: 1) “prediction as explanation”; and 2) environmental determinism.
First is criticism of predictive modelling as explanation. However, pre-
dictive modelling does not aim to explain patterns, rather it aims simply to

279
S.J. Kay, R.E. Witcher

identify patterns. It is the task of the archaeologist to explain and interpret

those patterns, not the model. The second criticism is that predictive model-
ling is “anti-historical” because it «assumes [patterns are] wholly a product of
the immediate surroundings of the individuals and communities» (Wheatley
2004). This argument relates to the observation that most variables used within
predictive models are environmental (for example, slope, distance to water,
etc.) and that social and cultural considerations are excluded. As a result, the
approach is deemed to be “environmentally deterministic”: constraints of the
physical environment replace human agency. However, this wider argument
is built on a confusion of correlation with causation. In common with any
other statistical technique, it is inherent in predictive modelling that statistical
association does not imply a causative relationship between variables. This
returns to the �rst point: predictive modelling as descriptive not explana-
tory; it does not inherently exclude the agency of knowledgeable individuals
or societies to structure their lives. Indeed, Wheatley (2004) himself notes
that «this is not to deny that correlative predictive models may be telling us
something about the behaviour of people in the past». North American ap-
plications have produced relatively powerful models based on environmental
variables; it would be wrong to dismiss this predictive power on the grounds
of how the results are (mis)interpreted.
In general, these arguments betray a series of broader misplaced and
latent concerns. Firstly, that any attempt to involve environmental variables
in an archaeological study is “determinism” by another name. Secondly,
that quanti�cation and statistical analyses are reductionist and, thirdly that
“prediction” is antithetical to free agency. Brie�y, these may be rebutted with
the following responses: discussion of environment does not presuppose
determination (for much more subtle approaches, including environmental
affordance, see Ingold 2000); statistical analyses, if appropriately used,
become rigorous, repeatable investigations to support – not replace – inter-
pretation; and, thirdly, as such studies merely describe, it is the responsibility
of the archaeologist, not the models, to ensure that individuals are granted
appropriate agency and that societies are given cultural autonomy. The best
solution is to place less emphasis on “prediction” per se, and to foreground
the approach as exploratory data analysis and pattern recognition, in other
words, a form of data modelling and characterisation.
In the context of the current research, some other speci�c criticisms
should also be addressed. Firstly, that predictive modelling is concerned only
with sites and has failed to take broader theoretical developments about off-
site activity into account. Whilst this is certainly a valid criticism, in relation
to the present case study, the vast majority of data used here derive from
site-based survey conducted �fty years ago, long before such theoretical and
methodological developments. Whilst not ideal, it would be wrong to discard

280
 Predictive modelling of Roman settlement in the middle Tiber valley

these data as inadequate; indeed, as emphasised above, many of these sites

no longer exist and thus these data form a unique documentary record which
cannot be re-collected. A related issue is criticism that the technique deals
only with a simple binary – site or no-site – and the possibility of more than
one site per unit of land is not addressed (in the current study, South Etruria
survey sites were recorded to the nearest 100 m, resulting in a spatial record-
ing unit of 10000 m2 which could theoretically contain more than a single
site). Whilst recent high-intensity survey suggests that more than one site per
10000 m2 is de�nitely a possibility in this area (Di Giuseppe et al. 2002), the
precision of the original survey recording of 100 m precludes more detailed
consideration at a higher resolution.
A further criticism has noted that many applications lack suf�cient
archaeological data and as a result fail to differentiate between sites of differ-
ent dates and types. For example, in his study of the island of Brač, Stančič
(2000) used 29 sites covering four centuries across 395 km². The present
application uses 288 villas dating to the �rst century AD to train the model
and then tests it against a reserve of a further 288 �rst century villas across
a total area of 1100 km².
A �nal issue concerns the anachronistic nature of much environmental
data used. It is important to include variables that are as chronologically
relevant as possible; hence in the present application, modern land use was
excluded from the �nal and most powerful model. However, such data should
not be dismissed entirely. Whilst they may not necessarily inform about
past settlement decisions, they may well shed light on recovery processes (in
particular, visibility, see Terrenato, Ammerman 1996). Indeed, predictive
modelling can be seen as an heuristic tool for exploring data and identifying
in�uencing factors (such as post-depositional or recovery bias) that present
signi�cant problems for �eld surveyors.

5. Methodology
As described above, the site data are stored in a relational database and
are linked, via SQL commands, with ESRI ArcView; the predictive models
are developed using the Arc-WofE extension (data have been continuously
migrated since the projects inception and are now analysed through ESRI
ArcGIS; future analysis will make use of ArcSDM3.1, Sawatzky et al. 2004).
The Weights of Evidence (WofE) methodology is part of a larger group of
multi-criteria decision-making techniques and is commonly used, for example,
for prospective mapping of mineral deposits. Goodchild (2007, 2008) pro-
vides an archaeological adaptation of multi-criteria decision-making which,
rather than predict site locations, seeks to model agricultural productivity
and population in the middle Tiber valley.

281
S.J. Kay, R.E. Witcher

The WofE model uses statistical associations between known sites called
training points (in this case, early imperial villas) and different map themes
(such as geology, aspect and slope) in order to calculate a set of weights. It
is therefore an inductive approach. These weights are then used to evaluate
every possible combination of the different map layers in order to produce a
single map (a unique conditions grid) showing probability of the presence of
a site. The variables which the model identi�es as important are considered
for any signi�cance in understanding ancient land use or perception of the
landscape (for an archaeological application of WofE, see Hansen n.d.; for
an alternative predictive methodology, also applied in the middle Tiber valley,
see Espa et al. 2006).
The study area comprises c. 1100 km2 in the middle Tiber valley, to the
immediate north of the city of Rome (Fig. 1). A Digital Elevation Model was
derived from contours and spot-heights from the Istituto Geogra�co Militare
1:25.000 map series. From this, maps of slope, aspect and topographical form
(ridge, peak, valley, etc.) were derived. Other themes include geology, modern
land use and rivers. Proximities to three “cultural” variables were also considered:
Roman consular roads, contemporary Roman towns and the city of Rome.
For each theme, its relative weight (or in�uence) on site location is cal-
culated by the WofE extension, by taking into account the spatial extent of
each theme’s attributes (for example, categories of geology) and the number
of training sites present on each. Statistics evaluate the signi�cance of the
association between these sites and each theme and its individual attributes.
For example, for geology, there is a strong aversion to alluvial areas; for
topography there is strong preference for ridges and other convex topo-
graphical forms such as hilltops; an aversion to areas less than 100 m from
watercourses; and an aversion to slope greater than �fteen degrees. The indi-
vidual attributes of different themes can be categorised into varying numbers
of classes (for example, four or eight classes of aspect) to assess the effects
of this generalization on statistical association; similarly, various �lters were
used to derive topographical features generalized to different spatial scales
(generally, see Wood 1996-2008). In each case, the classi�cation producing
the strongest association was used. Statistics were also calculated for themes
based on proximity or distance. As the size and number of classes affects the
output, these had to be carefully de�ned. Weights were graphed to identify
signi�cant cut off points. In the case of rivers, a simple binary theme of <100
m and >100 m to nearest watercourse was used; for proximity to roads, three
categories were used (<1 km, 1-3 km, >3 km); for proximity to towns, three
categories were used (<5 km, 5-10 km, >10 km); and for proximity to Rome,
four bands were used (<20 km, 20-30 km, 30-40 km, >40 km).
On the basis of these statistics, different combinations of themes were
used to generate unique combination grids, or probability surfaces (Kay,

282
 Predictive modelling of Roman settlement in the middle Tiber valley

Witcher 2005, tav. 1). The model was developed as an exploratory process,
with various themes introduced and excluded from the model in order to
identify those combinations which were most predictive. The WofE extension
includes a number of tests to ensure that the statistical assumptions of the
model are not violated; in all the examples described here, these assumptions
were upheld.

6. Results
The best (i.e. most powerful) model achieved with the environmental
and cultural coverages listed above utilised just three environmental themes:
geology (ten classes), topography (six classes) and slope (three classes). Par-
ticularly high probability combinations were level to gently sloping ground
located on the tops of tufo ridges and spurs. The resulting probability surface
is illustrated in Fig. 2. Darker shades indicate higher probability (for exam-
ple, the narrow, fertile ridges in areas A and B); lighter shades indicate lower
probability, most notably, the alluvial soils of the Tiber �oodplain show up
as pale areas.
The predictive model produced a highly complex and fragmented mo-
saic, in which areas of very low probability sit next to areas of high probability.
These results may suggest the very careful localised positioning of sites in
relation, particularly, to topographical form. The signi�cant in�uence of the
strongly dissected topography is re�ected in the thin, parallel strips of land
alternating between high and low probability.
Fig. 3 shows the cumulative percentage of background cells (i.e. random)
and site cells (i.e. the reserve villa sites) against the posterior probability value.
This demonstrates that both the random and the site groups comprise large
numbers of low probability cells and fewer of higher probability. However,
the slower accumulation of site cells indicates that a greater percentage of
sites occur on higher probability cells. Overall, the model has moderate pre-
dictive power, but clearly offers a better-than-chance method of predicting
site location. It effectively predicts c. 20% of villas in just c. 6% of the area,
weakening to c. 53% of villas in c. 26% of the area.
The environmental and cultural themes used in the model have pos-
sible signi�cance in terms of both past human behaviour (for example, site
location preference) and archaeological recovery (for example, visibility).
For example, people may have avoided building sites in valley bottoms due
to �ooding and on steep slopes due to erosion. However there are also post-
depositional possibilities: alluvium may cover sites located in valley bottoms,
whilst the lack of cultivation on steep slopes (there is little use of terracing in
this area) means sites are less likely to be discovered if they did exist. It is, of
course, not straightforward to distinguish between past settlement location

283
S.J. Kay, R.E. Witcher

Fig. 2 – Predictive model for early imperial villas

in the middle Tiber valley. For location, see Fig. 1.
The river Tiber (in black) runs north-south. A =
Central Ager Faliscus; B = area west of Veii; C =
Ager Foronovanus.

Fig. 3 – Comparison of cumulative percentages of background and sites

cells against WofE posterior probability.

284
 Predictive modelling of Roman settlement in the middle Tiber valley

decisions and post-depositional and recovery issues. Of the themes used in

the model, both cultural and environmental, it is clear that environmental
themes have more predictive power. However, this is not to argue that they
are more important in general than cultural factors, but that of the themes
selected here, they have a more important role to play in prediction of where
sites may be found today.
Land use and distance to Rome were employed in alternative models
and both found to have high predictive power. However, both were excluded
from further analysis as land use mapping refers to the modern landscape
and seemed highly likely to re�ect archaeological visibility (for example, pas-
ture was low probability and vineyards high). Similarly, whilst it is possible
that villa density was higher closer to Rome, archaeological activity/survey
intensity has generally been much higher closer to the city. The current model
therefore concentrates on those themes most clearly free of post-depositional
and recovery problems. However, future work will attempt to use these and
other themes speci�cally to distinguish between site location, post-deposition
and recovery issues.
The model obviously predicts where sites might be found if the same
survey methodology were employed again. As such, the model replicates
existing biases (for example, surveyors may have concentrated their work
in areas of high visibility where they believe they will achieve better results).
Nonetheless, it still highlights potential unevenness in survey data. For ex-
ample, the intensity of survey in the central Ager Faliscus (marked A on Fig.
2) and to the west of the city of Veii (B) is known to be particularly low. On
the basis of their environmental similarity with other areas, the model sug-
gests the probability of �nding villas in these areas is high. However, these
existing biases can be addressed through the integration of new and more
systematic �eldwork into the model. Indeed, this addresses another of the
criticisms made by Wheatley (2004), that predictive models are self-ful�lling
prophesies as they reinforce existing biases – however, if modelling is treated
as an iterative process, with new results added in, models can be constantly
re�ned (Rua 2009). With the current model, it is noticeable that there is
some difference in probabilities between the two banks of the Tiber in the
top third of Fig. 2. Survey in both areas is relatively limited and the model is
therefore more strongly in�uenced by discoveries in the southern half of the
study area. However, the area on the west bank (B) in southern Etruria is of
a similar nature in terms of topography, geology, land use, etc. to the better
surveyed areas to its immediate south. In contrast, the area to the east of the
river (C) is relatively unlike other surveyed areas of the Sabina Tiberina on
the eastern bank of the Tiber. The addition to the model of results from recent
survey work in this area (Gabrielli et al. 2003; Verga 2007) may improve
con�dence in the strength of this patterning.

285
S.J. Kay, R.E. Witcher

Perhaps the most striking result is the similarity between early imperial
villa location and some modern settlement as revealed by visualising 1:10.000
Carta Tecnica Regionale maps over the probability model (Kay, Witcher
2005, tav. 2). The model, based on Roman villas, also distinguishes very pre-
cisely between nucleated medieval centres on the one hand (low probability),
and their suburbs and sprawling discontinuous developments of the last thirty
years on the other (high probability). Modern land use was explicitly excluded
from the analysis and there is unlikely to be any correlation as a result of the
preferential discovery of material during house construction, as the bulk of the
data used pre-date such suburban developments. This might suggest possible
similarities in the landscape perceptions, valuations or motivations in�uencing
the settlement location decisions of both Roman villas and modern suburban
settlement. Speci�cally, in contrast with Etruscan and Medieval settlement,
these are open sites with little need or desire to nucleate. Most notably, they
are on ridges. There are advantages to this, such as drainage and expansive
vistas, but also disadvantages such as exposure to wind and inaccessibility of
water. In the latter context, the widespread presence of Roman cisterns in this
area is interesting. But probably most important is the fact that roads tend
to follow the ridges. The tufo landscape across much of this area has created
narrow ridges, divided by steep valleys. The consular road, the Via Flaminia,
follows such a ridge and avoids the need to cross any river for more than
thirty kilometres north of the Milvian Bridge outside Rome (Fig. 1). However,
this is only the most impressive example of countless other ridge roads (that
is, non-consular roads) which were excluded from this preliminary model
because they are unevenly mapped and so may have skewed results.
Finally, the model of early imperial villa location was applied to samples of
sites from other periods. Prediction of Etruscan and mid-republican sites was as
ef�cient as the prediction of sites of imperial date on which the model was based.
Further work is required to interpret this situation, but two (not necessarily
mutually exclusive) explanations for this similarity in site location parameters
can be postulated. First, that settlement in all three periods is similarly located
(despite such signi�cant events as the construction of consular roads and the
economic pressures resulting from the emergence of the imperial metropolis of
Rome). Second, that these patterns are largely the product of post-deposition
and recovery, for example issues of visibility, which may effectively homogenise
differences between the settlement preferences in different periods.
Assessment of the unevenness of the data is vital if other aspects are to
be explored and developed. For example, work by Goodchild (2007, 2008;
also Goodchild, Witcher in press) on the agricultural production potential
in the middle Tiber valley explores issues such as subsistence regimes, popula-
tion and carrying capacities of land units through the use of historical and
comparative evidence, as well as archaeological data. These issues can only

286
 Predictive modelling of Roman settlement in the middle Tiber valley

be reliably addressed through assessment of the completeness or otherwise of

existing patterns. Predictive modelling is one method which may help improve
understanding of the character of archaeological datasets.

7. Conclusion
In summary, the predictive power of the �nal model does not compare
to that produced by some of the North American models. For example, Dalla
Bona’s (2000, 94) model of sites in northern Ontario located c. 84% of known
sites within areas of high archaeological potential which comprised just c.
16% of the area. However, some key differences between North American
and European models can be identi�ed. Firstly, most US models deal with very
few sites. It is possible that there was greater selectivity of prime landscape
locations, whereas in a full and intensively exploited landscape such as that
on the doorstep of imperial Rome, it would seem likely that choice was more
about compromise. Further, in comparing the results presented here with North
American models, it is apparent that in the hinterland of ancient Rome we are
dealing with a more complex agricultural and territorial empire capable of sig-
ni�cantly altering the environment and its potential to support large, sedentary
populations (for example, cisterns for irrigated agriculture or inter-regional
exchange). Pressures to supply the metropolis with agricultural goods and the
desire or political need to live near Rome may have meant that environmental
variables were increasingly less in�uential on settlement location over time.
Lower predictive power may be one reason for the lack of popularity
of such work in the Mediterranean, but a more probable explanation is the
very different developmental pressures and Cultural Resource Management
processes. In particular, there are few areas (in Italy at least) about which ab-
solutely nothing at all is known archaeologically; further, the archaeological
record is densely-distributed; agricultural and developmental pressures are
great, and the extent of areas to be assessed is comparatively small and can
be subject to more intensive reconnaissance.
This study is a preliminary attempt to evaluate the possibilities of pre-
dictive modelling in the middle Tiber valley. In particular, it makes use of a
generic modelling package; whilst this provides a useful initial framework,
future work will seek to move away from a “blackbox” approach and seek to
increase control over the process. Nonetheless, the initial results suggest that
that predictive modelling is a useful heuristic tool to explore site location pref-
erences and archaeological recovery issues in this part of the Mediterranean.
Acknowledgements
This study draws upon the results of the Tiber Valley Project, and in particular on re-
search conducted by Leverhulme fellows, Helga Di Giuseppe and Robert Witcher; the project

287
S.J. Kay, R.E. Witcher

is directed by Helen Patterson of the British School at Rome. The authors would like to thank
all their project colleagues and the former Director of the British School, Andrew Wallace-
Hadrill. The Tiber Valley Project draws together scholars and ceramic specialists from twelve
British universities and many Italian institutions whose work is gratefully recognised here.
Thanks also to Helen Goodchild who read and commented on an earlier draft. Any mistakes
or misinterpretations in the present article are the responsibility of the authors.

Stephen J. Kay
The British School at Rome
Robert E. Witcher
Department of Archaeology
Durham University

REFERENCES

Chapman H.P. 2006, Landscape Archaeology and GIS, Stroud, Tempus.

Coarelli F., Patterson H. (eds.) 2008, Mercator Placidissimus. The Tiber Valley in Antiquity.
New Research in the Upper and Middle River Valley, Quaderni di Eutopia 8, Roma,
Quasar.
Crow J., Turner S. in press, Unlocking historic landscapes in the Eastern Mediterranean using
historic landscape characterisation, «Antiquity».
Dalla Bona L. 2000, Protecting cultural resources through forest management planning
in Ontario using archaeological predictive modelling, in Wescott, Brandon 2000,
73-100.
Di Giuseppe H., Sansoni M., Williams J., Witcher R. 2002, The Sabinensis Ager revisited:
a �eld survey in the Sabina Tiberina, «Papers of the British School at Rome», 70, 99-
150.
Ebert J.I. 2000, The state of the art in “inductive” predictive modelling: seven big mistakes
(and lots of smaller ones), in Wescott, Brandon 2000, 129-134.
Espa G., Benedetti R., De Meo A., Ricci U., Espa S. 2006, GIS based models and estima-
tion methods for the probability of archaeological site location, «Journal of Cultural
Heritage», 7, 3, 147-155.
Gabrielli R., Guidi A., Santoro P. 2003, Il progetto Galantina, in J.R. Brandt, X. Dupré
Raventós, G. Ghini (eds.), Lazio e Sabina 1, Roma, De Luca, 109-114.
Gaffney V., Stančič Z. 1991, GIS Approaches to Regional Analysis: A Case Study of the
Island of Hvar, Ljubljana, Znanstveni Institut Filozofske Fakultete.
García Sanjuán L., Wheatley D.W. (eds.) 2002, Mapping the Future of the Past. Managing
the Spatial Dimension of the European Archaeological Resource, Seville, Universidad
de Sevilla.
Goodchild H. 2007, Modelling Roman Agricultural Production in the Middle Tiber Valley,
Central Italy, University of Birmingham, Unpublished PhD thesis (http://etheses.bham.
ac.uk/175/).
Goodchild H. 2008, Modelling the productive potential of the middle Tiber valley, in
Coarelli, Patterson 2008, 769-796.
Goodchild H., Witcher R.E. in press, Modelling the agricultural landscapes of Republican
Italy, in J. Carlsen, E. Lo Cascio (eds.), Agricoltura e scambi nell’Italia tardo repub-
blicana, Bari-Roma, Edipuglia.
Hansen D.T. 2004, Describing GIS applications: Spatial statistics and weights of evidence ex-
tension in ArcView in the analysis of the archaeology sites in the landscape (http://www.
giscafe.com/technical_papers/Papers/paper054/, accessed 15/07/2004).

288
 Predictive modelling of Roman settlement in the middle Tiber valley

Ingold T. 2000, The Perception of the Environment. Essays on Livelihood, Dwelling and
Skill, London-New York, Routledge.
Kamermans H. 2000, Land evaluation as predictive modelling: a deductive approach, in
Lock 2000, 124-146.
Kay S., Witcher R.E. 2005, The Tiber Valley Project. The role of GIS and databases in �eld
survey data integration and analysis, «Archeologia e Calcolatori», 16, 113-137.
Lock G. (ed.) 2000, Beyond the Map. Archaeology and Spatial Technologies, Amsterdam,
IOS Press.
Lock G. 2003, Using Computers in Archaeology: Towards Virtual Pasts, London,
Routledge.
Lock G., Stančič Z. (eds.) 1995, Archaeology and Geographical Information Systems: A
European Perspective, London, Taylor and Francis.
Mehrer M.W., Wescott C.L. (eds.) 2006, GIS and Archaeological Site Location Modeling,
London, Taylor and Francis.
Patterson H. (ed.) 2004, Bridging the Tiber. Approaches to Regional Archaeology in the
Middle Tiber Valley, Archaeological Monographs of The British School at Rome 13,
London, The British School at Rome.
Patterson H., Bousquet A., Di Giuseppe H., Felici F., Fontana S., Witcher R., Zampini
S. 2004a, Le produzioni ceramiche nella media Valle del Tevere tra l’età repubblicana
e tardo antica, «Rei Cretariae Romanae Fautores Acta», 38, 161-170.
Patterson H., Di Giuseppe H., Witcher R.E. 2004b, Three south Etrurian “crises”: �rst
results of the Tiber Valley Project, «Papers of the British School at Rome», 72, 1-36.
Patterson H., Di Giuseppe H., Witcher R.E. 2004c, Progetto Tiber Valley. La Sabina e le
due rive a confronto, in A. Guidi, S. Ponchia (eds.), Ricerche archeologiche in Italia
e in Siria, Padova, Sargon, 67-88.
Rua H. 2009, Geographic Information Systems in archaeological analysis: a predictive model
in the detection of rural Roman villae, «Journal of Archaeological Science», 36, 2,
224-235.
Sawatzky D.L., Raines G.L., Bonham-Carter G.F., Looney C.G. 2004, ArcSDM2: ArcMAP
extension for spatial data modelling using Weights of Evidence, logistic regression, fuzzy
logic and neural network analysis (http://www.ige.unicamp.br/sdm).
Stančič Z., Veljanovski T. 2000, Understanding Roman settlement patterns through multi-
variate statistics and predictive modelling, in Lock 2000, 147-156.
Terrenato N., Ammerman A.J. 1996, Visibility and site recovery in the Cecina Valley Survey,
Italy, «Journal of Field Archaeology», 23, 91-109.
van Hove D. 2004, Time and experience: taskscapes within GIS, «Internet Archaeology», 16
(http://intarch.ac.uk/journal/issue16/vanhove_index.html, accessed 06/09/ 2007).
van Leusen P.M. et al. 2002, Understanding digital archaeological landscapes, in P. Attema,
G.J. Burgers, E. van Joolen, M. van Leusen, B. Mater (eds.), New Developments in
Italian Landscape Archaeology, BAR International Series 1091, Oxford, Archaeopress,
125-152.
van Leusen M., Kamermans H. (eds.) 2005, Predictive Modelling for Archaeological Herit-
age Management: A Research Agenda, Nederlandse Archeologische Rapporten 29,
Amersfoot, National Service for Archaeological Heritage.
Verga F. 2007, Ager Foronovanus I, Forma Italiae, vol. 44, Firenze, Leo S. Olschki Editore.
Verhagen Ph. (ed.) 2007, Case Studies in Archaeological Predictive Modelling, Archaeological
Studies Leiden University 14, Amsterdam, Leiden University Press.
Wescott K.L., Brandon R.J. 2000, Practical Applications of GIS for Archaeologists: A Pre-
dictive Modelling Kit, London, Taylor and Francis.
Wheatley D. 2004, Making space for an archaeology of place, «Internet Archaeology», 15
(http://intarch.ac.uk/journal/issue15/wheatley_index.html, accessed 15/07/2004).

289
S.J. Kay, R.E. Witcher

Wheatley D., Gillings M. 2002, Spatial Technology and Archaeology, London, Taylor and
Francis.
Witcher R.E. in press, Missing persons? Models of Mediterranean regional survey and ancient
populations, in A. Bowman, A. Wilson (eds.), Settlement, Urbanisation and Population,
Oxford Studies in the Roman Economy 2, Oxford, Oxford University Press.
Witcher R.E., Kay S.J. 2008, The Tiber Valley Project: the database and Geographical Infor-
mation System, in Coarelli, Patterson 2008, 417-423.
Wood J. 1996-2008, Landserf (http://www.landserf.org).

ABSTRACT

This paper discusses the results of an inductive predictive modelling experiment on

Roman settlement data from the middle Tiber valley, Italy. The study forms part of the British
School at Rome’s Tiber Valley Project, which since its inception in 1997 has been assessing
the changing landscapes of the Tiber Valley from protohistory through to the medieval period.
The aim of this present study is to broaden understanding of settlement patterns via predictive
modelling, and in particular to evaluate unevenness in �eld survey coverage, survey bias and
past settlement location preferences. The predictive modelling method chosen was an applica-
tion of the statistical Weights of Evidence extension for ESRI ArcView. The results highlight
associations between Roman settlement and environmental themes that provide moderate
predictive potential and suggest that further experimentation might prove valuable.

290