Journal of Archaeological Science 38 (2011) 2644e2657

Contents lists available at ScienceDirect

Journal of Archaeological Science

journal homepage: http://www.elsevier.com/locate/jas

Computerized morphological classification of ceramics

Avshalom Karasik a, *, Uzy Smilansky a, b,1
a
The Computerized Archaeology Laboratory, The Hebrew University of Jerusalem, Mount Scopus, Jerusalem 91905, Israel
b
Department of Physics of Complex Systems, The Weizmann Institute of Science, Rehovot 76100, Israel

a r t i c l e i n f o a b s t r a c t

Article history: A novel morphological analysis of ceramic assemblages is presented using objective, automatic and
Received 27 September 2010 computerized method for clustering and classification. The analysis is based on the entire shape of the
Received in revised form potsherds, as expressed by three mathematical representations of their profiles. The similarity between
27 May 2011
profiles is mathematically defined in terms of the above mentioned representations. Cluster Analysis and
Accepted 28 May 2011
Discriminant Analysis are used to reveal a hierarchical classification of pottery assemblages. The method
is illustrated here by a detailed analysis of an assemblage which was classified previously by a pottery
Keywords:
expert. A quantitative comparison of the two typologies reveals the power and archaeological sense of
Pottery classification
Automatic typology
the new method.
Computerized morphological analysis Ó 2011 Elsevier Ltd. All rights reserved.
Cluster analysis

1. Introduction excavation reports. It forms the basis for further research, where
temporal sequencing and spatial correlations are extracted by
Pottery analysis is an essential and elementary task in any post- comparing data from different strata within the same excavation or
excavation examination and in the ensuing research. Usually, this from different sites. More often than not, the only accessible
involves the documentation and classification of hundreds or even information about the finds is what stored in the reports, and thus,
thousands of fragments. It consumes a lot of time, and delays the typology selected by the excavator cannot be tested or
significantly the analysis of the finds. Further research may also reassessed.
require the comparison of assemblages emerging from the same or The crucial step in the irreversible data reduction process
from different sites, to examine temporal or spatial relations. The described above is the determination of the types on which the
pottery fragments are classified by their form, clay composition, classification is based. This elementary fact was recognized as early
production methods, decorations etc. Here, we shall exclusively as modern archaeology established its modus operandi, and the
focus on the morphology which is a prominent attribute in any literature abounds with books and articles which discuss classifi-
attempt to classify potsherds. We also assume that the potsherds to cation and typology from a large variety of viewpoints. In spite of
be analyzed were produced on the wheel, and thus possess axial decades of research and study, there are methodological as well as
symmetry. In other words, their shape is entirely characterized by practical issues which are not universally accepted. (Even the
their sections (profiles). Moreover, they are assumed to be indica- terminology is not yet converged and concepts such as e.g.,
tive rim fragments. typology, clustering, sorting, taxonomy, categorization, classifica-
The traditional approach to pottery study proceeds along the tion etc. are used interchangeably). In the present article we do not
following line. The fragments are classified into one of a set of pre- intend to delve into these debates (the interested reader is referred
determined distinct types, and for each excavation unit, only few to the literature, for instance: Adams and Adams, 1991; Aldenderfer
representative fragments per type are drawn and published. The and Blashfield, 1978; Arnold, 1999; Binford, 1972; Brown, 1982; Hill
rest of the assemblage is sometimes ignored or, in the more and Evans, 1972; Joukowsky, 1980; O’Brien and Lyman, 1999; Orton
fortunate cases, summarized as a numerical frequency table which et al., 1993; Read, 2007; Rice, 1987; Whallon and Brown, 1982).
shows the distribution of the types within the excavation units. Rather, we go directly to the subject of the present work, namely,
This condensed representation of the data is then published in the the introduction of a practical computer based method for sorting
pottery fragments according to their shapes. It organizes any given
* Corresponding author. Tel.: þ972 2 5882118; fax: þ972 2 5825548.
assemblage of potsherds in hierarchical clusters (Cluster Analysis),
E-mail address: [email protected] (A. Karasik). according to mathematically defined morphological similarity. This
1
Tel.: þ972 8 9343180; fax: þ972 8 9344109. data organization process will be referred to as classification or

0305-4403/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jas.2011.05.023
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2645

Fig. 1. Two profiles of a krater and a cooking pot (a), and their representative mathematical functions e radius (b), tangent (c) and curvature (d).

clustering. Segregated clusters are the computer generated types, change in a consistent way by emphasizing the relevant
and the hierarchical ordering determines the corresponding sub- features at each stage.
types. Thus, a typology is defined, and objects within each type - The classification scheme should be amenable to generaliza-
are similar enough, so that their shapes can be represented by tions, so that non morphological parameters (varnish, clay etc.)
a single template (prototype). In our analysis, the prototype can be could be incorporated.
generated as a proper average (mean) of the shapes of the frag- - Finally and most importantly, the resulting classification
ments belonging to the corresponding cluster. should make archaeological sense and pass the scrutiny of
It should be emphasized at the outset that our method is not expert ceramics analysts.
intended to replace the archaeological judgment or reasoning. It is
offered as a convenient research tool which provides a systematic The method proposed here satisfies most of the requirements
basis for further analysis and modifications, by making use of the above. So far, it was tested on a few assemblages which consisted
advantages of computer based technologies. between 300 to 3000 items. Such numbers are large enough to
Almost per definition, any computer based classification scheme reach statistically significant conclusions, using standard computer
is objective and reproducible. However, it will meet the demands of resources (PC or laptops). With more experience, and once the
modern archaeological research if it satisfies the following method gains the trust of a wider range of users, it will be made
requirements: available as a service to the entire community.
The paper is organized in the following way. In the next
- The procedure should be efficient and cost effective enough to section we discuss the essential question in any classification
analyze an assemblage of a few thousands items, using a stan- algorithm, namely e the definition of “similarity” or “distance”
dard PC. between profiles. Here we follow our previous publications
- It should enable hierarchical storage and retrieval of data, (Adan-Bayewitz et al., 2009; Gilboa et al., 2004; Karasik et al.,
which could be further incorporated into a large data-base 2005; Saragusti et al., 2005). We briefly review it here so as to
system. make this paper self contained, and to introduce the main
- Preliminary setting of prototypes or preliminary sorting by the concepts and notations. Next, we deal with the clustering method
investigators should be optional but not mandatory. and with the statistical checks for its consistency. As an ultimate
- The archaeologist should be able to use the algorithm inter- check, we compare the computer based classification with the
actively and inject his/her preference to modify the analysis classification proposed by Dr Ayelet Gilboa, the ceramics expert
parameters at all stages. who excavated and analyzed the assemblage. It consists of 358
- The classification should use the information stored in the Iron Age pottery from the Tel Dor excavation. All illustrations in
entire profile of the fragment and not only in discrete metrical the rest of the paper use this data as well.2 Finally, the method is
measures such as e.g., height, width, girth etc. critically discussed and some concluding remarks are offered in
- The procedure should be able to compare fragments of the last section.
different sizes.
- The classification should be hierarchical and flexible. Flexibility
should ensure that as the classification proceeds from coarser 2
We are grateful to Dr Gilboa and the Tel Dor expedition for their generous
to finer details, the criteria which define similarity should cooperation.
2646 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

2. The method artifacts (Gilboa et al., 2004). This is done by averaging one of the
representative functions (curvature, tangent or radius) along their
2.1. Mathematical representation of the profiles mutual sections and reconstructing its corresponding shape. This
simple procedure produces an average form, which does not exist
It is assumed that the pottery assemblage consists of indicative in the assemblage. Nevertheless, it is an authentic representation of
rim fragments of spherically symmetric vessels, whose shape is the common features of the vessels which belong to the group. The
completely defined by their profiles which are available for advantage of the mean profile is that it enables portraying the
further analysis as digitized images. They can be obtained from prototype of a certain typological group, and to underline its
scanned archaeological publications (Gilboa et al., 2004) or, as significant and characteristic features (see Fig. 11 and discussion
becoming more and more common, from 3D models of the thereafter).
objects (Karasik and Smilansky, 2008). Either way, the profiles are
assumed to be properly positioned, namely, that the axis of
symmetry of the fragments are correctly determined. The appli- 2.2. Comparing profiles e the quantitative definition of the distance
cability of our method relies on the data satisfying the above between profiles
requirements. This does not affect its applicability too severely
because the indicative rim fragments form the bulk of profiles Given the representation of the ceramic profiles in terms of
displayed in reports, and their proper positioning is determined the functions RðsÞ, qðsÞ and kðsÞ defined above, we must address
either by experienced archaeologists or by the algorithm which now the crucial question of how to define similarity or dissimi-
larity between any pair of profiles in a quantitative way. In other
deduced the profile from the 3D models. As was already
mentioned, our method is based on treating the cross-section words, we must define a “distance” between profiles. Mathe-
matically speaking, there are many legitimate ways for the defi-
profiles as entire curves rather than as objects specified by
nition of a distance, the choice is dictated by its relevance to the
selected discrete measures.
problem at hand, and to other aspects such as the ease of its
The points on the profile are specified by their distance along
implementation. For the sake of the simplicity of the exposition,
the profile from a pre-determined reference point (the arc-length)
let us assume for the time being that we use only RðsÞ to describe
which will be denoted by s throughout. For the sake of simplicity
the profile and we shall explain the definition of the distance
and to enable straight forward comparisons between different
between profiles using this function. The inclusion of the other
profiles (see below), we set the highest rim-point (s ¼ 0) as the
two functions will follow.
reference point on the curve. We assign negative/positive values of
The definition of the rim maximum as the reference point (s ¼ 0)
s to the points on the exterior/interior of the fragment, respectively
enables us to position the compared functions without any ambi-
(Fig. 1). It is at this initial point of the analysis that the requirement
guity (See Fig. 1b). However, since the potsherds are usually broken
of proper positioning of the profile is essential.
at different points, the corresponding R(s) functions are defined in
There are several ways to describe the profile as a function of s.
different intervals on the s axis. To compare the functions we must
As in previous publications (Adan-Bayewitz et al., 2009; Saragusti
truncate the domain of s to the largest interval in which the two
et al., 2005), we use three different mathematical representations
functions are defined, and denote its end-points by ½Smin ; Smax  as
to describe the morphology of the curves: the radius RðsÞ, the
shown in Fig. 1b. Note that the interval ½Smin ; Smax  as defined above
tangent qðsÞ and the curvature kðsÞ. Mathematically speaking, each
of the three representations of the profile stores the entire depends on the pair of fragments which are compared. An alter-
morphological information of the curve (but for shifts). They are in native choice would be to fix the comparison interval for the entire
one to one correspondence, and each can be fully reconstructed assemblage, in which case it is determined by the shortest frag-
from the other without any loss of information (Adan-Bayewitz ment. The two options have their advantages and disadvantages,
et al., 2009; Saragusti et al., 2005). They differ, however, in the and the computer code allows the user to choose one or the other.
sort of features to which they are most sensitive. The three func- In the following illustration we preferred to use the pair specific
tions for two rather similar profiles of early Iron Age krater and determination of the comparison interval.
cooking pot (Fig. 1a) are shown in Fig. 1bed (courtesy of the Tel Dor Given two profiles (denoted by a, b), we define their distance by
excavations, Israel). The feature which distinguishes between the vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
two radius functions is the vertical shift between the curves. It u Z Smax
h i2
u1
reflects the size difference between the two vessels. At the same dR ða; bÞ ¼ u
tL Ra ðsÞ  Rb ðsÞ UðsÞ ds
time, finer differences can hardly be detected. For instance, both the Smin
krater and the cooking pot have a thickened rim at the exterior,
which are represented by deep valleys in the tangent function where UðsÞ stands for a (positive) weight which allows to empha-
around s ¼ 2. The depth of these valleys and their exact position size different parts of the profiles. This additional freedom becomes
relatively to the rim-point, show different patterns that are very handy when the archaeologist decides e.g., that the rim shape
emphasized by the tangent function. Finer details can be detected is more significant than the base shape. In this case UðsÞ will be
with the curvature function (Fig. 1d). The most conspicuous larger near s ¼ 0. In the examples discussed here we chose
elements in this function are the peaks which correspond to the rim UðsÞ ¼ 1, which endows the same weight to all parts of the profile.
area of the cooking pot. The positive peak matches the sharpened Finally, L stands for the total (weighted) arc-length interval
R Smax
tip and the negative one corresponds to the indentation below the L ¼ Smin UðsÞds. The subscript R indicates that the distance is
exterior rim. This graded resolving power enables us to choose the defined with respect to the radius function. The intuitive meaning
scale of details the classification should emphasize. The represen- of the definition of the distance is quite clear: it is the root mean
tative functions of the original profile highlights the sensitivity of square deviation between the two radius functions. It is clear that
our method to detect and quantify differences between profiles at the distance between profiles vanishes if and only if they are
various scales. This ability is crucial for successful typological identical.
classification as will be discussed below. In a similar way we can define the distance between any two
Another great advantage of the numerical representations of the profile in terms of the two other representing functions qðsÞ
profiles is the ability to compute and illustrate the mean of several and kðsÞ:
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2647

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u hundred items. However, a large amount of the information stored
u Z Smax
h i2
u1 in the matrix may be redundant: if the profiles can be partitions to
dq ða; bÞ ¼ u
tL qa ðsÞ  qb ðsÞ UðsÞds ; similarity groups, the distances between profiles within the same
Smin group are very small, and the distance between members of a group
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u and a profile from any other group are all the same. Thus, clustering
u Z Smax
h i2
u1 may reduce the effective dimension (rank) of the matrix to
dk ða; bÞ ¼ u
tL ka ðsÞ  kb ðsÞ UðsÞds a number which is much smaller than the number of objects N. The
Smin following example will clarify this point. Imagine an assemblage
with N identical profiles, for which the distance matrix is all zeros.
Each of the definitions of a distance expresses different aspect of The rank of this matrix is zero and there is no relevant typological
similarity between the two vessels and can be used independently variability in it. A less trivial example is an assemblage with only 2
in morphological studies (Adan-Bayewitz et al., 2009; Gilboa et al., different shapes ða; bÞ, each represented by Na ; Nb identical indi-
2004; Karasik et al., 2005). However, in order to establish the viduals respectively. The matrix is of dimension ðNa þ Nb Þ and its
analysis on more than one representative function without bias elements are either zero or dða; bÞ. The rank of this matrix is two.
(considering that the three functions have different ranges and Thus, the rank of the distance matrix is directly related to the
units), a normalization factor has to be introduced so that the variability of the assemblage under study. The best way to avoid
different distances are considered on the same footing. For an redundancies and to focus on the relevant variables which repre-
assemblage with M objects we have defined the ensemble average sent the true variability of the assemblage is to employ PCA
hRi; hqi and hki in the following way: (Jackson, 1991; Jolliffe, 2002). This method transfers the original
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi information from the distance matrix into a new matrix, with no
u Samax
X M u u1 Z
loss of information. The columns of the new matrix are linear
1 u
hRi ¼ ðRa ðsÞÞ2 UðsÞds ; combinations of the original ones. They are linearly independent
M a ¼ 1 tLa and with descending magnitude of variability, computed as the
Samin
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi mean of squared distances of each value from the average of the
u Samax column. Usually, only a few columns suffice to represent most of the
X M u u1 Z
1 u variability that exists in the assemblage. Fig. 2 shows the accu-
hqi ¼ ðqa ðsÞÞ UðsÞds ;
2
M a ¼ 1 tLa mulated variability of the columns (1e358) of the distance matrix
Samin from the test-case, before and after employing PCA. Before the PCA
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi manipulation (left figure), the various columns have, approxi-
u Samax
X M u u1 Z mately, the same scale of variability and each of them contributes
1 u
hki ¼ ðka ðsÞÞ2 UðsÞds
M a ¼ 1 tLa similar additions to the accumulated variability (as can be deduced
Samin from the linear behavior of the graph). On the other hand, after the
transformation (right plot), the contribution of successive columns
Thus, when all three representative functions are combined, the (principal components) to the accumulated variability is not
modified definition of the distance between the profiles simply uniform. The first 10 components already represent more than 90%
sums the normalized distances of the three functions: of the total variability within the assemblage, which is demon-
u u uk strated by the steep incline at about index 10. On the other hand,
dða; bÞ ¼ R
$d ða; bÞ þ q $dq ða; bÞ þ $dk ða; bÞ:
hRi R hqi hki the plateau starting at index 20 shows that the corresponding
components are irrelevant and represent redundant information.
Here, uR ; uq ; uk are the (non negative) relative weights The actual number of components was determined so that the
assigned to the corresponding functions, with uR þ uq þ uk ¼ 1. following three requirements are met: 1. The included PC represent
They are introduced in order to adjust the relative importance of more than 90% of the variability. 2. Each selected component
the various representations in the definition of “similarity”. The corresponds to at least 2% of the total variability. 3. All the
weights provide us with the freedom to emphasize finer details as components which contribute more than 5% are included, even if
we proceed with the hierarchically clustering of a complex the cumulative variability exceeds 90%. Thus, the final variability
assemblage. that we employ varies according to the nature of the assemblage.
The definition of dða; bÞ provides us with a single number The above practical rules provide an optimal compromise for all the
which quantifies the distance between any pair of profiles. The assemblages we analyzed. They can be altered when the data calls
dependence of the distance on the weighting function UðsÞ, and on for it.
the weights uR ; uq ; uk enable the archaeologist to introduce his/ To summarize: the PCA procedure associates to each fragment
her judgment and preferences in the clustering algorithm, and test a vector of numerical coefficients with a dimension which is much
them by comparing classifications which are carried out with smaller than N. Yet it fully represents the morphological variability
different choices of the parameters. It is exactly here that the within the ceramic assemblage. Each such vector is now defined as
dialog between the archaeologist and the computer comes to the similarity vector of PCA coefficients of the corresponding vessel.
bear in a prominent way. Given an assemblage, a convenient This vector is the most concise, but nevertheless complete,
framework for expressing all of its inner comparisons is in terms description of the relations of a given fragment to the rest of the
of a symmetric distance matrix, which will be discussed in details assemblage, and therefore it provides the most efficient and robust
in the following section. basis for the subsequent analysis.

2.3. Investigating the distance matrices 2.3.2. B. Cluster analysis (CA) and the definition of typology
As was demonstrated above, the vectors of PCA coefficients are
2.3.1. A. Principal component analysis (PCA) ideally designed for typological classification. In order to compre-
A distance matrix for an assemblage with N fragments is hensively reveal the complete inner structure of similarities and
symmetric and therefore the number of independent data is their clustering within the assemblage, the PCA vectors have to be
N  ðN  1Þ=2 which can be very large for any assemblage of a few further manipulated. The most common and useful method is
2648 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

35 35

30 30

25 25
accumulated variability

accumulated variability
20 20

15 15

10 10

5 5

0 0
0 100 200 300 0 100 200 300
culumn index of the distance matrix culumn index of the PCA matrix

Fig. 2. Accumulated variability of the columns of a distance matrix before (left) and after (right) employing PCA. The total sum of columns variability in the two plots is the same.

Cluster Analysis (CA) (Aldenderfer and Blashfield, 1984). It provides was already mentioned and will be demonstrated in the next
a hierarchical way to investigate grouping in the data, simulta- section, it is recommended to generate the typology using several
neously over various scales, by creating a cluster tree. The cluster refinement steps. The first step should differentiate between the
tree is a very convenient tool to explore and visualize correlations global shapes of the assemblage using high weights of the radius
in a data set (see Fig. 3). In this tree, similar objects are placed on and the tangent functions when generating the distance matrix. In
one branch (cluster) and each branch represents a segregated the next steps, each branch that was defined as a type, is analyzed
group. Clusters at one level of correlation are joined at the next separately with different weight parameters (emphasizing the
higher level. The height of each branching point is proportional to curvature over the tangent and the radius), in order to reveal finer
the distance between the two objects being connected. Every line at distinctions and sub-types. In this analysis each branch is treated as
the bottom of Fig. 3 (a ‘leaf’ in the tree) represents one sherd, and an independent assemblage and its PCA coefficients are computed
the various leaves are connected to a branch which are connected ab initio, based only on the distances between the members of the
to lower branches and so on, until the root of the tree (which branch. In this way, the PCA parameters which carry most influence
represent the entire assemblage). Note that by this convention, our on the next level of classification are updated according to the
trees grow with their leaves down and with their root up. relevant morphological features within each branch. The procedure
The algorithm which constructs the tree is based on the ends after a set number of refinement steps or when the groups are
distances between the corresponding vectors of parameters. Using smaller than a pre-determined number. In this way we assure that
the PCA coefficients for the Cluster Analysis forms the tree in the general shapes are separated according to parameters which are
most comprehensive and accurate way which reflects the most relevant to their differentiation. Nevertheless, their sub-types
morphological similarities. are based on finer details that are more appropriate and significant
The structure of the tree and its clusters serve for the definition for such a task. We use 60% of the level of the tree as a default
of the typology. Any horizontal line that is drawn at a given height height which sets the number of its sub-types. Establishing higher
intersects a finite number of branches. The number of intersection or lower thresholds will mainly affect the number of sub-divisions
points sets the number of types for the typology at a given level. needed to reach a certain resolution of classification, but it does not
The lower the line is drawn the more delicate the typology is and alter the shape of the tree.
vice versa, higher lines produce rougher classifications. The two The cluster tree defines various levels of classification based on
extreme cases are when each sherd is defined as a type, or when its hierarchic branches. The clustering status of each branch is
there is only one type which comprises the entire assemblage. summarized by a ‘type-code’. The number of refinement steps
There is no way to determine in advance at which level the line equals the number of digits in this ‘type-code’. Each digit stands for
should be drawn, and it depends on the nature of the assemblage, the index of the branch in the corresponding step. The ordering of
on the question under study and also on the research strategy of the the indices for the various branches is arbitrary. For instance, in the
investigator. The self-explanatory graphic representation of the tree analysis depicted in Fig. 3 the first refinement partitions the
helps to establish the decision. Nevertheless, this subjective deci- assemblage into three main groups. The corresponding type-codes
sion still produces objective results, since the chosen level will of these groups are 100, 200 and 300. In the second step, each of
always produce the same typology. This is a very important these branches was re-analyzed independently, using new set of
advantage over the traditional method of typological analysis, weights to produce their sub-branches (e.g., 110, 120, 210, 220, 230,
which is based on the human eye as the sole tool to evaluate a pre- 310, 320 and 330). The third subdivision creates a more refine
determined number of types and where each object belongs. As segregation as can be seen on the cluster tree. If, at a certain point,
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2649

200

100 300

220

230 210
221 222
230
110 120 310 330 320
213 330 322
321
214 211 212 110 122 315
121 123 313 311 314 312

Fig. 3. A typical cluster tree for 358 fragments from the early Iron Age at Tel Dor on which most of the type-codes are marked up to the third level.

a branch has already less leaves than the pre-determined level, its assigns. Repeating this process many times each object appears
digit at the current step remains zero (see for example type-codes approximately half of the times in the sample set, and acquires
330, 230 and 110 in Fig. 3). Fig. 3 shows the type-codes of a tree a corresponding number of labels. If the resulting labels distribute
which was constructed in three rounds that produces several types narrowly at a single group, for each object in the assemblage, the
and sub-types. Their correspondence with the archaeological underlying CA and the resulting typology is considered significant.
shapes will be discussed in the sequel. In practice, the DA analysis is considered significant if the label for
each of the objects is the same for more than 50% of the trials.
2.3.3. C. Discriminant analysis (DA) As an example, we apply this procedure to test the validity of the
The Discriminant Analysis (DA) method enables us to assert the clustering provided by the ‘type-codes’ 100, 200 and 300, shown in
statistical significance of the typology offered by the PCA and the Fig. 3. The DA algorithm was run 1000 times. In each run the
subsequent cluster analysis. The DA method can be briefly computer randomly chose half of the assemblage (179 fragments),
described in the following way (Klecka, 1982). Given an assemblage and built a discriminant function according to their corresponding
of N objects, each specified by a vector of P parameters, and suppose PCA coefficients. Then, the function was applied on the remaining
that the objects were clustered into G groups (with G, P << N) in half of the objects and the classification labels were saved for each
a way which reflects the relationships within the parameters run. The final product of this procedure is the distribution of the
specifying the objects. The DA algorithm assigns a number to each labels for every fragment. A distinct dichotomy between the type-
object which is a linear function of the P parameters. This function codes can be demonstrated when any single fragment is classified
is defined in such a way that it provides the best distinction back to its original group in most of the trials. On the other hand,
between the G pre-determined groups. This means that the ranges a substantial mixture between the classifications of the DA algo-
of the DA results for the various G clusters would have maximum rithm would indicate that the grouping is not significant (see for
separation. The separated ranges provide a DA label to each object instance Fig. 7). It is important to note that the DA algorithm uses
which can be compared with the original assignment. If the DA the PCA coefficients derived for one set of weights uR ; uq ; uk , with
labeling is statistically consistent with the original classification, no refinement steps. However, the final ‘type-codes’ represent
one can conclude that the classification is significant and reflects complicated grouping in which the weights were modified several
the true nature of the assemblage. Otherwise, either the clustering times. Therefore, in order to test the soundness of the CA ‘type-
is faulty or the set of parameters does not provide sufficient reso- codes’ the DA algorithm should be applied on each refinement step
lution to warrant the typology. Standard applications of the DA in separately. The groups and sub-groups of each digit in the ‘type-
the archaeological context could use any set of numerical param- code’ should be tested using the same weights which were used for
eters to specify the objects, and the groups may correspond to the the corresponding refinement step. Fig. 4 summarizes the results
traditional Archaeological types such as jars, bowls, kraters, etc. for the three main ‘type-codes’ e 100, 200 and 300. The number of
In the present work the objects are specified by their PCA times a profile was classified at any type-code is displayed as
coefficients, and the CA provides the segregation of the objects into a black bar for every fragment. If a fragment was always classified to
groups, by associating all the objects sharing the same type-code the same group, then a single bar is plotted and it covers the full
with a single group. To test if the type-codes cluster the object in height of the corresponding type-code. When a fragment was
a significant way, we take advantage of the fact that the number of classified to several type-codes, then a corresponding number of
objects, N, is much larger than the number of type-codes, so that bars are displayed and their heights are proportional to the clas-
each group consists of many objects. We partition the N objects sifications percentages. The pattern of the black bars demonstrates
randomly into two sets each consisting of N/2 objects. Then, we that most of the fragments were recognized properly. There are
apply the DA analysis to one of the sets (the training set) to obtain only two fragments which were not identified correctly both
the DA labeling function, and use it to label the objects in the other belong to ‘type-code’ 200. Their black bars ratio outside of the
set of N/2 objects (the sample set) e thus, every object in the sample expected rubric is substantial. Their corresponding drawings are
set acquires a label which this particular DA labeling function also illustrated in the figure.
2650 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

0 10 20 cm

0 10 20 cm

Fig. 4. DA results for the main three type-codes. The black bars in each rubric represent the ratio of classifications to the corresponding type-code. The arrows point to the drawings
which correspond to the two outliers bars.

As will be demonstrated in the next section, the lower vessel fits constructing a new branch, while the DA looks for maximum
very well also into ‘type-code’ 100, as is the case in about 30% of the difference between pre-determined groups. Second, as was
times of the DA runs. The upper drawing is a very unusual shape in the mentioned before, the vector of the PCA coefficients depends on the
assemblage, its vertical neck is similar to many of the fragments that given assemblage. The same vessel in a different assemblage might
belong to group 300 but its radius is larger and fits better to group have totally different PCA vector. The DA analysis in each run is based
200. The DA results testify that this vessel is an outlier in its group. on only half of the assemblage. Therefore, the DA results of several
Fig. 5 shows the DA results for the second digit of the ‘type-codes’. objects show some differences from their original CA classification.
Here, each of the three main groups was tested separately and again, These objects are less stable in their type-codes and would be the
most of the fragments were recognized correctly. However, in 14 firsts to move when the conditions are changed (when more data is
cases the highest black bar is outside of the expected rubric which added or when using different weights).
implies that each of these profiles, which were classified into Fig. 6 shows the DA results for the third digit of the ‘type-codes’.
a certain group by CA, proved to be better fitting in another group Again, each of the five ‘type-codes’, for which the last digit is not
with discriminant analysis. The discrepancy between the CA and the zero, was checked for the significance of its sub-divisions. It is
DA can be explained as follow: First, the two methods are based on important to note that some of the ‘type-codes’ groups have only
different principles. The CA searches for maximum similarity when a few members (for instance 211 N ¼ 6, 312 N ¼ 2 and 315 N ¼ 5).

330

320

310

310 320 330

230
DA labeling

220

210

210 220 230

120

110

110 120
Original type-codes

Fig. 5. DA results for the second level of type-codes. The black bars in each rubric represent the ratio of classifications to the corresponding type-code.
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2651

Fig. 6. DA results for the third level of type-codes for which the third digit is not zero. The black bars in each rubric represent the ratio of classifications to the corresponding
type-code.

This increases the statistical uncertainty of the DA procedure and To summarize, the investigation of the distance matrix starts
contributes to the global error rate. Nevertheless, most of the with PCA which generates the most relevant parameters for the
fragments were recognized correctly and only in several cases the analysis. It continues with a sophisticated version of cluster analysis
highest black bar is outside of the expected rubric. that uses a refinement procedure in order to relate between the
An agreement between these two independent methods, as is level of the cluster and its most significant characteristics. The
the case for most of the assemblage here, strengthens the classifi- output is a self-explanatory cluster tree, from which the complete
cation results and enables to define the ‘core’ of each type-code in relations and similarities within the assemblage are graphically
comparison to its outliers. illustrated. The typology is generated by hierarchical type-codes
Perhaps the best way to appreciate the advantage of the DA which correspond to the various branches of the tree. The validity
results is by repeating the analysis using a less discriminating and the significance of the type-codes are tested using DA which
parameter set to describe the objects. Fig. 7 shows the DA results for also detects the outliers in each group.
the same groups as in Fig. 5 but this time the PCA algorithm uses
only the curvature function. Although most of the fragments were 3. Test-case
classified correctly to their original type-codes, the error rate is
high (42%). The curvature function is very powerful in detecting Although the classification method which was described above
minute changes between similar objects but it can also wrongly seems consistent and provides a reliable typology, it is very hard to
associate shapes with different sizes and orientation. This test assess its archaeological relevance. At the end of the day, the groups
emphasizes once again why it is important to use all three repre- and the sub-groups which are defined by the algorithm must be
sentative functions in the classification procedure. examined by the archaeologists, especially because the weights

330

320

310

230
DA labeling

220

210

120

110

110 120 210 220 230 310 320 330

Original type-codes

Fig. 7. DA results similar to Fig. 5 but when the algorithm used only the distances of the curvature function as input.
2652 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

70

Bowls
Kraters
60 Cooking pots
Jugs
Jars

Number of vessels 50

40

30

20

10

0
110 121 122 123 211 212 213 214 221 222 230 311 312 313 314 315 321 322 330
Type-Code

Fig. 8. Distribution of Gilboa’s family types according to the type-codes of the automatic typology.

parameters are subject to their judgments: there is no a-priori pot and Bowl, with several further definitions for sub-types but
unique and optimal settings for the weights, and one should always with no strict typological boundaries (Gilboa, 2001). Fig. 3 shows
test and adjust them to match the idiosyncratic features of the the final cluster tree and type-codes for this assemblage after
assemblage at hand. In the present section we shall discuss the applying the following weights at the three refinement steps (given
analysis of the assemblage mentioned previously, and show, among in percents for uR ; uq ; uk respectively):
other things, how different weights may affect the final results. The Step 1: [40, 50, 10]; Step 2: [25, 50, 25]; Step 3: [0, 45, 55].
main reason for choosing the Tel Dor assemblage to demonstrate The chosen weights emphasize the radius over the curvature in
our method is that it was previously analyzed by Dr Ayelet Gilboa e the first step and vice versa in the third iteration. We found the
an expert archaeologist. This way we could compare classifications tangent function very indicative for typological analysis and
based on the computerized procedure and the archaeological therefore its weight is always w50%.
instincts and tradition, and to elaborate on the differences between Fig. 8 displays the distribution of Gilboa’s types within the
the two independent techniques. various type-codes. Most of Gilboa’s bowls (the open most shape),
The benchmark assemblage consists of 358 fragments from the together with only one krater belong to type-code (100). The semi-
early Iron Age at Tel Dor, Israel (Gilboa, 2001). According to Gilboa, closed vessels e the kraters and the cooking pots e are the main
the assemblage comprises of 5 vessel types: Jar, Jug, Krater, Cooking component in the middle group of type-codes (200). Similarly, the

34 35 36 37 38

Type-Code 123 29 30 31 32 33

24 25 26 27 28

Type-Code 122
19 20 21 22 23

14 15 16 17 18

Type-Code 121 9 10 11 12 13

5 6 7 8

1 2 3 4

Type-Code 110

0 10 20 cm

Fig. 9. Several examples of drawings for each of the sub-types in type-code 100.
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2653

18 19 20 21 22 23

Type-Code 214

12 13 14 15 16 17

Type-Code 213
7 8 9 10 11

Type-Code 212
1 2 3 4 5 6

Type-Code 211

0 10 20 cm

Fig. 10. Several examples of drawings for each of the sub-types in type-code 210.

third group (300) is consistent only with closed containers e the While the first two refinements classified the assemblage
jars and the jugs. In the sub-divisions resulting from the second consistently with Gilboa’s types, it is mainly the third step which
refinement, one finds most of the cooking pots in type-codes generates further refinements. For instance, type-code (110)
(211)e(214), while the majority of the kraters accumulated in the comprises mainly larger bowls with thickened rim (Fig. 9: 1e8). The
type-code (220). The remaining cooking pots and kraters are rest of the bowls (often called ‘boring bowls’ at Tel Dor) are divided
associated with type-code (230). Likewise, type-code (330) into three sub-types: (121) e has no rim treatment; (122) e with
contains almost only jugs, while there is just one jug in (320) along elaborated rim treatment; and (123) e the nonstandard examples
with 38 jars. On the other hand, type-code (310) consists of both (Fig. 9). The first two also correlate almost exclusively with Gilboa’s
jugs and jars. This is consistent with Gilboa’s typological assertion sub-types definitions (BL 33 or 20e26).
that “. it is possible that part of the rims classified as jugs (GR 6) The sub-divisions of type-code (210) show minor differences
actually belong to such jars (SJ 1)” (Gilboa, 2001). between the cooking pots. The rims in this category are thickened
It is important to note, again, that changing of the default level at on the outside, and they are distinguished by details, such as the
which we cut the tree into branches will mainly affect the coding inclination of the lip or its relative thickness. Fig. 10 shows that
system but the shape of the tree and the final typology will remain type-code (211) displays the classic rim shape of Iron Age cooking
almost the same. For instance, lowering this level down to 33% will pots, where the inclination of the upper part is almost vertical,
define five main branches at the first step (instead of 3). This may fit unlike for instance type-code (212) where it is leaning inwards.
better to Gilboa’s 5 types and will reduce the number of digits to two, A glance at type-code (213) shows larger variability of shapes.
but it will not change much the final groups. On the other hand, by However, their common feature is evident from the analysis of the
keeping the level at 60% one can see more clearly the morphological corresponding curvature functions, demonstrated by the mean
similarity between kraters and cooking pots as semi-closed vessels, curvature functions in Fig. 11. A crucial issue for a good agreement
and between jugs and jars. We found this level of 60% as a good between two curvature graphs is the position of the positive peaks
compromise between the two needs of reducing the number of digits and negative valleys. Shifted peaks and valleys contribute
and keeping the hierarchic nature of the tree in the ‘type-codes’. substantially to the distance between the corresponding profiles.

Type-code 211
Type-code 212
type-code 211 Type-code 213
type-code 212
0.1 Type-code 214

type-code 213 type-code 214

0 10 20 cm
0.05
k(s)

-0.05

-0.1
-50 -40 -30 -20 -10 0 10
Arc-length (mm)

Fig. 11. The mean curvature functions of the four sub-types in type-code 210. The reconstructed shapes out of the mean representations are plotted at the top left corner.
2654 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

14 15 16

cooking pots
10 11 12 13

7 8 9

4 5 6

1 2 3

kraters

0 10 20 cm

Fig. 12. The two groups within type-code 230, kraters (1e9) and cooking pots (10e16).

Apparently, the relative locations of those peaks along the arc- unusual shapes of kraters also according to Gilboa (compare to
length are the main characteristic feature of type-code (210). other kraters in Fig. 13). The rims here are much thicker and pro-
Fig. 11 shows the mean curvature of the type-codes (211)e(214). jected to the exterior part of the vessel in a way which is more
Their corresponding shifted peaks and valleys are the main influ- similar to the common cooking pot than the common krater at Dor.
ence for the sub-grouping (particularly because uk was set to 55%, It is probably Gilboa’s rich experience and her knowledge of the
in the third step). This also explains why there are two different function of these vessels which led her to call them kraters.
kraters (Fig. 10,13,14) within type-code (213). Their very thick rims Morphologically speaking, these kraters have different parameters
follow a similar pattern of peaks and valleys as with the curvature than the rest of the kraters in the assemblage. The automatic
function of the rest of the group. classification pointed at this phenomenon but the archaeologist is
Type-code (230) is, according to Gilboa’s typology, a mixture of certainly free to modify the typology by taking into consideration
cooking pots and kraters (Fig. 12). However, it comprises the other features such as function, color, clay etc.

25 26 27 28 29

Type-Code 222
20 21 22 23 24

16 17 18 19

11 12 13 14 15

6 7 8 9 10

1 2 3 4 5

Type-Code 221

0 10 20 cm

Fig. 13. Several examples of drawings for each of the sub-types in type-code 220.
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2655

30 31 32 33 34 35 36 37 38 39

Type-Code 322
Type-Code 330
Type-Code 321
23 24 25 26 27 28 29

11 12 13 14 15 16 17 18 19 20 21 22

Type-Code 314 Type-Code 315

1 2 3 4 5 6 7 8 9 10

Type-Code 311 Type-Code 312 Type-Code 313

0 10 20 cm

Fig. 14. Several examples of drawings for each of the sub-types in type-code 300.

Most of the kraters in the assemblage were assigned to type- A very interesting and important issue regarding computerized
code (220), and the high accumulation in type-code (222) can be classification is the influence of the weights parameters uR ; uq ; uk
explained by their great homogeneity. They differ, however, in the on the final typology. We have tested many different weights
finer details like the inclination of the wall below the thickened rim. setting since a-priory the choice of the weights is arbitrary.
The inward leaning angle of type-code (222) is larger, and there are Nevertheless, several typologies, in which we did not introduce
some rims with exterior overhang in type-code (221) (Fig. 13). extreme parameters, revealed similar clusters of fragments.
Nevertheless, one should always remember that the two groups are Though, there were always differences and some vessels mean-
rather close in shape. dered between the groups. Obviously, the corresponding type-
The third main branch (300) comprises exclusively of the closed codes and the cluster trees are also different from one set of
vessels in the assemblage e jars and jugs. Considering the discus- weights to another. Fig. 15 shows the distribution of Gilboa’s types
sion above regarding the lack of confidence in the dichotomy according to the type-codes obtained with equal weights for all
between the two, there is no point to try and follow Gilboa’s defi- three representative functions in all three refinement steps.
nitions. Nevertheless, type-codes 320 and 330 seem to maintain Generally, the first division (1. bowls; 2 kraters and cooking pots; 3
Gilboa’s original division into jars and jugs respectively. As can be jars and jugs) remain the same. The main difference is the greater
seen in Fig. 14, the main differences between the groups are details mixture between the kraters and the cooking pots in the sub-types
like the length of the neck, its radius and orientation, the thickness of the second group. This can be explained by the reduced weight of
of the rim and the convexity of the interior neck. the curvature function in the third refinement. This reduction
The results show fine grouping which corresponds quite well to causes the small details of the rim, which characterize each type to
the archaeologist’s classification and in several cases even improve it. exert less weight on the final typology.

45
Bowls
40 Kraters
Cooking pots
Jugs
35 Jars

30
Number of vessels

25

20

15

10

0
110 121 122 123 211 212 220 231 232 233 311 312 313 314 321 322 323 324 330
Type-Code

Fig. 15. Distribution of Gilboa’s family types according to the type-codes of the automatic typology when using equal weights for all three functions e radius, tangent and curvature.
2656 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657

To test the statistical significance of the typology we used importance in ceramic studies, other parameters such as e.g.,
Discriminant Analysis (DA). Fig. 4 demonstrated that most of the varnish, chemical composition and provenience provide important
assemblage was recognized correctly and classified back to its typological clues. We plan to incorporate this information in our
principal ‘type-code’ ((100), (200) and (300)). The average height of algorithm.
wrong classifications bars can serve as an indication for the quality So far we did not discuss the question of adding more data or
of the clustering type-codes. After the second refinement this value updating existing drawing items to an assemblage. The straight
was 5.2%, implying that on average 94.8% of the times each frag- forward approach would be to re-analyze the extended assem-
ment was classified correctly to its genuine type-code (Fig. 5). As blage. This may not be critical for small or medium size collections
a comparison, the same test was applied to Gilboa’s typology of 5 [the analysis of an assemblage of 3000 fragments on a dual-core PC
main types e bowls, kraters, cooking pots, jugs and jars. This time, took approx. 15 min]. However, the main obstacle in dealing with
the wrong classifications mean ratio was 9.8%, almost double than large assemblages is not the computing time, but the memory
before, although there are fewer groups. needed for the analysis of large matrices. The computational
As a final comment we would like to stress that we have applied resources required for the classification grow quadratically with the
the same method on several large assemblages (up to 3000 frag- number of items. For a standard PC the critical number of items is of
ments) with similar rate of success and under stringent scrutiny of order 104. The problem can be overcome by taking advantage of the
the archaeological relevance (to be published separately). hierarchical nature of the typology and the refinement procedure.
An initial typology can be defined using only part of the assemblage
4. Summary, critical discussion and future research which will be selected at random. In the next step, each vessel
directions which was not chosen is to be compared to the mean shapes of the
various type-codes (see examples for mean shapes in Fig. 11). After
This paper describes a novel method for computerized ceramic the most similar type-codes are selected, each of them could be
typology and classification. The method is based on a distance reclassified separately, to include the new data. Another solution
function which is defined in terms of three mathematical repre- may be to use other methods, such as machine learning and arti-
sentations of the profile e radius, tangent and curvature given as ficial intelligence to classify new data into the existing type-codes.
functions of the arc-length. The distance function is defined in such The successful classifications obtained so far encourage us to
a way that it takes the entire morphology of the profiles into further develop this method. After gaining more experience we
account with no bias. Nevertheless, it allows the archaeologists the shall create a user friendly version of the code which will be
freedom to adjust the parameters and weights of the three func- available to the archaeological community. Thus, the method is
tions in order to fit them to the assemblage under study and to the offered as objective tool that saves time, reveals significant details
goal of the research. Still, the various adjustments maintain the and helps substantially to the mandatory never-ending task of
objectivity and the reproducibility of the analysis. The internal pottery analysis. By that, we believe that it fulfills and reaches
relations within the assemblage are summarized in the distance further than the hope expressed by J.A. Brown almost 30 years ago:
matrix. Three steps are followed in the computerized typology
“These methods, which fall under my definition of automatic
recommended in the present work: The PCA provides the most
classification, hold out the hope that typological procedures can be
economical characterization of the correlations within the data. An
routinized to the point they can be reduced to an algorithm”
advanced Cluster Analysis which uses refinement steps constructs
(Brown, 1982: 183).
a hierarchical tree from which the types and sub-types can be
defined. Then, DA tests the significance of the resulting typological
classification. The advantages over the traditional method can be
Acknowledgment
summarized as follow:
We would like to express our gratitude to Ilan Sharon and Ayelet
1. Objectivity e All profiles are compared algorithmically, inde-
Gilboa for allowing us to use archaeological material from the Tel
pendently of any subjective judgment.
Dor excavation and for their interest and help. The discussion with
2. Data Completeness e Each of the three representative func-
Ayelet Gilboa on her classification was of extreme importance. We
tions stores the entire morphological information of the profile,
are grateful to Leore Grosman and Talia Goldman for their support,
unlike a discrete set of measurements.
interest and comments on the manuscript. Support from the ISF
3. Comparison Completeness e The distance function was defined
grant 168\06 is acknowledged with thanks. A.K. is indebted to the
so that the entire profile is taken into consideration in the
Nathan Rotenstreich Scholarship for Outstanding Ph.D. Candidates,
computation, albeit different parts of the profile can be
Israel Council for Higher Education, for their support.
weighted differently using the weight function UðsÞ.
4. Hierarchical typology e The refinement steps procedure
creates hierarchical classification which represents various References
scales of the relevant information.
Adams, W.Y., Adams, E.W., 1991. Archaeological Typology and Practical Reality: A
5. Adjustable refinement e The weights (uR ; uq ; uk ) can be Dialectical Approach to Artifact Classification and Sorting. Cambridge Univer-
modified along the steps of the typology, to control the effect of sity Press, Cambridge.
different shape representations. Adan-Bayewitz, D., Karasik, A., Smilansky, U., Asaro, F., Giauque, R.D., Lavidor, R.,
2009. Differentiation of ceramic chemical element composition and vessel
morphology at a pottery production center in Roman Galilee. Journal of
The method was demonstrated on an early Iron Age assemblage Archaeological Science 36, 2517e2530.
from Tel Dor, Israel, and its results were compared to its traditional Aldenderfer, M.S., Blashfield, R.K., 1978. Cluster analysis and archaeological classi-
fication. American Antiquity 43, 502e505.
classification. It has been shown that automatic classification can
Aldenderfer, M.S., Blashfield, R.K., 1984. Cluster Analysis. Sage University Paper.
follow the traditional archaeological concepts of typological clas- Arnold, P.J.I., 1999. On typologies, selection, and ethnoarchaeology in ceramic
sification and even improve its resolution. It does so without bias, in production studies. In: Chilton, E.S. (Ed.), Material Meanings e Critical
a fast, objective and reproducible way. Approaches to the Interpretation of Material Culture. The University of Utah
Press, Salt Lake City, pp. 103e117.
Our method classifies the assemblage based on morphological Binford, L.R., 1972. An Archaeological Perspective. Seminar Press, New-York and
information only. Even though the shape information is of prime London.
 A. Karasik, U. Smilansky / Journal of Archaeological Science 38 (2011) 2644e2657 2657

Brown, J.A., 1982. On the structure of artifact typologies. In: Whallon, R., Brown, J.A. Karasik, A., Smilansky, U., Beit-Arieh, I., 2005. New typological analyses of hole-
(Eds.), Essays on Archaeological Typology. Evanson. mouth jars from the early bronze age from Tel Arad and southern Sinai. Tel Aviv
Gilboa, A., 2001. Southern Phoenicia During Iron Age IIa in the Light of the Tel Dor 32, 20e31.
Excavations. The Institute of Archaeology: The Hebrew University, Jerusalem. Klecka, W.R., 1982. Discriminant Analysis. Sage University Paper.
Gilboa, A., Karasik, A., Sharon, I., Smilansky, U., 2004. Towards computerized O’Brien, M.J., Lyman, R.L., 1999. Seriation, Stratigraphy and Index Fossils: The
typology and classification of ceramics. Journal of Archaeological Science 31, Backbone of Archaeological Dating. Kluwer Academic/Plenum Publishers,
681e694. New-York.
Hill, J.N., Evans, R.K., 1972. A Model for Classification and Typology. In: Clarke, D.T. Orton, C., Tyres, P., Vince, A., 1993. Pottery in Archaeology. University Press,
(Ed.), Models in Archaeology. Methuen and Company, London, pp. 231e273. Cambridge.
Jackson, J.E., 1991. A User’s Guide to Principal Components. John Wiley and Sons. Read, D.W., 2007. Artifact Classification: A Conceptual and Methodological
Jolliffe, I.T., 2002. Principal Component Analysis, second ed. Springer Series in Approach. Left Coast Press, Walnut Creek, California.
Statistics. Rice, P.M., 1987. Pottery Analysis: A Sourcebook. University Press, Chicago.
Joukowsky, M., 1980. A Complete Manual of Field Archaeology. Prentice-Hall, Saragusti, I., Karasik, A., Sharon, I., Smilansky, U., 2005. Quantitative analysis of
Englewood Cliffs, N.J. shape attributes based on contours and section profiles in archaeological
Karasik, A., Smilansky, U., 2008. 3D scanning technology as a standard archaeo- research. Journal of Archaeological Science 32, 841e853.
logical tool for pottery analysis: practice and theory. Journal of Archaeological Whallon, R., Brown, J.A., 1982. Essays on Archaeological Typology. Center for
Science 35, 1148e1168. American Archaeology Press, Evanstone.