Journal of Applied Geophysics 178 (2020) 104069

Contents lists available at ScienceDirect

Journal of Applied Geophysics

journal homepage: www.elsevier.com/locate/jappgeo

Joint interpretation of seismic refraction tomography and electrical

resistivity tomography by cluster analysis to detect buried cavities
Alessandra Carollo ⁎, Patrizia Capizzi, Raffaele Martorana
Università degli Studi di Palermo, Dipartimento delle Scienze della Terra e del Mare (DiSTeM), Palermo, Italy

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

Article history: In the last few years, the geophysical methods of seismic refraction tomography (SRT) and electrical resistivity
Received 24 January 2020 tomography (ERT) are among the most used geophysical techniques for the reconstruction of subsoil geometries,
Received in revised form 12 May 2020 for the investigation of underground cavities and also for the archaeological prospecting. However, the main dis-
Accepted 13 May 2020
advantage of each geophysical method is the difficulty of final interpretation of the data. In order to eliminate ar-
Available online 16 May 2020
tifacts and generally improve the reliability and accuracy of geophysical interpretation, it is useful to perform a
Keywords:
joint approach of different geophysical methods, also introducing the a priori information. In this work, it is
SRT shown the integrated study of seismic refraction tomography and electrical resistivity tomography techniques,
ERT the two geophysical methods are tested on both synthetic and real data and the integration of data is useful in
Joint interpretation detecting buried cavities and also evaluate their geometric characteristics. Likelihood parameters has been de-
K-means cluster analysis fined and tested, in order to help recognizing voids from other lithological structures. Finally, a statistical ap-
Modeling proach based on cluster analysis of the P-wave velocity, the density of the seismic rays and the electrical
Cavity resistivity of the synthetic and experimental models was used. Multi-space cluster distribution maps were
built, allowing to better define and interpret the anomalies of the subsoil.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction different nature on independent parameters and to correlate them dur-

ing the interpretative process.
The presence of cavities buried (voids like tunnels, underground ca- It is also essential that the geophysical survey is designed and chosen
nals, underground mines, caves, tombs) in highly populated areas cre- on the basis of a priori information of the geological context, making it
ates serious problems in terms of hazard and risk for the stability of possible to obtain a rough estimate of depth and size of the voids to
the infrastructure and individual protection. It is therefore essential to evaluate any differences in the properties physico-mechanical proper-
identify these areas to minimize the geological risk and implement an ties of the materials involved (Cardarelli et al., 2010). Today electrical
optimal plan for the use of the soil. Several case studies, from around resistivity tomography (ERT) is the most widely method used for the
the world, show the effectiveness of geophysical methods in obtaining detection of underground voids (Capizzi et al., 2005) and also for ar-
information on the structures and geometries present in the subsoil chaeological research (Cardarelli et al., 2008).
and constitute an excellent tool for the detection of buried cavities However, this technique is less sensitive to electrical and electro-
(Imposa et al., 2018; Maraio and Bruno, 2015; Orfanos et al., 2008). magnetic disturbances when compared to other geophysical methods
The determination of dimension and depth of the cavities and of the (Van Schoor, 2002; McDonald and Davies, 2003; Cardarelli et al.,
contrast in physical properties between the cavity and surrounding 2008). Consequently, in recent years, the technique of electrical tomog-
media are fundamental aspects for understanding the validity of a par- raphy has been joined more and more often to the technique of seismic
ticular geophysical method (Fasani et al., 2013). Until today, different refraction tomography (SRT) in order to obtain more robust interpreta-
geophysical prospecting methods have been used for the detection of tions (Sheehan et al., 2005; Riddle et al., 2010). The use of combining
underground voids, and their success is strongly influenced by their electrical and seismic data derives from the observation that there is
non-invasiveness, the ability to pre-investigate and data resolution. an indirect correlation between the electrical resistivity and the velocity
The ambiguities of the inverse models suggest the integrated use of dif- of the seismic waves, through a direct correlation of both parameters
ferent geophysical methodologies in order to obtain information of with the porosity (Archie, 1942; Wyllie et al., 1956), caused by the influ-
ence on both these geophysical parameters by the structure of the pores
⁎ Corresponding author.
of the materials (Meju et al., 2003). This is the main assumption for a
E-mail addresses: [email protected] (A. Carollo), [email protected] joint inversion of seismic refraction and electrical resistivity tomogra-
(P. Capizzi), [email protected] (R. Martorana). phy. This can be very useful for the identification of cavities and tunnels

https://doi.org/10.1016/j.jappgeo.2020.104069
0926-9851/© 2020 Elsevier B.V. All rights reserved.
2 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

(Dahlin et al., 1999), because it allows to reduce the intrinsic ambigui- approach is tested on a site of limestone quarries, on which a speleolog-
ties of each method (Linder et al., 2010), at the same time improving ical survey was previously made.
the resolution and compatibility of the velocity and resistivity models
(Gallardo and Meju, 2004).
A correct management of a multiparametric dataset must take into 2. Materials and methods
account the difficulty of defining valid theoretical or at least empirical
relations between the parameters. These are in facts influenced by the 2.1. Choice of the synthetic models
great variability of the chemical and physical conditions within the
Earth (Lees and VanDecar, 1991). It is therefore necessary to allow a cer- Synthetic modeling means creating a simplified representative
tain degree of independence between each type of observed data, model of a subsoil portion. The most effective way to represent with a
modeling separately each geophysical parameter in the forward prob- good approximation the complex distributions of a physical parameter
lem. In this way it is possible to impose mathematical constraints that in the subsoil is to perform a discretization in homogeneous cells, each
guide the joint inversion towards models with similar spatial distribu- characterized by constant values of the physical parameters considered.
tions of anomalies (Gallardo and Meju, 2003, 2004, 2007, 2011). How- In this way the synthetic model is not a complete representation of the
ever, often, the parameters considered (for example, seismic velocity real model, since it is characterized and described by a finite number of
and electrical resistivity) have different distributions in the subsoil elements, but it is possible to decide the resolution degree of the model
and, in these cases, the conjugate inversions can lead to serious by varying the dimensions of the cells. By solving a forward problem, it
misinterpretations. is possible to calculate the predicted data according to the analyzed
An approach that is alternative to the joint inversion involves the use physical laws (elastic wave propagation, electric field, etc.). The forward
of post-inversion techniques for univariate independent models, in modeling allows to study potentialities and limits of the analyzed meth-
order to find relationships between the different observable parameters odology. Similarly with other authors (Cardarelli et al., 2010; Fasani
to identify different lithological structures. Following this approach, et al., 2013), in this study synthetic models that represented different
local empirical relationships between different parameters can be cavities were considered, parametrizing the velocity of the pressure
defined. seismic waves and the electrical resistivity. These synthetic models
Cluster analysis is a multivariate analysis technique in which statis- were used to test the properties, efficacy and limits of the joint use of
tical units can be combined using an optimization criterion, minimizing seismic refraction tomography (SRT) and electrical resistivity tomogra-
the parametric distance within each group and at the same time maxi- phy (ERT) for the detection of sub-surface cavities, using post-inversion
mizing it. The parametric distance is quantified by measures of similar- cluster analysis techniques. Moreover, a detailed synthetic modeling
ity and/or dissimilarity between defined statistical units. Thus, with the was also performed, based on a real situation in which high-resolution
cluster analysis method, it is possible to identify within a set of objects speleological surveys were available, in order to compare predicted
some subsets called clusters, which tend to be similar within them, and experimental results, to optimize the inversion and interpretation
based on their level of similarity. In this way, each cluster contains ele- phase. All the synthetic models of seismic refraction tomography were
ments that are homogeneous to each other and which have a high inter- created by Surfer® software (Golden Software), while the synthetic
nal coherence (ie, minimum intra-cluster distance) and high external data were calculated, processed and inverted with the Rayfract® soft-
heterogeneity (ie, maximum intercluster distance) (Barbarito, 1999). ware (Intelligent Resource, Inc.). The resistivity models were created
In general, clustering is an excellent tool whenever you want to catalog with the RES2DMOD® software and inverted with the RES2DINV®
a certain amount of information in meaningful and manageable groups. software.
Over the years, clustering methods have been applied to a wide variety 2D synthetic models, synthetic models were created with different
of research fields, the scholar Hartigan (1975) provides a comprehen- number of cavity and blocks of highly cohesive lithological material
sive list of the numerous published studies reporting the results of var- (high seismic velocity and resistivity values). The electrical resistivity
ious cluster analysis applications. models were constructed using a value of 10,000 Ωm for cavities and
Several authors used the cluster analysis between the seismic veloc- 500 Ωm for the background material. While, refraction seismic models
ity and the electrical resistivity distribution (i.e. Dell'Aversana, 2001; were constructed (Rayfract, 2010a; Rayfract, 2010b)using a value of
Gallardo and Meju, 2003; Bottari et al., 2018a). K-means cluster analysis 330 m/s for cavities and 5000 m/s to simulate the presence of a highly
has often been applied in order to correlate different physical properties cohesive material (box size 6 m × 6 m). The background of the seismic
between them, for example seismic velocity and electrical resistivity models has been divided into layers to create a vertical velocity gradi-
(Bottari et al., 2018b), velocity and attenuation of the electromagnetic ent, with vp values between 400 m/s near the surface up to 3000 m/s
waves in georadar data (Tronicke et al., 2004), georadar attenuation in depth. The layers have variable thicknesses: for the first four layers
and magnetic field intensity (Scudero et al., 2018), datasets of seismic near surface the thickness is 2 m, the underlying layer instead has a
refraction tomography (SRT) and controlled-source audio thickness of 7 m (bottom at depth of 15 m) and the last layer reaches
magnetotelluric (CSAMT) (Di Giuseppe et al., 2014). the depth of 30 m coinciding with lower boundary of the model. The dif-
The use of post-inversion techniques applied to ERT and SRT data for ference between seismic and electrical background models derives from
the investigation of cavities has been successfully proposed using differ- the observation that stone rocks often show surface layers of alterations
ent types of clustering approaches, in order to facilitate the phase of in- that cause a decrease in seismic velocity that does not correspond to a
terpretation by providing a clearer imaging of the representation of the measurable variation in their high resistivity values.
subsoil (Meju et al., 2003; Gallardo and Meju, 2004; Orfanos and The synthetic seismic coverage and refraction tomographies were
Apostolopoulos, 2013; Di Giuseppe et al., 2014; Kotyrba and Schmidt, created, processed and inverted with Rayfract® software (Intelligent
2014; Hellman et al., 2017). Resource, Inc.), while the synthetic resistivity models have been created
In this paper the application of an algorithm based on non- and inverted by the RES2DMOD® and the RES2DINV® software,
hierarchical cluster analysis (k-means) is proposed, together with the respectively.
use of likelihood parameters, for the joint interpretation of SRT and The synthetic models (Fig. 1) were created starting from the same
ERT data, acquired with identical sensor array (electrodes and geo- background subsoil and inserting in it anomalous zones that repre-
phones). The methodology is tested on synthetic models that represent sented empty or filled cavities or different anomalous zones.
cavities at different abscissas and depths, together with structures of Model A (Fig. 1 a), represents a single square section cavity. In the
similar shape but with physical parameters not indicative of the pres- middle of the section a square-shaped has been inserted (6 m × 6 m),
ence of cavities, and the advantages are discussed. Finally, the proposed having its top at depth of 6 m. This anomaly simulates an air-filled cavity
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 3

Fig. 1. Simulated models representing different combinations of cavities (vp = 330 m/s, ρ = 105 Ωm) and compact calcarenitic blocks (vp = 5000 m/s, ρ = 5000 Ωm): a) model A
representing a buried cavity with its top at depth of 6 m; b) model B representing three cavities, with their top at depths respectively of 4 m, 6 m and 8 m; c) model C representing a
calcarenitic compact block (left) with top at a depth of 4 m, and two cavities (middle and right), with their top at depths respectively of 6 m and 8 m; d) model D representing a
calcarenitic compact block (middle) with top at a depth of 6 m, and two cavities (left and right) with their top at depths respectively of 6 m and 8 m; e) model E representing a
calcarenitic compact block (right) with top at a depth of 8 m, and two cavities (left and middle) with their top at depths respectively of 4 m and 6 m; f) model F representing the
Foderà Quarry section, including four cavities of irregular shapes, with top at the same depth of 6 m.

characterized by a P-waves velocity of 330 m/s and a resistivity of 105 Finally, model F (Fig. 1 f), is a 2D section obtained by a detailed syn-
Ωm. thetic modeling based on high-resolution speleological survey carried
Model B (Fig. 1 b) represents three cavities and shows three anom- out the Foderà Quarry (Fig. 2 a) near Marsala, in Sicily (Fig. 2 b), that
alous square-shaped zones (6 m × 6 m) located between 29 m and has been chosen as field test site in order to compare predicted and ex-
35 m, between 45 m and 51 m, and between 61 m and 67 m respect perimental results. A map of the Quarry (Fig. 2 c) was plotted thanks to a
to left limit of the profile, and having the top at depths respectively of 3D speleological survey carried out in 2016. Starting from this survey a
4 m, 6 m and 8 m. The three anomalies are characterized by P-wave ve- section of the subsoil (AB) was obtained and used as a base to simulate
locity equal to 330 m/s and electrical resistivity equal to 105 Ωm. the synthetic model F; the profile corresponds to the location of the
Model C represents two cavities and a compact calcarenitic block electric and seismic tomography line carried out in field.
(Fig. 1 c). It presents the same three anomalous square-shaped zones Model F has been set on the same background considered for the
as the model B. However, the most superficial block (on the left) repre- other models. Moreover, it must be said that the range of seismic
sents a compact calcarenitic block with a P-wave velocity of 5000 m/s waves velocity and resistivity values has been chosen according to the
and a resistivity of 5000 Ωm. The other two deeper blocks, on the characteristic lithology of the quarry area: being a compact calcarenitic
other hand, represent two cavities with P-wave velocities equal to rock, vp reaches a maximum value of 3000 m/s and ρ is about 500 Ωm.
330 m/s and resistivity equal to 105 Ωm. The section obtained from the speleological survey intersects four
Model D, similarly to the model C, represents two cavities and a cavities that are sketched in the synthetic model, all at 6 m depth,
compact calcarenitic block (Fig. 1 d). The difference is that the compact with variable thicknesses that reach a maximum depth of 21 m from
calcarenitic block, with vp = 5000 m/s and ρ = 5000 Ωm, is the square- the countryside level. Starting from the left, the first cavity is located
shaped zone in the middle of the model, and the left and right square- from 28 m to 34 m from the origin of the profile, it has a thickness of
shaped zone represent air-filled cavities (vp = 330 m/s and ρ = 105 13 m and reaches a depth of 19 m; the second cavity is located from
Ωm). 45 m to 50 m, it has a thickness of 11 m and reaches a depth of 17 m;
Model E, similarly to models C and D, represents two cavities and a the third cavity, located from 58 m to 65 m, has an irregular shape
compact calcarenitic block (Fig. 1 e). In this case the calcarenitic block, and thickness ranging from a minimum of 2 m to a maximum of 8 m,
with vp = 5000 m/s and ρ = 5000 Ωm, is the square-shaped zone in it reaches a depth of 10 m; the fourth cavity located from 70 m to
the right of the model, and the left and middle square-shaped zone rep- 100 m, also has irregular shape and variable thickness up to a maximum
resent air-filled cavities (vp = 330 m/s and ρ = 105 Ωm). of 15 m, reaching a depth of 21 m. The values of the geophysical
4 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

Fig. 2. a) A picture of the Foderà Quarry, in which three tunnel entrances are visible; b) Ubication of “Parco delle Cave” near the city of Marsala, in Sicily; c) Map of the Foderà Quarry,
obtained after a 3D speleological survey, showing the trace A-B of the joint SRT and ERT survey.
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 5

parameters assigned to these cavities are the same as for the other syn- 2.2. Simulating joint-interpreted seismic refraction and electrical resistivity
thetic models: vp = 330 m / s and ρ = 105 Ωm. tomographies
At the beginning, the model obtained from the speleological survey
did not include the fourth cavity on the right, added later on the basis Starting from the aforementioned models, the execution of joint-
of the results of the field surveys and subsequently identified by spele- interpreted SRT and ERT were simulated, considering the same posi-
ologists. This is a clear example of how in field the choice of position and tions for the geophones and the electrodes, coinciding with the surface
total length of the array does not always turn out to be optimal. nodes of the model cells.

Fig. 3. Model A: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the likelihood
parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities are indicated by black solid lines. The joint interpreted section is contoured by a dashed white line.
6 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

For the seismic refraction survey 48 geophones were considered, and is ideal for making the most of the capabilities of modern multi-
2 m spaced, and 25 shots distributed along equal distances of 2 m, channel resistivity-meters. In fact, it allows a good resolution and reli-
with left and right offset equal to 2 m, for a total length of the seismic ability of the survey, according with a not too long acquiring time due
line equal to 100 m. This array should allow a good resolution of the in- to an adequately high ratio between the number of measures and that
verse model, with a quite homogeneous coverage. of current dipoles (Martorana et al., 2017a).
Seismic refraction data was inverted using Rayfract® (Intelligent Re- Inversion of predicted apparent resistivity data was performed using
source, Inc.) ver. 3.12. It uses the WET (Wavepath Eikonal Traveltime) the RES2DINV™ and RES3DINV™ software applications (Loke, 2013).
method (Schuster and Quintus-Bosz, 1993; Rohdewald, 2016). The The same set of damping and inversion parameters and the same cell
same set of damping and inversion parameters and the same cell sizes sizes were used for all the ERTs, in order to correctly compare the re-
were used for all the data sets related to different synthetic models, sults. Considering the large contrast in the apparent resistivity relative
aiming to obtain a reliable comparison among all the tomographies. A to adjacent measurements, a further model refinement was performed,
manual picking of each track was made to estimate the first arrival with mesh size equal to a half of the electrode distance. Moreover, a
times of P-waves. As a first step, synthetic traveltime data has been mesh grid with two nodes per electrode spacing was chosen in the for-
inverted using the Delta-t-V method (Gibson et al., 1979) which auto- ward modeling routine, in order to accurately calculate the theoretical
matically generates an initial 1D model directly from the chosen values. An L1 norm, iteratively reweighted least squares method
traveltimes (Gebrande and Miller, 1985; Rohdewald, 2011), to identify (Wolke and Schwetlick, 1988) was used to obtain models capable to
small features and velocity inversions. Subsequently, an iterative back- highlight sharp boundaries with a high resistivity contrast. We consider
projection of the wavefronts using a finite-difference solution to the this to be the best choice if we want to correctly outline the shape of a
eikonal equation (Qin et al., 1992) was applied to the Delta-t-V results. cavity.
The electrical resistivity survey was simulated considering 48 elec-
trodes, 2 m spaced, for a total length of 94 m. The inline dipole-dipole 2.3. Likelihood parameters for cavity presence
array was considered, with dipole length a ranging from 1 to 4 times
the electrode spacing and dipole-dipole distance n ranging from 1 to 8 The size, the depth of burial of the cavity and the contrast of the
times the electrode spacing. This data set comprises n. 945 measures physical properties between the host rock and the cavity, are aspects

Fig. 4. Model A: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.

 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 7

that play an important role for the evaluation of the geophysical and SRT data shows a region with high resistivity and low seismic ve-
method for the detection of cavities. ERT is a well-established and locity, this is considered as a region in which a vacuum is present
widely used method to detect cavities; however, the information ob- (Sheehan et al., 2005). Vice versa, if an anomaly shows low resistivity
tained from SRT is capable of eliminating some potential ambiguities and high seismic velocity, the presence of the vacuum in the subsoil
in the interpretation of ERT data. Both geophysical methods have can be excluded. In fact, therefore, the two correlated physical prop-
been applied to both synthetic and real data, allowing to identify erties whose high values can indicate the presence of cavities are the
and evaluate the geometric characteristics of the cavities analyzed electrical resistivity and the seismic slowness sₚ, inverse of the seis-
(Cardarelli et al., 2010). In general, if the combined inversion of ERT mic velocity vₚ.

Fig. 5. Model B: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the likelihood
parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities are indicated by black solid lines. The joint interpreted section is contoured by a dashed white line.
8 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

To detect a body in the subsoil it is also necessary to have adequate The second one, here named P2, is the ratio between P1 and the nor-
coverage of seismic rays. In the case of a cavity the seismic rays will con- malized seismic ray density d:
centrate in the area around the cavity and will decrease at the cavity it-


self. Moreover, in order for the SRT to be able to detect a cavity, it is P2 ρ; sp; d ¼ N ðLog ρÞ  N sp =N ðdÞ ð3Þ
essential that the seismic rays penetrate a depth greater than that of
the cavity base, to be thus critically refracted to the surface, if the veloc- These likelihood parameters are able to help to differentiate with
ities of the underlying layers increase with depth. more precision the cavity (values close to 1) from other structures
With the above considerations in mind, we choose two likelihood that could imply high values of electrical resistivity or high values or
parameters that could highlight the presence of cavities. These parame- seismic slowness but not both these contemporary events. P1 and P2
ters have been estimated starting from the normalization of the physical were calculated for each cell of the inverse models.
quantities considered. The normalized values of each quantity is calcu-
lated as follows: 2.4. Cluster analysis

NðxÞ ¼ ðx−xmin Þ=ðxmax −xmin Þ ð1Þ In cluster analysis there are two types of classification algorithms:
hierarchical algorithms and non-hierarchical algorithms. As for hierar-
chical algorithms, each statistical unit is part of a larger group, which
where x is the physical quantity involved and xmin and xmax are the min-
is itself contained in an even larger group, until a single group contain-
imum and maximum values in the inverse model.
ing the entire population is obtained. For non-hierarchical algorithms it
The first parameter, here named P1, is the product between the nor-
is necessary to decide the number of clusters to be obtained and, based
malized logarithm of resistivity ρ and the normalized P-wave slowness
on this number, the algorithm tries to obtain the best possible grouping.
sp = 1 / vp:
Generally, cluster analysis does not require an a priori model of inter-


 pretation (Fabbris, 1983). The centroid-based clustering technique can
P1 ρ; sp ¼ N ðLog ρÞ  N sp ð2Þ
be theoretically considered as an optimization problem that proceeds

Fig. 6. Model B: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.

 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 9

according to an iterative process: in the preliminary phase it is neces- the number of clusters, k, and the coordinates of the initial centroid.
sary to find the k cluster centers and to associate each object to its clus- This aspect is considered one of the major problems of these algorithms,
ter, so that the distances squared by the centroid of the cluster are kept because an inappropriate choice of k can produce poor results. It is also
to a minimum; subsequently it is needed to calculate the new averages important to do diagnostic research to determine the number of clusters
to be the centroids of the observations in the new clusters. The algo- in the data set.
rithm reaches its final optimum (local) when the assignments no longer In this work, a statistical approach based on non-hierarchical cluster
change and the clustering procedure is interrupted; however, in some analysis (k-means) has been applied, as it is the most efficient approach
cases there is no certainty of finding the global optimum. Typically, in computational terms, less influenced by the anomalous values and al-
centroid-based algorithms require to identify and specify in advance lows a statistical unit to change your cluster during the iterative process

Fig. 7. Model C: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the likelihood
parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities and compact blocks are indicated by black and white solid lines respectively. The joint interpreted section is
contoured by a dashed white line.
10 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

(Martorana et al., 2017b). Moreover, the final results obtained by this resistivity. The choice of the optimal number of clusters (k) have been
non-hierarchical approach are certainly simpler to interpret because optimized maximizing R2 parameter, taking into account the intra-
they are represented by a single partition, which is interpreted on the cluster (DEVin) and inter-cluster (DEVout) variances:
basis of its hypotheses. The idea of using cluster analysis in geophysical
exploration is that the structures in the subsoil are easily distinguishable R2 ¼ DEVout=DEVt; DEVt ¼ DEVin þ DEVout ¼ 1 ð5Þ
using a set of different physical properties compared to a single prop-
erty. To perform cluster analysis on data, an algorithm implemented
in MATLAB® is used. This algorithm aims to group units into classes in Passing from k + 1 to k groups (i.e. the aggregation phase) DEVin in-
a single solution, based on optimization criteria, minimizing variance creases, while obviously DEVout decreases. At each iteration groups are
within the cluster and at the same time maximizing the distance be- aggregated among them so there is the least increase in variance within
tween clusters. The data acquired in this study are favorable to cluster the groups.
analysis, since the values of the propagation velocity of the wave P, The first phase of elaboration consists in the independent derivation
the density of the seismic rays and the electrical resistivity are available of the resistivity inversion model and the inversion model of seismic ve-
on coincident sections with comparable resolution. The distance of each locities; the values obtained are processed by the algorithm proceeding
element from the initial nucleus and from the nuclei obtained after each with the first iteration by choosing the coordinates of these quantities,
iteration was calculated from the weighted sum of the Euclidean dis- to be represented in a common graph x-z, where each node of this
tances of all the parameters considered: grid has a value of electrical resistivity, seismic velocity and seismic
ray density. At the end of the iterations, an x-z profile is generated
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi r
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  r
ffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
2 2 2 2 2 where the nodes belonging to the profile are colored according to the
D¼a dx þ dy þb dC þc dV þd dρ ð4Þ
starting cluster.
With clustering it is also possible to analyze the different parameters
where a, b, c and d are weights, dx, dy, dC, dV and dρ are respectively the by plotting them within a same 3D multiparametric graph, delimiting
differences between the spatial coordinates x and y, the density of the the limits of the data group areas and facilitating the reading of the de-
seismic rays, the propagation velocity of the P wave and the electrical tected information and obtaining a more complete view. Generally, the

Fig. 8. Model C: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.

 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 11

graphical representation of the cluster occurs through the use of a color seismic data required 20 iterations to reach a misfit value of 4 μs
code that encloses a group of clusters with similar values. (Schuster, 1993). The distribution of the ray path density (Fig. 3
a) shows a decrease at the cavity and an increase at the lower limit of
3. Synthetic test results the cavity. The Seismic Refraction Tomography (Fig. 3 b), shows an evi-
dent low velocity at the cavity, even if the inverted values are greater
3.1. Model A than those of the theoretical model. The electrical resistivity tomography
(Fig. 3 c) show high values of resistivity at the cavity, low values even if
In Fig. 3 the results of the inversion of simulated data obtained lower than those of the theoretical model. Both SRT and ERT identify the
starting from the model A are showed. The inversion processing of the cavity, even if with a wider anomaly in ERT (Riddle et al., 2010).

Fig. 9. Model D: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the likelihood
parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities and compact blocks are indicated by black and white solid lines respectively. The joint interpreted section is
contoured by a dashed white line.
12 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

The distribution of the likelihood parameter P1 (Fig. 3 d) and P2 depth of the cavities increases; the deeper cavity, located at 8 m
(Fig. 3 e) shows that the cavity area is well delimited, with both values depth, could be identified with more difficulty because of a not so
close to one at the cavity. high resistivity contrast and more smoothed boundaries; low resistivity
The cluster analysis performed considering the geophysical parame- values delineate the volume surrounding the cavities.
ters of the electrical resistivity, the propagation velocity of the P waves The distribution of the likelihood parameter P1 (Fig. 5 d) and P2 (Fig. 5
and the coverage of the seismic beams, assigns each cell of the model e) are able to discriminate the three cavities to which low values of con-
to a specific cluster characterized by similar trends of the aforemen- ductivity, P-wave velocity and ray path density correspond. In this case
tioned parameters. Depending on the choice of number k of the classes, the identification of the deeper cavity is more accurate than the inversion
the algorithm returned different patterns (Fig. 4). The cavity is well of the only SRT o ERT, especially if we look at the distribution of P2.
identified if k ≥ 3. If k = 5 we can also differentiate the vertical gradient The cluster analysis is obviously dependent on the choice of the
of seismic velocity. value of k classes. With k = 2 (Fig. 6 a) the image discriminates two
large clusters the delimit quite well only the bottom of the cavities.
For k = 3 (Fig. 6 b) there is a cluster that delimit the area underneath
3.2. Model B the cavities, yet information about the cavities is insufficient or even
missing for the deeper cavity.
The results of the inversion of simulated data obtained starting from If k = 4 (Fig. 6 c) all the cavities are well isolated, identified with a
the model B are showed in Fig. 5. After 20 iterations a misfit of 6 μs was single cluster. With k = 5 (Fig. 6 d) another cluster defines the outer
obtained between predicted and measured travel times of P-waves. The edges of the image.
distribution of the ray path density (Fig. 5 a) shows evident decreases
corresponding to the three cavity zones located at the different depths;
while a high ray path density zone, between 14 m and 20 m of depth, is 3.3. Model C
placed underneath the cavities marking the lower limits of these. The
SRT (Fig. 5 b) shows a decrease in velocity at the three cavities. ERT The results of the inversion of simulated data obtained starting from
(Fig. 5 c) shows high resistivity values that however decrease as the the model C are showed in Fig. 7. After 50 iterations a misfit of 12 μs was

Fig. 10. Model D: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 13

obtained between predicted and measured travel times of P-waves. 6 m and at 8 m of depth where the cavities are located; a high seismic
There is a decrease in the ray path density (Fig. 7 a), corresponding to velocity zone is showed in depth and tends to expand towards the sur-
the two cavities; while a high density area is evident at 4 m of depth face in correspondence to the simulated compact calcarenitic block is lo-
(corresponding to the simulated compact calcarenitic block), but it cated. ERT (Fig. 7 c) shows high resistivity values for the cavity at a
also propagates at depths between 11 m and 20 m below the cavities, depth of 6 m and low values and strongly blunted limits for the cavity
delineating the lower limits. SRT (Fig. 7 b) shows the decrease of vp at located at 8 m depth. It also well highlights the calcarenitic compact

Fig. 11. Model E: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the
likelihood parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities and compact blocks are indicated by black and white solid lines respectively. The joint
interpreted section is contoured by a dashed white line.
14 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

block at 4 m depth, with higher resistivity values compared to adjacent while a high density area is located at the center of the image, where the
deeper cavities. ERT therefore fails to differentiate the calcarenitic com- calcarenitic block is located, starting from a depth of 6 m to about 18 m,
pact block from the cavities. and spreading below the two cavities, marking their lower limits. SRT
In the anomalous zones the likelihood parameter P1 (Fig. 7 d) shows (Fig. 9 b) detects the decrease in seismic velocity in the two cavities,
values between 0.7 and 1 progressively decreasing with depth and with while maximum velocity values are at the center of the image corre-
increasingly blunt limits; in the same zones the range of variation of P2 sponding to the position of the calcarenitic block. ERT (Fig. 9 c) well
(Fig. 7 e) is smaller (0.85–1). In this case the parameters P1 and P2 do not identifies the most superficial cavity at 4 m depth with the highest resis-
allow to discriminate between the cavities and the compact calcarenitic tivity values; but it associates low values of resistivity and very
block. smoothed anomalies both to the cavity at 8 m of depth and to the
Even for this model the cluster analysis is dependent on the choice of calcarenitic block at 6 m of depth.
the value of k classes (Fig. 8). The three anomalous zones are identified The likelihood parameter P1 (Fig. 9 d) shows values close to 1 corre-
if k N 2 (Fig. 8 a), however only by choosing 5 clusters it is possible to dis- sponding the shallower cavity (4 m of depth), while similar values, be-
criminate the two cavities from the compact calcarenitic block that is tween 0.5 and 0.7 are showed by the cavity at 8 m depth and the
identified by a specific cluster, while the two cavities are identified by calcarenitic block at 6 m depth. The likelihood parameter P2 (Fig. 9
another cluster (Fig. 8 d). e) shows values close to one for the cavity at 4 m depth and very high
for that at 8 m depth. Vice versa, the calcarenitic block is characterized
3.4. Model D by high parameter values, but a heterogeneous pattern that does not
allow a clear identification of its shape.
In Fig. 9 the results of the inversion of simulated data obtained Images of spatial distribution of clusters for this model (Fig. 10)
starting from the model D are showed. The inversion processing of are less clear than those relating to the previously discussed model:
the seismic data required 50 iterations to reach a misfit value of the three anomalous zones are identified starting from k ≥ 4
12 μs. but, even for this model, the cavities and the calcarenitic block are
The distribution of the ray path density (Fig. 9 a) highlights two discriminated by two different clusters only if we chose k = 5
areas of low density in the approximate correspondence of the cavities; (Fig. 10 e).

Fig. 12. Model E: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 15

3.5. Model E of the calcarenitic block, at a depth of 8 m. this area is not limited to
the calcarenitic block but extends deep beneath the base of the cavities
The results of the inversion of simulated data obtained starting from up to 20 m depth. SRT (Fig. 11 b) shows high velocity values that are lo-
the model E are showed in Fig. 11. After 50 iterations a misfit of 11 μs cated at the calcarenitic compact block but are not limited by this, ex-
was obtained between predicted and measured travel times of P- tending laterally outside it. Non-significant decreases of P-wave
waves. Looking at the distribution of the ray path density (Fig. 11 velocity are found at the cavities while low values are instead found in
a) it's easy to identify two areas of low density at the cavities at 4 m the cavity areas at 4 m and 6 m depth. ERT (Fig. 11 c) discriminates
and 6 m depth; while a high density zone is found in correspondence quite well with high resistivity values the two surface cavities at 4 m

Fig. 13. Model F: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the
likelihood parameter P1; e) imaging of the likelihood parameter P2. Boundaries of cavities are indicated by black solid lines. The joint interpreted section is contoured by a dashed
white line.
16 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

and 6 m of depth, while to the calcarenitic compact block it associates a nevertheless it also presents for the portion investigated a decrease,
highly smoothed area with low values of resistivity. Both parameters P1 smaller than the other cavities, of the ray path density. An area of
and P2, but a little better P2, discriminate well the cavities showing high ray coverage is located below the cavities between 14 m depth
values close to one in their correspondence. On the other hand, the com- and 23 m depth, which perfectly delimits the base of the voids. SRT
pact calcarenitic block is poorly highlighted (Fig. 11 d, e). (Fig. 13 b) shows seismic velocity decreases in correspondence to
Starting from k = 3 the cluster analysis (Fig. 12) includes the two all the cavities. ERT (Fig. 13 c) highlights cavities with four areas of
cavities in a specific cluster and well delimits their shape. Instead, the high resistivity (values higher than 5000 Ωm are in correspondence
calcarenitic block, due to its depth, is never clearly identified, even if with the greatest cavity). The distribution of likelihood parameters
for k ≥ 4 the images show a cluster that encloses it but that also extends P1 and P2 (Fig. 13 d, e) shows a general distribution of values close
in depth to areas of high ray path. to 1 both superficially than within the cavities. In this case the bot-
tom and the lateral boundaries of the cavities are delimited quite
3.6. Model F well, but not the top.
Depending on the choice of the value of k classes, the cluster analysis
Fig. 13 shows the results of the inversion of simulated data obtained returned different representation images. With k = 2 (Fig. 14 a) the
starting from the model F. The inversion processing of the seismic data image identifies two different clusters, almost including the cavities in
required 20 iterations to reach a misfit value of 25 μs. Considering the a single cluster that comprises also the shallower subsoil. With k = 3
distribution of the ray path density (Fig. 13 a), three zones of low (Fig. 14 b) the cavities are specifically represented by a separate cluster
density, located from 28 m to 34 m, from 45 m to 50 m, from 58 m even if its shape do not extend to the real depths of the cavities. With
to 65 m with respect to the origin, correspond to the three cavity k = 4 and k = 5 (Fig. 14 c, d) the deep extension of the cavities is not
zones all located at a depth of 6 m. The fourth cavity extends also be- yet well identified by the cluster, however cluster analysis generates a
yond the coverage area, and therefore cannot be fully investigated, new cluster that extends beneath the cavities delimiting their bottoms.

Fig. 14. Model F: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 17

4. Test site data analysis thicknesses, with calcareous granules and carbonate cement. The rock
is vacuolar, highly porous and poorly fractured.
Experimental data were acquired at the test site of Parco delle Cave The Foderà quarry (Fig. 2), recently defined also as “Parco delle
of Marsala (Italy, Sicily) to compare theoretical and experimental data, Cave”, is now used for touristic and cultural activities.
optimize the inversion process and simplify the step of interpretation The quarries of this area are mainly open-air quarries, with almost
of the same data. regular shapes and sub-vertical excavation fronts. Most of these
Since the time of the Phoenicians up to the present day this area has quarries have now been filled, even if partially, with waste materials
been exploited for the extraction of the Marsala calcarenite (Pleistocene coming from excavation faces of the same quarry or nearby cavities.
Medio Inf.). This calcarenite is rather homogeneous and generally well- In the Park there are also several underground cavities, located at
cemented, light yellow in color, poorly fossiliferous, with high varying depths with respect to the ground level, made by excavation

Fig. 15. Foderà Quarry: results of inversion of simulated P-waves travel-times and apparent resistivity data. a) ray path density d; b) SRT of P-waves velocity vp; c) ERT; d) imaging of the
likelihood parameter P1; e) imaging of the likelihood parameter P2. The joint interpreted section is contoured by a dashed white line.
18 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

of tunnels sustained by pillars. In more recent times, particularly in the software used for speleological activities. In particular, the Leica X310
post-war period, the latter were mostly abandoned due to various laser distance gauge was used with precision in measurements of
causes, as the huge cost of excavating the material, or the chemical- 2 mm and 0.5° RMS for angular measurements. The calcarenitic cavities
physical degradation of the calcarenite that generates instability phe- of the Park of the Caves are almost all accessible (there are at least 3 en-
nomena like collapses or deformations of the pillars and vaults trances), in fact the relief carried out has allowed to determine the var-
(Bonamini et al., 2013). This situation of abandonment led to an in- ious configurations of the voids, the depths and their dimensions,
crease in the exploitation of open-air quarries, allowing to intercept allowing to rebuild the plant of the area and the reference section
the hypogeal areas. (Fig. 2 c).
The method of open excavation involves the removal of the surface The refraction seismic tomography investigation was carried out
alteration blanket that generally presents discontinuity surfaces, infil- using 48 geophones with 2 m intervals between the geophones. Along
tration zones or physical alterations due to the long exposure from ex- the same alignment in the NE-SW direction, an electrical resistivity to-
ogenous agents. And it allows to exploit the underlying rock with mography was acquired with 48 electrodes, putting the electrodes in
more appropriate characteristics for building purposes. The cultivation the same positions as the geophones. The inversion of the seismic data
in underground instead, a technique currently in disuse, is practiced in (after 50 iterations and a misfit value of 0.012 ms) has allowed to obtain
order not to damage the upper ground. Access to the tunnels is through the seismic rays density d and of P-waves velocity vp. The results of the
wells that are also used for lighting, ventilation of the tunnels and exit inversion are showed in Fig. 15. These distribution maps are character-
routes for the excavated sections. The progress procedure takes place ized by a low seismic ray density and low P-wave velocity values in cor-
with the formation of underground tunnels and halls of significant respondence with the cavity blocks (Fig. 15 a, b). Also, the electrical
size, paying particular attention to leaving massive pillars supporting resistivity model (Fig. 15 c) identifies three of the four cavities present
the rocky vaults. This technique is particularly decisive when the in the real model. The two proposed correlation parameters are able
lithotype is of poor quality and therefore difficult to use. to differentiate with greater precision the cavity (values close to
The underground quarry chosen as field test develops in a series of 0) from the high compact rock (high values), except for the cavity posi-
tunnels and connected rooms, most of which have been the subject of tioned at the end of acquisition line, due to the low resolution of the data
a phase of hypo geographical topographic surveys with tools and in this area (Fig. 15 d, e).

Fig. 16. Foderà Quarry: results of cluster analysis by choosing different number of classes k. a) k = 2; b) k = 3; c) k = 4; d) k = 5.
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 19

It must be said that the complexity of the shape of the tunnels and ERT carried out in situ is also disturbed by variations in the resistivity
chambers of the Foderà Quarry implies that the 2D inversions are af- of the rock, and by the detritus, which are not foreseen by the theoreti-
fected by the ghost anomalies generated by the three-dimensionality cal model.
of the intercepted structures and by the presence of other structures Finally, the results of a non-hierarchical cluster analysis algorithm
close to the survey lines (Martorana et al., 2018). Moreover, the 2D for a number of clusters from 2 to 5 are showed (Fig. 16). With k = 2

Fig. 17. Synthetic data (top) and experimental ones (bottom) related to the Foderà Quarry, plotted in the multiparameter space (seismic ray density, P-waves velocity and electrical
resistivity). Yellow squares represent data relative to cavities. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
20 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069

(Fig. 16 a) the map identifies two different clusters, grouping almost all because it is less influenced by abnormal values, and allows a statistical
cavities in a single cluster. From k = 3 to k = 5 (Fig. 16 b, c, d) the first unit to change its cluster during the iterative process. Depending on the
two cavities are grouped in the same cluster, while the third cavity is as- choice of the number of clusters to be identified (values between 2 and
sociated to a different cluster. 5), cluster distribution maps have been constructed in the multi-
parameter space, allowing to define certain variability limits for the se-
5. Discussion lected parameters, for synthetic and experimental data. Finally, experi-
mental data show that electrical and seismic tomographies is not
Results of syntethic tests show that the detectability of the voids and influenced by the presence of cavities without lateral continuity.
the compact blocks of limestone is strongly influenced by the depth of
the target and by its lateral position with respect to the length of the to-
mography. This obviously applies both considering the seismic velocity Declaration of Competing Interest
and the electrical resistivity as the investigation parameter. The electri-
cal resistivity certainly gives higher contrasts than the seismic velocity, The authors declare that they have no known competing financial
but it cannot allow to distinguish a cavity from a compact block of interests or personal relationships that could have appeared to influ-
calcarenite with sufficient reliability, especially if this is close enough ence the work reported in this paper.
to the surface, to generate an evident anomaly. On the other hand, seis-
mic refraction tomography, although generating anomalies of different References
sign (positive if in the presence of compact blocks, negative if in the
Archie, G.E., 1942. The electrical resistivity log as an aid in determining some reservoir
presence of voids), does not accurately identify the geometries of the
characteristics. I. Pet. Technol. 5. https://doi.org/10.2118/942054-G.
structures, due to the intrinsic limitations of the method. Barbarito, L., 1999. L’analisi di settore: metodologia e applicazioni. Milano, Franco Angeli
The use of the likelihood parameters can help the interpretation, fa- Ed.
Bonamini, M., Di Maggio, C., Lollino, P., Madonia, G., Parise, M., Vattano, M., 2013. Study of
cilitating the distinction between the voids (in which the values of P1
anthropogenic sinkholes in the Marsala area (wester Sicily) through fiel surveys and
and P2 tend to 1), from other structures that produce lower values, but numerical analyses of instability processes in underground quarries. GeoItalia 2011,
also their yield depends on position and depth of the target. The P2 VIII Forum Italiano di Scienze della Terra. IT 74-74. http://hdl.handle.net/10447/
shows contrasts greater than P1, if in the presence of voids, but a 60742.
Bottari, C., Albano, M., Capizzi, P., D’Alessandro, A., Doumaz, F., Martorana, R., Moro, M.,
lower geometric precision. Saroli, M., 2018a. Recognition of earthquake-induced damage in the Abakainon ne-
The cluster analysis is able to give good results when choosing a suf- cropolis (NE Sicily): results from geomorphological, geophysical and numerical anal-
ficiently high number of k (k ≥ 4 if it is necessary to discriminate be- yses. Pure Appl. Geophys. 175 (1), 133–148. https://doi.org/10.1007/s00024-017-
1653-4.
tween voids and compact blocks, also taking into account the Bottari, C., Capizzi, P., Cavallaro, D., Coltelli, M., D’Alessandro, A., Lodato, L., Martorana, R.,
stratification). However it must always be compared with the tomo- Pisciotta, A., Scudero, S., 2018b. Coseismic damage at an archaeological site in Sicily,
graphic images to define better the shape of the structures. Italy: evidence of Roman earthquake surface faulting. Surv. Geophys. 39 (6),
1263–1284. https://doi.org/10.1007/s10712-018-9482-2.
The comparison between the results of filed survey and model F, Capizzi, P., Cosentino, P.L., Fiandaca, G., Martorana, R., Messina, P., 2005. 2D GPR and
based on the speleological survey, confirm the effectiveness of the pro- geoelectrical modelling: tests on man–made tunnels and cavities. 11th European
posed methods, but highlight other problems not considered in the syn- Meeting of Environmental and Engineering Geophysics of the Near Surface Geosci-
ence Division of the EAGE, Near Surface 2005; Palermo; Code 104313, B036 1–4.
thetic models, with regard to the lithological heterogeneities and the
https://doi.org/10.3997/2214-4609-pdb.13.B036 ISBN 90-737-8191-4.
three-dimensionality of the cavities that can determine for them assign- Cardarelli, E., Fischanger, F., Piro, S., 2008. Integrated geophysical survey to detect buried
ment to a different cluster. structures for archaeological prospecting. A case-history at Sabine Necropolis (Rome,
All cluster data relative to the cavity have been highlighted, for both Italy). Near Surf. Geophysics 6 (1), 15–20. https://doi.org/10.3997/1873-
0604.2007027.
synthetic data and experimental ones (yellow squares) (Fig. 17). To fa- Cardarelli, E., Cercato, M., Cerreto, A., Di Filippo, G., 2010. Electrical resistivity and seismic
cilitate the comparison between these two graphs the values of the refraction tomography to detect buried cavities. Geophys. Prospect. 58 (4), 685–695.
three components (seismic ray density, P-waves velocity and electrical https://doi.org/10.1111/j.1365-2478.2009.00854.x.
Dahlin, T., Bjelm, L., Svensson, C., 1999. Use of electrical imaging in site investigations for a
resistivity) have been normalized. In this way it was possible to high- railway tunnel through the Hallandsas Horst, Sweden. Q. J. Eng. Geol. Hydrogeol. 32,
light how the cavities, considering the values of the three parameters 163–172. https://doi.org/10.1144/GSL.QJEG.1999.032.P2.06.
analyzed, are arranged in the same region of the multiparameter Dell’Aversana, P., 2001. Interpretation of seismic, MT and gravity data in a thrust belt in-
terpretation. First Break 19 (6), 335–341. https://doi.org/10.1046/j.1365-
space favoring a rapid identification of the cavities present in the sub- 2397.2001.00158.x.
soil. The region in the multiparametric space where the cavities are lo- Di Giuseppe, M.G., Troiano, A., Troise, C., De Natale, G., 2014. K-means clustering as tool
cated is characterized by high values of electrical resistivity and low for multivariate geophysical data analysis. An application to shallow fault zone imag-
ing. J. Appl. Geophys. 101, 108–115. https://doi.org/10.1016/j.jappgeo.2013.12.004.
values of seismic ray density and P-waves velocity.
Fabbris, L., 1983. Analisi esplorativa di dati multidimensionali (Cleup ed).
Fasani, G.B., Bozzano, F., Cardarelli, E., Cercato, M., 2013. Underground cavity investigation
6. Conclusions within the city of Rome (Italy): a multi-disciplinary approach combining geological
and geophysical data. Eng. Geol. 152, 109–121. https://doi.org/10.1016/j.
enggeo.2012.10.006.
This work shows how the joint study of seismic refraction tomogra- Gallardo, L.A., Meju, M.A., 2003. Characterization of heterogeneous near-surface materials
phy and electrical resistivity tomography techniques leads to the deter- by joint 2D inversion of dc resistivity and seismic data. Geophys. Res. Lett. 30 (13).
mination of cavities buried in the subsoil, minimizing possible https://doi.org/10.1029/2003GL017370.
Gallardo, L.A., Meju, M.A., 2004. Joint two-dimensional DC Resistivity and Seismic travel
interpretative ambiguities and producing the most robust results at time inversion with cross-gradients constraints. J. Geophys. Res. 109 (B3). https://
the same time. doi.org/10.1029/2003JB002716.
In order to facilitate discrimination between voids and other litho- Gallardo, L.A., Meju, M.A., 2007. Joint two-dimensional cross-gradient imaging of seismic
travel time data for structural and lithological classification. Geophys. J. Int. 169 (3),
logical structures, two likelihood parameters that correlate the physical 1261–1272. https://doi.org/10.1111/j.1365-246X.2007.03366.x.
quantities has been defined. The use of these parameters facilitates the Gallardo, L.A., Meju, M.A., 2011. Structure coupled multiphysics imaging in geophysical
distinction between voids and other structures but their effectiveness sciences. Rev. Geophys. 49, RG1003. https://doi.org/10.1029/2010RG000330.
Gebrande, H., Miller, H., 1985. Refraktionsseismik Angewandte Geowissenschaften II ed F
depends on the position and depth of the investigate target.
Bender (Stuttgart: F Enke). pp. 226–260.
The cluster analysis performed on static units defined by electrical Gibson, B.S., Odegard, M.E., Sutton, G.H., 1979. Nonlinear least-squares inversion of
resistivity values, P-wave velocities, and seismic density on coincident traveltime data for a linear velocity-depth relationship. Geophysics 44 (2),
sections, seems to be an effective method to characterize structures 185–194. https://doi.org/10.1190/1.1440960.
Hartigan, J.A., 1975. Clustering Algorithms. John Wiley & Sons, Inc.
near the surface in the presence of complex geological settings. The Hellman, K., Ronczka, M., Günther, T., Wennermark, M., Rücker, C., Dahlin, T., 2017. Struc-
use of the non-hierarchical clustering algorithm has been chosen turally coupled inversion of ERT and refraction seismic data combined with cluster-
 A. Carollo et al. / Journal of Applied Geophysics 178 (2020) 104069 21

based model integration. J. Appl. Geophys. 143, 169–181. https://doi.org/10.1016/j. Orfanos, C., Apostolopoulos, G., Amolochitus, G., Leontarakis, K., 2008. Integrated geo-
jappgeo.2017.06.008. physical approach for the detection of underground voids in a Construction site.
Imposa, S., Grassi, S., Di Raimondo, S., Patti, G., Lombardo, G., Panzera, F., 2018. Seismic re- 70th EAGE Conference & Exhibition, Rome, Italy, Extended Abstracts, P097.
fraction tomography surveys as a method for voids detection: an application to the Qin, F., Luo, Y., Olsen, K., Cai, W., Schuster, G.T., 1992. Finite-difference solution of the
archaeological park of Cava Ispica, Sicily, Italy. Int. J. Architect. Herit. 12 (5), eikonal equation along expanding wavefronts. Geophysics 57 (3), 478–487. https://
806–815. https://doi.org/10.1080/15583058.2017.1419311. doi.org/10.1190/1.1443263.
Kotyrba, B., Schmidt, V., 2014. Combination of seismic and resistivity tomography for the Rayfract, 2010a. Palmer 2010 Syncline Model construction and forward modeling, with
detection of abandoned mine workings in Münster/Westfalen, Germany: improved Rayfract® 3.25 and Golden Software Surfer® 11. Downloaded from. http://rayfract.
data interpretation by cluster analysis. Near Surf. Geophysics, 2014 12 (3), com/tutorials/palmfig9.pdf.
415–425. https://doi.org/10.3997/1873-0604.2013056. Rayfract, 2010b. Basement Step Mode construction and forward modeling, with
Lees, J.M., VanDecar, J.C., 1991. Seismic tomography constrained by Bouguer gravity Rayfract® 3.25 and Golden Software Surfer® 11. Downloaded from. http://rayfract.
anomalies: applications in western Washington. Pure Appl. Geophys. 135, 31–52. com/tutorials/step.pdf.
https://doi.org/10.1007/BF00877007. Riddle, G.I., Hickey, C.J., Schmitt, D.R., 2010. Subsurface tunnel detection using Electrical
Resistivity Tomography and Seismic Refraction Tomography: a case study. 23rd
Linder, S., Paasche, H., Tronicke, J., Niederleithinger, E., Vienken, T., 2010. Zonal coopera-
EEGS Symposium on the Application of Geophysics to Engineering and Environmen-
tive inversion of crosshole P-wave, S-wave, and georadar traveltime data sets.
tal Problems (pp. cp-175). European Association of Geoscientists & Engineers,
J. Appl. Geophys. 72 (4), 254–262. https://doi.org/10.1016/j.jappgeo.2010.10.003.
pp. 552–562. https://doi.org/10.3997/2214-4609-pdb.175.SAGEEP064.
Loke, M.H., 2013. Tutorial: 2-D and 3-D Electrical Imaging Surveys. www.geotomosoft. Rohdewald, S.R., 2011. Interpretation of first-arrival travel times with wavepath eikonal
com. traveltime inversion and wavefront refraction method. Symposium on the Applica-
Maraio, S., Bruno, P.P.G., 2015. Near-surface voids in the Neapolitan Volcanic Tuff (Italy) tion of Geophysics to Engineering and Environmental Problems. Society of Explora-
detected by seismic refraction tomography. Near Surface Geoscience 2015-21st tion Geophysicists. https://doi.org/10.4133/1.3614086.
European Meeting of Environmental and Engineering Geophysics (Vol. 2015, no. 1, Rohdewald, S.R., 2016. Rayfract Online Help [Online]. Intelligent Resource Inc. Available:.
pp. 1-5). European Association of Geoscientists & Engineers. https://doi.org/ http://www.rayfract.com.
10.3997/2214-4609.201413748. Schuster, G.T., Quintus-Bosz, A., 1993. Wavepath eikonal traveltime inversion: theory.
Martorana, R., Capizzi, P., D’Alessandro, A., Luzio, D., 2017a. Comparison of different sets of Geophysics 58, 1314–1323. https://doi.org/10.1190/1.1443514.
array configurations for multichannel 2D ERT acquisition. J. Appl. Geophys. 137, Scudero, S., Martorana, R., Capizzi, P., Pisciotta, A., D’Alessandro, A., Bottari, C., Di Stefano,
34–48. https://doi.org/10.1016/j.jappgeo.2016.12.012 (ISSN: 0926-9851). G., 2018. Integrated Geophysical Investigations at the Greek Kamarina Site (Southern
Martorana, R., Capizzi, P., Avellone, G., D’Alessandro, A., Siragusa, R., Luzio, D., 2017b. As- Sicily, Italy). Surv. Geophys. 39 (6), 1181–1200. https://doi.org/10.1007/s10712-018-
sessment of a geological model by surface wave analyses. J. Geophys. Eng. 14 (1), 9483-1.
159–172. https://doi.org/10.1088/1742-2140/14/1/159. Sheehan, J.R., Doll, W.E., Watson, D.B., Mandell, W.A., 2005. Application of Seismic Refrac-
Martorana, R., Capizzi, P., Carollo, A., 2018. Misinterpretation caused by 3D effects on 2D tion Tomography to karst cavities. US Geological Survey Karst Interest Group Pro-
Electrical Resistivity Tomography: tests on simple models. 2nd Conference on Geo- ceedings, Rapid City, South Dakota, 12–15 September 2005, 29–38.
physics for Mineral Exploration and Mining (Vol. 2018, no. 1, pp. cp-566).European Tronicke, J., Holliger, K., Barrash, W., Knoll, M.D., 2004. Multivariate analysis of cross-hole
Association of Geoscientists & Engineers. https://doi.org/10.3997/2214- georadar velocity and attenuation tomograms for aquifer zonation. Water Resour.
4609.201802560 Res. 40, W01519. https://doi.org/10.1029/2003WR002031.
McDonald, R., Davies, R., 2003. Integrated geophysical surveys applied to karstic studies. Van Schoor, M., 2002. Detection of sinkholes using 2D electrical resistivity imaging.
First Break 21 (10). J. Appl. Geophys. 50, 393–399. https://doi.org/10.1016/S0926-9851(02)00166-0.
Meju, M.A., Gallardo, L.A., Mohamed, A.K., 2003. Evidence for correlation of electrical re- Wolke, R., Schwetlick, H., 1988. Iteratively reweighted least squares algorithms, conver-
sistivity and seismic velocity in heterogeneous near-surface materials. Geophys. gence analysis, and numerical comparisons. SIAM J. Sci. Stat. Comput. 9 (5),
Res. Lett. 30, 1373. https://doi.org/10.1029/2002GL016048. 907–921. https://doi.org/10.1137/0909062.
Wyllie, M.R.J., Gregory, A.R., Gardner, L.W., 1956. Elastic wave velocities in heterogeneous
Orfanos, C., Apostolopoulos, G., 2013. Multiparameter analysis of geophysical methods for
and porous media. Geophysics 21 (1), 41–70. https://doi.org/10.1190/1.1438217.
target detection: the unified geophysical model approach. Geophysics 78 (6),
IM1–IM13. https://doi.org/10.1190/geo2012-0285.1.