See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/325982488

Three Geostatistical Methods for Hydrofacies Simulation Ranked Using a

Large Borehole Lithology Dataset from the Venice Hinterland (NE Italy)

Article  in  Water · June 2018

DOI: 10.3390/w10070844

CITATIONS READS

6 178

4 authors, including:

Mattia Marini Fabrizio Felletti

University of Milan University of Milan
57 PUBLICATIONS   275 CITATIONS    108 PUBLICATIONS   1,242 CITATIONS   

SEE PROFILE SEE PROFILE

Giovanni Pietro Beretta

University of Milan
48 PUBLICATIONS   332 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computational Ichnology View project

Late Palaeozoic-early Mesozoic continental palaeontology and stratigraphy from Italy View project

All content following this page was uploaded by Mattia Marini on 26 June 2018.

The user has requested enhancement of the downloaded file.

 water
Article
Three Geostatistical Methods for Hydrofacies
Simulation Ranked Using a Large Borehole Lithology
Dataset from the Venice Hinterland (NE Italy)
Mattia Marini * ID
, Fabrizio Felletti ID
, Giovanni Pietro Beretta and Jacopo Terrenghi
Earth Science Department ‘Ardito Desio’, University of Milan, 20133 Milan, Italy; felletti.fabrizio@unimi.it (F.F.);
giovanni.beretta@unimi.it (G.P.B.); Jacopo.terrenghi@unimi.it (J.T.)
* Correspondence: mattia.marini@unimi.it; Tel.: +39-340-586-5715




Received: 25 April 2018; Accepted: 21 June 2018; Published: 25 June 2018


Abstract: A large borehole lithology dataset from the shallowly buried alluvial aquifer of the Brenta
River Megafan (NE Italy) is used in this paper to model hydrofacies with three classical geostatistical
methods, namely the Object-Based Simulation (OBS), the Sequential Indicator Simulation (SIS), and
the Truncated Gaussian Simulation (TGS), and rank alternative output models. Results show that,
though compromising with geological realism and rendering a noisy picture of subsurface geology,
the pixel-based TGS and SIS are better suited than OBS for their ease of conditioning to closely spaced
boreholes, especially in fine-scale simulation grids. In turn, SIS appears to provide better prediction
and less noisy hydrofacies models than TGS without requiring assumptions about relationship
among operative facies, which makes it particularly suited for use with large borehole lithology
datasets lacking detail and quality consistency. Flow simulation on a test volume constrained with
numerous boreholes indicates the SIS hydrofacies models feature well-connected sands forming
relatively fast flow paths as opposed to TGS models, which instead appear to carry a more dispersed
flow. It is shown how such a difference primarily relates to ‘noise’, which in TGS models is so
widespread to translate into a disordered spatial distribution of K and, consequently, a nearly
isotropic simulated flow.

Keywords: alluvial porous aquifer; hydrofacies; geostatistical simulation; borehole lithology database;
aquifer assessment; sand connectedness

1. Introduction
Groundwater flow modelling in alluvial aquifers requires prior reconstruction of hydraulic
conductivity (K) variability. However, because the scale of heterogeneity of alluvial lithofacies can
be much finer than typical inter-borehole spacing [1–5], understanding how K varies in space from
sparse hydraulic tests is inherently impractical. Alternatively, when numerous continuous-core
boreholes are available, one can resort to geostatistics for simulating lithofacies at first step, and then
obtain the K-field by assigning appropriate K to simulated lithofacies [6–9]. Geostatistics deals with
prediction of any property (including discrete variables such as, for example, lithofacies) exhibiting
some degree of spatial continuity [10]. The variable value is estimated at unsampled sites weighting
observed values against some metric of spatial continuity. After building the grid, i.e., discretizing
the model domain into cells (or pixels, i.e., finite elements of fixed size and geometry), and feeding
it with observed values (upscaling of input conditioning data), lithology can be simulated in 3D
with two different geostatistical approaches, namely object-based and pixel-based techniques [11].
Object-Based Simulation (OBS) is suitable when boreholes are sparse but ancillary observations
(e.g., geophysics) provide information on geobodies’ shape and lithofacies [12–14]. OBS populates

Water 2018, 10, 844; doi:10.3390/w10070844 www.mdpi.com/journal/water

Water 2018, 10, 844 2 of 23

the grid by inserting sets of cells with pre-defined shape and facies (i.e., objects) conditionally to
borehole data and facies fractions. Multiple object types mimicking the component elements of a
depositional system (e.g., channels, levees, lateral splays of both fluvial and deep-water systems,
etc.) can be used, defined by a set of numerical descriptors of their 3D geometry (i.e., cross-sectional
and plan-view sizes and shapes, azimuth orientation) borrowed from depositional analogs [15].
Although producing geologically realistic and visually attractive facies results, OBS is inefficient
at honoring dense borehole information [11], which may preclude achieving full conditioning to
input data. Conversely, pixel-based techniques simulate facies cell by cell and fully conditionally to
input data, which makes them best suited where numerous boreholes cross geobodies with complex
geometry. Classical pixel-based algorithms, such as the Sequential Indicator Simulation (SIS) and the
Truncated Gaussian Simulation (TGS), rely on the two-point statistic from variograms as a metric
of facies autocorrelation and use Kriging to estimate the most probable facies [16]. In SIS, operative
facies are transformed into indicator variables (i.e., a variable taking value of 1 if present or 0 if
not present at a given cell of the grid) and kriged individually using facies-specific variograms.
Facies are then assigned to cells weighting kriging results against facies cumulative density functions
(CDF) from input data. Alternatively, TGS assumes facies to be sequentially ordered and converts
them into a continuous random variable with Gaussian distribution prior to kriging. Subsequently,
it assigns facies code to cells by truncating kriged values with thresholds adjusted to local facies
CDF. TGS capability has been recently extended implementing multiple-gaussian random functions
(Pluri-Gaussian Simulation [17]), handling complex facies patterns. Differently from variogram-based
methods, the Multiple-Point Statistics relies on statistics extracted from training images prepared by
the modeler to inform the algorithm on expected geometries and facies patterns [1,2,4]. Although
the design of Pluri-Gaussian and the Multiple-Point Statistics algorithms provide the potential for
better capturing the heterogeneity of some deposits than classical algorithms, their use of fixed facies
transition rules and identity between facies and geobodies requires defining with care operative facies,
which is typically impractical when working with large borehole lithology datasets [18,19], geophysical
logs [20], and penetrometer tests. A drawback of pixel-based methods is that facies realizations may
be prone to short-scale noisy variations (i.e., isolated cells taking outlier facies codes) and poor
reproduction of input facies proportions, which can be overcome post-processing realizations with
cleaning algorithms [21,22]. The method is based on maximizing a-posteriori facies probability, which
is calculated at each cell of the grid weighting facies occurrences within a neighborhood by closeness
to the cell being considered. The case study of this paper is an alluvial aquifer from the shallow
subsurface of the Brenta River Megafan (BRM) of NE Italy, which is particularly valuable for practical
testing of hydrofacies geostatistical modelling because of the availability of about 2000 continuous-core
boreholes [23,24]. In this study, lithology information is used for modelling hydrofacies with three
classical methods, namely the Object-Based Simulation (OBS), the Truncated Gaussian Simulation
(TGS) and the Sequential Indicator Simulation (SIS), and ranking modelling ranked based on how well:
(a) input data are honored by simulation; (b) hydrofacies are predicted at unsampled locations; and (c)
the hydrostratigraphic model from expert judgment is matched. Selected models are then used for
running a particle tracking experiment to show the likely implications on groundwater flow modelling
of simulating hydrofacies with one algorithm in place of another.

2. Geological and Hydrogeological Framework

Located in the Venetian-Friulian Plain of NE Italy (Figure 1a), the study area rests in the distal
sector of the Brenta River Megafan (BRM), one of a series large alluvial fans (megafans sensu [25])
developed during the Würm glaciation (Late Pleistocene) because of an increased fluvial export of
glacial debris from major Alpine valleys.
Water 2018, 10, 844 3 of 23
Water 2018, 10, x FOR PEER REVIEW 3 of 22
Water 2018, 10, x FOR PEER REVIEW 3 of 22

Figure 1. (a) Simplified geomorphological sketch of the Friulan-Venetian Plain outlining the main
Figure 1. (a) Simplified geomorphological sketch of the Friulan-Venetian Plain outlining the main
Figure
alluvial1.fans
(a) of
Simplified geomorphological
late Quaternary sketch
age (modified, of the
from Friulan-Venetian
[25]). Plain(2)outlining
Legend: (1) Brenta; the main
Montebelluna; (3)
alluvial fans of late Quaternary age (modified, from [25]). Legend: (1) Brenta; (2) Montebelluna;
alluvial
Piave, and (4) Adige; (b) Detail of the study area with outline of the ‘Petrolchimico’ and location(3)
fans of late Quaternary age (modified, from [25]). Legend: (1) Brenta; (2) Montebelluna; of
(3) Piave, and (4) Adige; (b) Detail of the study area with outline of the ‘Petrolchimico’ and location of
Piave, andused
boreholes (4) Adige; (b) Detail of
for geostatistical the studyand
modelling area with outline of the ‘Petrolchimico’ and location of
validation.
boreholes used for geostatistical modelling and validation.
boreholes used for geostatistical modelling and validation.
The deglaciation following the Würm Last Glacial Maximum (LGM; 26–19 kyr BP) resulted in a
The deglaciation following the
the Würm Last Glacial Maximum (LGM;
(LGM; 26–19 kyr
kyr BP)
BP) resulted
resulted in aa
dramaticdeglaciation
The decrease of following
sediment Würm
supplyLast to Glacial
megafans Maximum
with consequent 26–19pedogenization in
and
dramatic
dramatic decrease
decrease of sediment
of sediment supply to megafans
supply with consequent pedogenization and development
development of the ‘Caranto’ paleosol [26],toas megafans
well as a swift withsea consequent
level rise and pedogenization
landward shift and
of
of the ‘Caranto’
development ofpaleosol
the [26],
‘Caranto’ as well as
paleosol a swift
[26], sea
as level
well as rise
a and
swift landward
sea level shift
rise of
and coastal depositional
landward shift of
coastal depositional environments. The shallow subsurface stratigraphy of the model area comprises,
environments.
coastal The shallow
depositional subsurface stratigraphy
environments. of the model area comprises, fromarea
topcomprises,
to bottom:
from top to bottom: (i) a rather thin The shallow
(typically subsurface
a few meters) andstratigraphy of the layer
discontinuous model of anthropogenic
(i) a
from rather
top to thin (typically
bottom:a(i)maximuma few meters)
a rather thin and
(typically discontinuous layer of anthropogenic deposits, reaching
deposits, reaching thickness of aca.
few
10meters)
m in the and discontinuouschemical
‘Petrolchimico’ layer of anthropogenic
industrial site
adeposits,
maximum thickness
reaching of
a maximumca. 10 m in the ‘Petrolchimico’ chemical industrial site (Figure 1b) as a
(Figure 1b) as a result of pastthickness of ca. 10 m
land reclamation in the(ii)
works; ‘Petrolchimico’ chemicalpost-LGM
a few meters-thick industrial site
unit,
result
(Figureof 1b)
pastas land result
reclamation works; (ii) a few meters-thick post-LGM unit, comprising the Caranto
comprising the aCaranto of past
paleosol land
and thereclamation
Late glacial works; (ii)
to Holocenea few meters-thick
sediments relatedpost-LGM unit,
to the alluvial
paleosol
comprisingand the Late
the BRM
Caranto glacial to Holocene sediments related to the alluvial network of the BRM and
network of the andpaleosol
the lagoon and ofthe Late and
Venice; glacial
(iii)to
ca.Holocene
10 to oversediments
40 m of LGM related to the
alluvial alluvial
sediments
the lagoon
network of Venice;
of the BRM and (iii)
andeither ca.
the lagoon 10 to over
of Venice; 40 m
and of LGM alluvial
(iii)interstadial
ca. 10 to over sediments of
40 m ofofLGM BRM, which
alluvial overlay
sediments
of BRM, which overlay conformably onto older deposits the Würm glaciation or
either
of BRM,conformably
which ontoeither
overlay older conformably
interstadial deposits
onto of the
older Würm glaciation
interstadial deposits or
ofunconformably
the Würm onto the
glaciation or
unconformably onto the Montebelluna megafan (Figure 2).
Montebelluna
unconformablymegafanonto the(Figure 2).
Montebelluna megafan (Figure 2).

Figure 2. West-looking view of the model domain with top and bottom bounding stratigraphic
Figure
surfaces2. West-looking view of thepenetrating
model domain with top and bottom deposits
bounding stratigraphic
Figure 2. and the available
West-looking viewboreholes the Late
of the model domain with top Glacial Maximum
and bottom of the
bounding stratigraphic Brenta.
surfaces
surfaces and the available boreholes penetrating the Late Glacial Maximum deposits of the Brenta.
and the available boreholes penetrating the Late Glacial Maximum deposits of the Brenta.
Because LGM deposits were not buried by younger deposits over most of the Venice’s
BecausetheLGM
hinterland, deposits
present-day were not
landscape buried
of the studybyarea
younger deposits
represents over
the relic most of the
morphology Venice’s
of BRM. This
hinterland, the present-day landscape of the study area represents the relic morphology
consists of a plain with topographic gradients in the range 0.4–2‰ crossed by a series ofof BRM. This
consists of a plain with topographic gradients in the range 0.4–2‰ crossed by a series of
Water 2018, 10, 844 4 of 23

Because LGM deposits were not buried by younger deposits over most of the Venice’s hinterland,
theWater 2018, 10, x FOR
present-day PEER REVIEW
landscape of the study area represents the relic morphology of BRM. This consists4 of 22 of

a plain with topographic gradients in the range 0.4–2‰ crossed by a series of morphological ridges
morphological ridges with azimuth orientations in the range 100–150° and widths of a few hundreds
with azimuth orientations in the range 100–150◦ and widths of a few hundreds of meters, representing
of meters, representing the infill of aggradational fluvial channels developed during the latest LGM.
the infill of aggradational fluvial channels developed during the latest LGM. The depositional model
The depositional model of the studied part of BRM is one of an outer sandy alluvial fan featuring a
of network
the studied part of BRM
of distributary is onechannels
leveed of an outer sandy
(Figure alluvial fan onto
3) ‘wandering’ featuring a network
a mud-prone of distributary
floodplain with
leveed channels (Figure 3)
scattered peatlands [24,25]. ‘wandering’ onto a mud-prone floodplain with scattered peatlands [24,25].

Figure 3. (a) Likely plan-view topology of the aquifer sandbodies in the shallow subsurface of the
Figure (a) (modified,
study3.area Likely plan-view topology
after [27]); of the cross-section
(b) Geological aquifer sandbodies in the
illustrating the shallow subsurface
lithofacies of the
heterogeneity
study area (modified, after [27]); (b) Geological cross-section illustrating the lithofacies
of the Late Pleistocene alluvial deposits addressed in this study (modified, after [24]). heterogeneity
of the Late Pleistocene alluvial deposits addressed in this study (modified, after [24]).
Typically, channel-fill sand bodies have widths (across-stream) of a few hundred of meters and
Typically,
maximum channel-fill
thickness sand
of a few bodies
meters, have
albeit widths (across-stream)
amalgamation of subsequent ofchannels
a few hundred of meters
may result in larger and
compositethickness
maximum sand bodies. By ameters,
of a few hydrostratigraphic standing point,
albeit amalgamation the study channels
of subsequent area is typified by a in
may result
shallow,
larger poroussand
composite aquifer with By
bodies. thickness in the range of 10–20
a hydrostratigraphic standinghosted mainly
point, in the area
the study sand-prone fill of
is typified by a
the BRM
shallow, paleo-channel
porous network.
aquifer with Overall,
thickness such
in the an aquifer
range of 10–20 consists
hostedofmainly
a NW-SE elongated
in the geobody
sand-prone fill of
thewith
BRM anpaleo-channel
estimated volume of 3.5 ×Overall,
network. 108 m3 and average
such hydraulic
an aquifer conductivity
consists of 2 ×elongated
of a NW-SE 10−5 m/s, encased
geobody
in a low-conductivity (in the range of 5 × 10 −7–1 × 10 −8 m/s) silty-clayey
with an estimated volume of 3.5 × 10 m and average hydraulic conductivity of 2 × 10−5 m/s,
8 3 background. Borehole
correlation in densely investigated sites (e.g., the Petrolchimico; Figure 3b) reveals how such
relatively thin aquifer can locally fringe into a multi-storey of smaller string-like sand bodies,
Water 2018, 10, 844 5 of 23

encased in a low-conductivity (in the range of 5 × 10−7 –1 × 10−8 m/s) silty-clayey background.
Borehole correlation
Water 2018, 10, x FOR PEERinREVIEW
densely investigated sites (e.g., the Petrolchimico; Figure 3b) reveals 5how of 22
such relatively thin aquifer can locally fringe into a multi-storey of smaller string-like sand bodies,
corresponding
corresponding to to single
single channel-fills
channel-fills [27].
[27]. Where
Where the
the thickest,
thickest, the
the aquifer
aquifer consist
consist instead
instead of
of aa stack
stack ofof
subsequent amalgamated
subsequent amalgamated channel-fills. Beyond aquifer heterogeneity, understanding groundwater
heterogeneity, understanding groundwater
flow
flow in
in the
the study
study area
area is
is made
made challenging
challenging by by superimposition
superimposition of of coastal
coastal (e.g.,
(e.g., tidal
tidal oscillations
oscillations and
and
fresh/saline waters interface)
fresh/saline waters interface) and anthropogenic
anthropogenic forcings, including an extensive man-made drainage
man-made drainage
network,
network, heavy
heavy groundwater withdrawals,
withdrawals, as well as numerous open wells which result result in
in artificial
artificial
connectedness
connectedness of of sand-prone
sand-prone aquifer
aquifer bodies
bodies [27–30].
[27–30].

3.
3. Materials
Materials and
and Methods
Methods
The
The input
input data
data used
used in
in this
this study
study is borehole lithology
is borehole from the
lithology from Provincia di
the Provincia Venezia subsurface
di Venezia subsurface
database [24]. Since the shallow subsurface of the study area is composed of non-lithified
database [24]. Since the shallow subsurface of the study area is composed of non-lithified sediments, sediments,
the
the database originally comprised a total of 94 different ‘soil’ types classified based on fractions of
database originally comprised a total of 94 different ‘soil’ types classified based on fractions of
dominant and accessory grain size classes in compliance with geotechnical classification
dominant and accessory grain size classes in compliance with geotechnical classification standards standards [31].
Therefore, no other
[31]. Therefore, objective
no other criteria
objective but dominant
criteria graingrain
but dominant size could be adopted
size could be adoptedfor grouping the full
for grouping the
range of soil types into operative facies. Such grouping yielded three main categories,
full range of soil types into operative facies. Such grouping yielded three main categories, namely namely sands,
silts
sands,and clays,
silts andplus peats
clays, plusaspeats
accessory facies (Figure
as accessory 4), whose
facies (Figure 4), character reflects reflects
whose character sedimentation in very
sedimentation
diverse depositional processes and environments (Table
in very diverse depositional processes and environments (Table 1). 1).

Figure 4. Breakdown of operative facies by component soil types named after the sedimentary clastic
Figure 4. Breakdown of operative facies by component soil types named after the sedimentary clastic
rock classification of [32].
rock classification of [32].
Table 1. Dominant depositional process and environment, abundance in the borehole dataset (as
Table 1. Dominant depositional process and environment, abundance in the borehole dataset
percent of cumulative thickness), and hydraulic conductivity (K) of the operative facies used in
(as percent of cumulative thickness), and hydraulic conductivity (K) of the operative facies used
geostatistical modelling. The facies ordering in the table reflects the transition rule adopted in
in geostatistical modelling. The facies ordering in the table reflects the transition rule adopted in
modelling facies with the Truncated Gaussian Simulation.
modelling facies with the Truncated Gaussian Simulation.
Facies Depositional Environment Depositional Process Abundance (%) Mean K (m/s)
sands fluvial channel point bar migration 43.2
Abundance 2.00 × 10−5
Facies Depositional Environment Depositional Process Mean K (m/s)
silts levee overbanking (%)
27.5 5.00 × 10−7
clays
sands floodplain
fluvial channel overbanking
point bar migration 28.4
43.2 1.00
2.00× × −8 −5
1010
silts
peats levee
peatland overbanking
plant debris accumulation 27.5
0.9 5.00× ×
3.50 −7 −7
1010
clays floodplain overbanking 28.4 1.00 × 10−8
peats peatland plant debris accumulation 0.9 3.50 × 10−7
Because operative facies have relatively narrow conductivity ranges (Figure 5a) they also
represent as many hydrofacies.
Because operative
Therefore, facies have
in the following relatively
text narrow
the terms faciesconductivity ranges
and hydrofacies (Figure
will 5a) interchangeably
be used they also represent
to
as many hydrofacies.
refer to the same lithotype associations. Data analysis and geostatistical simulation were
accomplished with Petrel 2014™ by Schlumberger™, run on a PC workstation equipped with a 4-
core 8-thread Intel Core i7-4790 CPU (3.60 GHz), a Nvidia Quadro P2000 GPU and 16 Gb of ram.
From the full dataset of 2153 boreholes penetrating BRM, a subset of 1615 randomly selected
boreholes (ca. 75% of the full dataset; input dataset, hereafter) was used to condition the simulation,
whereas the reminder 538 boreholes (validation dataset, hereafter) were used for validating results.
The simulation grid top was built interpolating the base of the Caranto paleosol, previously identified
in ca. 1000 boreholes [23,24], whereas the grid base was set at a depth of 30 m from ground level. The
intervening volume (17.45 × 27.55 × 0.032 km) was then discretized into 349 × 552 × 164 = 30,871,428
cells with a plan-view mesh size of 50 × 50 m and layering of 0.2 m. Operative facies from both the
input and the validation datasets were then upscaled to the simulation grid by assigning to each of
 facies proportions from the input dataset were assumed to closely reflect the BRM composition and
thus used for informing simulation on probability of drawing facies at any cell of the grid. Among
the facies modelling algorithms provided in Petrel 2014™, the choice was made to use three classical
geostatistical algorithms, namely the Object-Based Simulation (OBS), the Truncated Gaussian
Simulation (TGS), and the Sequential Indicator Simulation (SIS). OBS was accomplished using three
Water 2018, 10,types,
objects 844 namely fluvial channels, levees, and peat accumulations, whose facies composition and6 of 23
shape parameters are detailed in Table 2.

Figure 5. (a) Experimental set up of the particle tracking simulation with location of the injector and
Figure (a) Experimental
the5.receiving set up
wells; (b) Box of (logarithmic
plots the particle tracking simulation
scale) of hydraulic with location
conductivity of the injector
K measurements and the
from
receiving
[26] bywells; (b) Box
operative plots
facies. Box(logarithmic scale) 25th
boundaries indicate of hydraulic
percentileconductivity K measurements
and 75th percentile, from [26]
whiskers indicate
by operative
maximumfacies. Box boundaries
and minimum indicate
values, whereas the25th percentile
line and and
the cross 75ththe
within percentile, whiskers
box indicate indicate
the median
maximum
and theand
mean minimum values, whereas the line and the cross within the box indicate the median
values, respectively.
and the mean values, respectively.
Table 2. Order of insertion, component facies, and numerical descriptors of shape, size, and layout of
the depositional object used in the Object-Based Simulation.
Therefore, in the following text the terms facies and hydrofacies will be used interchangeably
Insertion Shape Size, Layout and Orientation
to refer to the same Order/Replacement
Object Type Facies
lithotype associations. Data analysis and geostatistical simulation were
Min Mean Max Drift 1
Cross-Sectional Plan-View
accomplished with Petrel Rules 2014™ by Schlumberger™, run on a PC workstation (m) equipped (%) with
width - 300 - 1
a 4-core 8-thread Intel Core i7-4790 CPU (3.60 GHz), a Nvidialength Quadro P2000 GPU infinite
and 16 Gb
of ram.fluvial From sands
the full dataset
after peats; of 2153half-pipe
can replace boreholes string-like,
penetrating BRM, thickness a subset 1 of 1615
2.5 5 randomly
0.3
channel peats sinuous orientation (°) - 110 - 0.1
selected boreholes (ca. 75% of the full dataset; input dataset, hereafter)
sinuosity amplitude was - used -to condition
500 0.3
the simulation, whereas the reminder 538 boreholes (validation wave-length -
dataset, hereafter) 2500 were - 0.5
used for
0.63 × channel
width 0.5
validating results. The simulation grid top was built interpolating the base of the Caranto
width paleosol,
in tandem with/on both
length
previously levee identified
silts in ca.
sides 1000 boreholes
of channels; can wedge
string-like,
[23,24], whereas the grid base was set atinfinite a depth of 30 m
sinuous 0.63 × channel
from ground level. The replace peats
intervening volume (17.45 × 27.55 × 0.032 thickness
km) was then 0.5
thick. discretized into
orientation/sinuosity same as channels
349 × 552 × 164 = 30,871,428 cells with a plan-view mesh size of 50 × 50 m and layering of 0.2 m.
last, fills in gaps
Operative faciesclays
background from both between the input and the validation datasets were
previously n/a then upscaled to the simulation
inserted objects
grid by assigning to each of the cells penetrated by at least one borehole width
the facies
100
occurrence
250 2000
closest
-
to
the cell mid-point.
peat
peats
Provided the lack of elliptical
first to be inserted
significant trends
elliptical
in facies distribution, global and vertical
length 100 625 5000 -
accumulation thickness
(i.e., layer by layer) facies proportions from the input dataset were assumed to0.2-closely
orientation (°)
0.5
110
1
reflect
-
-
the
0.1
BRM
composition and thus used for informing simulation on probability of drawing
Note: 1 Drift is a multiplier expressing the tolerance allowed to the algorithm to adapt sizes of inserted facies at any cell of
the grid. Among the facies modelling
objects to conditioning borehole data. algorithms provided in Petrel 2014™, the choice was made to
use three classical geostatistical algorithms, namely the Object-Based Simulation (OBS), the Truncated
GaussianPeats were inserted
Simulation (TGS),first, andfollowed by fluvial
the Sequential channels Simulation
Indicator and associated levees
(SIS). OBS which
waswere let to
accomplished
erode previously inserted objects. Gaps between inserted
using three objects types, namely fluvial channels, levees, and peat accumulations, whose objects were then assigned to clays. OBSfacies
composition and shape parameters are detailed in Table 2.
Water 2018, 10, 844 7 of 23

Table 2. Order of insertion, component facies, and numerical descriptors of shape, size, and layout of the depositional object used in the Object-Based Simulation.

Shape Size, Layout and Orientation

Object Type Insertion Order/Replacement
Facies
Rules Min Mean Max Drift 1
Cross-Sectional Plan-View
(m) (%)
width - 300 - 1
length infinite
thickness 1 2.5 5 0.3
fluvial channel sands after peats; can replace peats half-pipe string-like, sinuous
orientation (◦ ) - 110 - 0.1
sinuosity amplitude - 500 - 0.3
wave-length - 2500 - 0.5
width 0.63 × channel width 0.5
in tandem with/on both sides length infinite
levee silts wedge string-like, sinuous
of channels; can replace peats thickness 0.63 × channel thick. 0.5
orientation/sinuosity same as channels
last, fills in gaps between
background clays n/a
previously inserted objects
width 100 250 2000 -
length 100 625 5000 -
peat accumulation peats first to be inserted elliptical elliptical
thickness 0.2 0.5 1 -
orientation (◦ ) - 110 - 0.1
Note: 1 Drift is a multiplier expressing the tolerance allowed to the algorithm to adapt sizes of inserted objects to conditioning borehole data.
Water 2018, 10, 844 8 of 23

Peats were inserted first, followed by fluvial channels and associated levees which were let to erode
previously inserted objects. Gaps between inserted objects were then assigned to clays. OBS was then
run assigning honoring priority to either the input data (OBS-b) or the objects shape parameters (OBS-g)
of Table 2. While in the former mode the simulation is flexible on using object shape parameters so to
best honor input data, in the latter it uses shape parameters more stringently. Pixel-based modelling
was accomplished using the random function-based version of TGS and the classical GSLIB version of
SIS [16]. The used TGS algorithm was preferred to the sequential version from GSLIB [16] because it
was faster and more accurate in honoring input facies proportions [33]. Both TGS and SIS were run with
standard settings, choosing the ordinary kriging mode (which assumes constant mean in the search
neighborhood only) so to account for local variability in facies distribution within the model domain.
In TGS, operative facies were assumed to laterally pass each into another following the facies ordering
of Table 1 as to reflect deposition in contiguous depositional environments. Because pixel-based
realizations were found to be prone to noise and poor reproduction of input facies proportions, they
were cleaned with the Maximum a Posteriori Selection algorithm (MAPS) by applying a neighborhood
search of only two cells along the three axes (i.e., 100 × 100 m in plan-view and 0.4 along the vertical
axis) and assigning a greater weight to conditioning facies from boreholes [21]. Because OBS was found
to be quite demanding computationally (ca. 27 h for each realization as opposed to a few hours for
pixel-based realizations), a set of only 25 equiprobable realizations was simulated with each algorithm.
Probability of each operative facies was then calculated as the number of favorable outcomes (i.e.,
a cell is assigned to that facies) at each cell of the grid. Calculation of SIS and TGS probability models
was repeated after cleaning realizations with MAPS. Alternative facies models were then ranked
based on the attained degree of conditioning to input boreholes, how closely they reproduce facies
proportions from input data and predict facies at unsampled locations [11]. As proposed by [34],
closeness of prediction of facies f 1–n can be calculated at cells u a where f has been encountered in
validation boreholes as:
Cf − pf
Crel
f = (1)
pf
where p f is the probability (or relative proportion) of facies f at the layer of the grid to which u a
belongs, and:
C f = E{ p(u a ; f )| true = f }, f = 1, . . . , F (2)

where p is the probability calculated from the full realization set. In the ideal case of complete
information about facies in the subsurface, p = 1 at any grid cell. By definition, Crel
f has a lower bound
of −1 and takes negative values where estimated probability is less than relative facies proportions
(p f ). Conversely, where Crel
f takes positive values it expresses some improvement of facies prediction.
The connectedness of alternative facies models was quantified calculating the number of cells populated
with sands physically connected by sharing at least one of their bounding faces (connected sands,
hereafter). Lastly, to see whether hydrofacies models from alternative modelling approaches translated
into significantly different hydraulic conductivity scenarios, a particle tracking experiment was carried
out on a densely investigated test volume of 3.5 × 3.5 × 0.02 km (Figure 5a). The particle tracking test
was run on one facies realizations from each modelling approach by: (i) assigning mean K values to
each facies based on statistics of K measurement compiled from literature (Figure 5b); (ii) simulating
steady-state groundwater flow with MODFLOW [35] by applying a radially symmetric gradient of ca.
1.6 × 10−4 (corresponding to an head difference of 0.4 m between the center and the corners of the
model); and (iii) using MODPATH [36,37] to track time and layer of arrival at the four test volume
corners of particles (200 per layer) inserted at the test volume center (Figure 5a).
Water 2018, 10, 844 9 of 23

Water 2018, 10, x FOR PEER REVIEW 8 of 22

4. Results
4. Results
4.1. Facies Models
4.1. Facies Models
4.1.1. Object-Based Simulation
4.1.1. Object-Based Simulation
It is recalled that OBS was run in two different modes, namely assigning honoring priority to
It is recalled
either borehole that OBS
data (OBS-b) orwas run in two
geometry different modes,
of depositional namely
objects assigning
(OBS-g). honoring
As a result, priority
two to sets
different
either borehole data (OBS-b) or geometry of depositional objects (OBS-g). As a result, two different
of realizations were obtained which, though both quite realistic in terms of layout of fluvial channel
sets of realizations were obtained which, though both quite realistic in terms of layout of fluvial
network (Figure 6a,b), show very dissimilar structures.
channel network (Figure 6a,b), show very dissimilar structures.

Figure 6. Cont.
Water 2018, 10, 844 10 of 23
Water 2018, 10, x FOR PEER REVIEW 9 of 22

Figure 6. Fence diagrams (left) and representative cross-sections (right) through sample facies
Figure 6. Fence
realizations diagrams
obtained with(left) andObject
(a,b) the representative cross-sections
Based Simulation assigning(right) through
honoring prioritysample
to eitherfacies
realizations
boreholeobtained with geometry;
data or object (a,b) the Object
(c,d) theBased Simulation
Truncated Gaussianassigning
Simulationhonoring
prior andpriority to either
after noise
cleaning,
borehole datarespectively; and (e,f) the(c,d)
or object geometry; Sequential Indicator Gaussian
the Truncated Simulation Simulation
prior and after noise
prior andcleaning.
after noise
Vertical
cleaning, exaggerationand
respectively; is 200×.
(e,f) the Sequential Indicator Simulation prior and after noise cleaning.
Vertical exaggeration is 200×.
With this regard, it can be noted how OBS-b tends to populate the grid with numerous channels
with a quite low shape variability and small average sizes, resulting in a way more intricated and
With this regard,
horizontally layeredit facies
can be noted how
structure OBS-btotends
compared OBS-gto(Figure
populate
6a), the
whichgrid with uses
instead numerous channels
less objects
with with
a quite low shape variability and small
a much higher shape variability (Figure 6b). average sizes, resulting in a way more intricated and
horizontally layered facies structure compared to OBS-g (Figure 6a), which instead uses less objects
with 4.1.2.
a much Truncated
higher Gaussian Simulation(Figure 6b).
shape variability
The experimental variogram of the normal-score continuous facies used in TGS was typified by
4.1.2.aTruncated Gaussian
slight horizontal Simulation
anisotropy with azimuth orientation of 110° N, a ratio of horizontal to vertical
short-scale ranges ofvariogram
The experimental ca. 70, andofathe
‘nested’ structure continuous
normal-score which required using
facies usedanin
exponential and a by
TGS was typified
spherical function for fitting the short and the long-scale ranges, respectively
◦ (Table 3).
a slight horizontal anisotropy with azimuth orientation of 110 N, a ratio of horizontal to vertical
short-scale ranges of ca. 70, and a ‘nested’ structure which required using an exponential and a
Table 3. Parameters of the theoretical variograms used as input for the Sequential Indicator
sphericalSimulation
function and
for the
fitting the short
Truncated andSimulation
Gaussian the long-scale ranges,
(last row). respectively
Information reported(Table 3). refer
in brackets
to the second structure (if any) used for fitting the experimental variogram.
Table 3. Parameters of the theoretical variograms used as input for the Sequential Indicator Simulation
Ranges (m)
and the Truncated
Operative Gaussian
Facies Major Simulation (lastFunction
Direction (°) row). Information
Type 1 reported in brackets refer to the second
Sill
Major Minor Vertical
structure (if any) used for fitting
sands
the experimental
105
variogram.
exp (exp) 0.7 (0.3) 215 (4500) 210 (1500) 3.7 (16.4)
silts 180 exp (exp) 0.7 (0.3) 70 (3100) 35 (3000) 2.6 (17.8)
clays Major 110 Function exp (exp) 0.7 (0.3) 50 (3100) Ranges (m) 2.7 (25)
50 (1300)
Operative Facies Type 1 Sill
peats Direction (◦n/a
) exp 1.0 Major500 500
Minor 1.1Vertical
continuous variable 2 110 exp (sph) 0.6 (0.4) 90 (10,700) 30 (10500) 1.3 (9.5)
sands 105 exp (exp) 0.7 (0.3) 215 (4500) 210 (1500) 3.7 (16.4)
Note:
silts exp and sph are
1
180abbreviations
exp for exponential
(exp) 0.7 and
(0.3)spherical
70model
(3100)functions used for fitting
35 (3000) 2.6 (17.8)
the
claysexperimental variogram;
110 2 normal-score
exp (exp) continuous
0.7 (0.3)variable50into which operative
(3100) 50 (1300)facies are
2.7 (25)
peats
converted in TGS. n/a exp 1.0 500 500 1.1
continuous variable 2 110 exp (sph) 0.6 (0.4) 90 (10,700) 30 (10,500) 1.3 (9.5)
Note: 1 exp and sph are abbreviations for exponential and spherical model functions used for fitting the experimental
variogram; 2 normal-score continuous variable into which operative facies are converted in TGS.
Water 2018, 10, 844 11 of 23

On visual
Water 2018, 10,inspection, TGS models lack facies clusters with well-defined shapes and
x FOR PEER REVIEW appear
10 of 22
remarkably noisy (Figure 6c), though simulated facies occur following the adopted transition rule
(right side On visual 6c).
of Figure inspection,
Also, itTGS models lackthat
is noteworthy facies clusters
noise with well-defined
is accompanied with a shapes and appear
poor matching of facies
remarkably noisy (Figures 6c), though simulated facies occur following the adopted transition rule
proportions and, most notably, an underestimation of sands with respect to fraction for input data
(right side of Figure 6c). Also, it is noteworthy that noise is accompanied with a poor matching of
(Figurefacies
7). Though
proportionssmall, such
and, mostbias against
notably, an sands might potentially
underestimation of sands reflect on aquifer
with respect assessment,
to fraction for inputwhich
suggests
datafacies
(Figurerealizations should
7). Though small, suchbebias
cleaned
againstfrom
sandsnoise.
might Mildly cleaning
potentially facies
reflect on realizations
aquifer assessment,results
indeedwhich
in better facies clustering, especially in cross-section (Figure 6d), as well as better
suggests facies realizations should be cleaned from noise. Mildly cleaning facies realizations reproduction
of input proportions (Figure 7). One last observation concerns the highly layered structure
results indeed in better facies clustering, especially in cross-section (Figure 6d), as well as better of TGS
reproduction of input proportions (Figure 7). One last observation concerns
realizations, which likely descend from the long-scale range (of ca. 10 km!) accounting for c. 40% ofthe highly layered
structure sill
the variogram of TGS
(Table realizations,
3). which likely descend from the long-scale range (of ca. 10 km!)
accounting for c. 40% of the variogram sill (Table 3).

Figure 7. (a) Box plots showing the variability of simulated facies fractions from the Object-based
7. (a) Box
FigureSimulation plots
with showing
honoring the assigned
priority variability of simulated
to either facies
borehole data fractions
(OBS-b) from
or object the Object-based
geometry (OBS-
Simulation
g), thewith honoring
Truncated priority
Gaussian assigned
Simulation to either
(TGS) and theborehole data
Sequential (OBS-b)
Indicator or object(SIS).
Simulation geometry (OBS-g),
For TGS
the Truncated Gaussian
and SIS, fractions Simulation
before and after(TGS) and the are
noise cleaning Sequential Indicator
plotted for comparison;Simulation (SIS).
(b) Vertical For TGS and
proportion
curves ofbefore
SIS, fractions simulated
andfacies
after from
noisesample realizations.
cleaning In both
are plotted for graph sets, the (b)
comparison; boldVertical
dashed proportion
line indicatescurves
facies fractions
of simulated from input
facies from sample borehole data.
realizations. In both graph sets, the bold dashed line indicates facies
fractions from input borehole data.
Water 2018, 10, 844 12 of 23

4.1.3. Sequential Indicator Simulation

Apart from peats, for which input data were too few for computation of accurate variograms,
fitting of the experimental variograms of the reminder facies required using two nested exponential
functions for fitting the short-scale and long-scale. While theoretical variograms of both sands and
clays clearly show a WNW-ESE-oriented horizontal anisotropy (Table 3), that of silts have a very slight
horizontal anisotropy. In addition, ranges from theoretical variograms indicate sands have much
higher lateral correlatability than other facies. Lastly, the ratio of horizontal to vertical short-scale
ranges varies from ca. 20 for clays to a maximum of ca. 50 for sands, suggesting a relatively high
degree of facies layering, though smaller than seen earlier for the normal-score variable adopted in
TGS (Table 3). Using these variograms results in relatively large clusters of cells populated with sands
(Figure 6e) which, despite being patchy-looking in plan-view, tends to be elongated in a N-S direction,
thereby imparting a likewise anisotropy to the reminder facies belts. Especially on cross-sections, it is
apparent that SIS tends to render less fuzzy and noisy facies clusters compared to TGS (cfr. Figure 6c,e),
which at densely investigated sites (e.g., the Petrolchimico) results in faithful reproduction (compare
Figure 3b with sections on left-hand side of Figure 6e) of interpreted lithofacies distributions [26].
Cleaning SIS realizations from noise with the same settings used for TGS results in crispier facies
cluster’s boundaries (Figure 6f) with very little impact on facies proportions (Figure 7).

4.2. Conditioning to Input Data and Facies Prediction at Validation Boreholes

In modelling facies with OBS, changing the honoring priority setting had a relatively small effect
on the actual conditioning to input data, with a maximum of ca. 55% of input cells honored by OBS-b as
opposed to ca. 45% of OBS-g. After peats, which were nearly always honored by simulation, the sands
(i.e., fluvial channel-fills) represent the best honored facies, followed by silts (i.e., levees associated to
channels). In OBS-b, this is achieved by inserting channel and levee objects that show much less shape
variability and are smaller and more numerous than those inserted by OBS-g (Table 4).

Table 4. Report of modelling parameters by object type, including fraction of conditioning cells honored
in the simulation, of two sample facies realizations of BRM from the Object-based Simulation with
honoring priority assigned to either borehole data (OBS-b) or object geometry (OBS-g). Values in
brackets are standard deviations.

Object Type OBS-b OBS-g

honored cells (%) 50 45
number of inserted items 2525 1002
fluvial channels
width 125 (5) 300 (0)
thickness 1.15 (0.3) 2.83 (0.81)
honored cells (%) 35 33
number of inserted items 2525 1002
levees
width 75 (3.23) 189 (0)
thickness 0.72 (0.04) 1.15 (0)
honored cells (%) 100 99
number of inserted items 957 931
peats
maximum length 967 (401) 906 (423)
minimum length 645 (399) 608 (411)

The shape size and variability of the objects inserted can be viewed to reflect the fact that when
the honoring priority is assigned to borehole data, the algorithm will preferentially insert small
objects because easier to fit to conditioning cells and facies proportions. Another observation is that,
independently from the honoring priority settings, OBS models do not replicate satisfactorily either
global facies fractions (e.g., up to ca. 4% departure from target fractions for silts) or vertical trends
(Figure 7). As a result, there is generally a poor reproduction of input facies proportions. Conversely,
Water 2018, 10, 844 13 of 23

Water 2018, 10, x FOR PEER REVIEW 12 of 22

because of the way pixel-based work, all realizations from SIS and TGS fully match facies from
becauseHowever,
boreholes. of the wayinputpixel-based
facies work, all realizations
proportions from SISsatisfactorily
were matched and TGS fullyonlymatch
by facies from
SIS realizations
boreholes. However, input facies proportions were matched satisfactorily only by SIS
(departures <1% for all facies), whereas a systematic underestimation of sands (up to ca. 5%) in favor ofrealizations
(departures <1% for all facies), whereas a systematic underestimation of sands (up to ca. 5%) in favor
silts and clays affects TGS realizations (Figure 7). Mildly cleaning TGS realizations from noise slightly
of silts and clays affects TGS realizations (Figure 7). Mildly cleaning TGS realizations from noise
improves simulated sand fractions, mostly at the expenses of silty and clays, while very little changes
slightly improves simulated sand fractions, mostly at the expenses of silty and clays, while very little
are to be seen
changes areafter
to becleaning
seen afterSIS realizations
cleaning (right-hand
SIS realizations box plots
(right-hand in Figure
box plots 7). 7).
in Figure Closeness
Closenessofoffacies
prediction (C rel ) at cells crossed by validation boreholes (Figure 8) provides a mean for quantitatively
facies prediction
f ( ) at cells crossed by validation boreholes (Figure 8) provides a mean for
ranking modellingranking
quantitatively resultsmodelling
based onresults
fairness of geostatistic
based on fairness ofestimation.
geostatistic estimation.

Figure 8. Box plots of values of closeness of prediction of sands (top), silts (middle), and clays
Figure 8. Boxfrom
(bottom) plotsthe
of alternative
values of closeness of prediction
facies simulation of sands
approaches (top),insilts
adopted this(middle),
study. Forand claysTGS,
SIS and (bottom)
fromvalues
the alternative facies simulation approaches adopted
calculated after noise cleaning are reported for comparison.in this study. For SIS and TGS, values
calculated after noise cleaning are reported for comparison.
Since the validation cells for the facies peats where too few for a sound assessment, only the
three most abundant facies are considered here. It is recalled that, following Equation (2), values
Since the validation cells for the facies peats where too few for a sound assessment, only the three
greater than 0 signifies that geostatistical prediction does better than simply predicting byrel abundance
most abundant facies are considered here. It is recalled that, following Equation (2), C f values greater
(and thus probability) of each facies in the input dataset. Therefore, in the box plots Figure 8
than 0 signifies that geostatistical prediction does better than simply predicting by abundance (and
distributions of with positive median indicate that the used modelling approach provided some
thusprediction
probability) of each facies
improvement in atinleast
the 50%
inputofdataset.
the cellsTherefore,
where f hasinbeenthe box plots Figure
encountered 8 distributions
in a validation
of Crel with
borehole
f positive median indicate that the used modelling approach provided some
and the greater the 25th percentile, the median, and the 75th percentile, the better the overall prediction
improvement in at least
facies prediction. 50%are
If sands of the cells where
considered f has
first, it been
can be encountered
noted how all thein a validation
used modellingborehole
approachesand the
greater the 25th percentile, the median, and the 75th percentile, the better the overall facies prediction.
If sands are considered first, it can be noted how all the used modelling approaches (Figure 8a) yield
Water 2018, 10, 844 14 of 23

some prediction
Water 2018, 10, ximprovement
FOR PEER REVIEWin at least of ca. 50% of the validation cells, with SIS providing 13 of 22 by
far the best results with a median of 0.3. Prediction is even better in the case of clays (Figure 8c),
(Figure 8a) yield some prediction improvement in at least of ca. 50% of the validation cells, with SIS
with SIS providing again the best results with a median of 0.9. Conversely, the ability to predict silts
providing by far the best results with a median of 0.3. Prediction is even better in the case of clays
resulted to be much worse, with OBS and TGS populating validation cells with other facies in most of
(Figure 8c), with SIS providing again the best results with a median of 0.9. Conversely, the ability to
rel
the cases (e.g.,
predict median
silts resultedis to
negative)
be much and SISwith
worse, returning
OBS and a positive yet very
TGS populating small median
validation cells with otherof C f .
value
Finally,facies
oneinlast
mostobservation regards
of the cases (e.g., theissensitivity
median of prediction
negative) and SIS returningto cleaning
a positive yetfrom
very noise pixel-based
small median
realization,
value which
of .results
Finally, in a last
one weak to fair improvement
observation for someofof
regards the sensitivity the facies
prediction to (e.g., sands
cleaning fromandnoiseclays in
pixel-based realization, which results in a weak to fair improvement for some of
TGS and clays in SIS), counterbalanced by worsening for some others (e.g., silts in TGS and sands andthe facies (e.g., sands
silts inand clays
SIS). Whilein TGS and clays
in TGS theseinchanges
SIS), counterbalanced
correlate with bylikewise
worsening for someofothers
changes (e.g., silts
simulated in TGS
facies fractions,
and sands and silts in SIS). While in TGS these changes correlate with likewise changes of simulated
which is explained by the fact that MAPS tends to disfavor the least abundant facies [21], in SIS no
facies fractions, which is explained by the fact that MAPS tends to disfavor the least abundant facies
such correlation is seen, possibly because of the too subtle changes of simulated facies fractions after
[21], in SIS no such correlation is seen, possibly because of the too subtle changes of simulated facies
noise cleaning.
fractions after noise cleaning.

4.3. Sand
4.3.Probability Models
Sand Probability Models
Results fromfrom
Results the alternative facies
the alternative faciesmodelling
modelling approaches adopted
approaches adopted in this
in this study
study can
can be be ranked
ranked
based based on ability
on ability to depict
to depict the the extent
extent andlikely
and likelytopology
topology of ofhighly
highlyconductive geobodies.
conductive This is
geobodies. done
This is done
here focusing on sands only, since with a K of at least two orders of magnitude greater
here focusing on sands only, since with a K of at least two orders of magnitude greater than other facies than other
facies
it might hostitmost
mightofhost
the most of the groundwater
groundwater flow. 9,
flow. In Figure In the
Figure 9, the boundary
boundary of the
of the main main sandbody
aquifer aquifer in
sandbody in the shallow subsurface [27] is overlain onto alternative sand probability maps for
the shallow subsurface [27] is overlain onto alternative sand probability maps for comparison.
comparison.

Figure 9. Sand probability maps at a depth of 10 m below ground level calculated from the Object-
9. Sand
FigureBased probability
Simulation withmaps at a priority
honoring depth ofassigned
10 m belowto (a)ground
boreholelevel
datacalculated from
and (b) object the Object-Based
geometry, the
Simulation with Gaussian
Truncated honoringSimulation
priority assigned
prior (c) to
and(a)after
borehole data
(d) noise and (b)and
cleaning, object
the geometry, the Truncated
Sequential Indicator
Gaussian Simulation
Simulation prior
prior (c) and
(e) and afterafter (d) noise
(f) noise cleaning,
cleaning. and thelines
The dashed Sequential
are the Indicator
boundariesSimulation prior (e)
of the main
aquifer
and after sandbody
(f) noise from The
cleaning. [27].dashed lines are the boundaries of the main aquifer sandbody from [27].
Water 2018, 10, 844 15 of 23

Disregarding the applied honoring priority, maps from OBS (Figure 9a,b) are typified
Water 2018, 10, x FOR PEER REVIEW 14 of 22
by a
relatively low overall probability, indicating a considerable uncertainty in locating sands. Yet, assigning
the honoring priority tothe
Disregarding object geometry
applied honoringresults at least
priority, in some
maps from clustering
OBS (Figure of 9a,b)
high-range valuesbyparallel
are typified a
relatively
to axis and within low the
overall probability,
boundaries indicating
of the a considerable
aquifer sandbody uncertainty
(Figure in locating
9b). Although it cansands. Yet,
be speculated
that aassigning the honoring
greater number priority to object
of equiprobable geometry
realizations results
might at least
have in some
resulted clustering
in some of high-range
improvement of sand
values parallel to axis and within the boundaries of the aquifer sandbody (Figure
probability models, most of the uncertainty associated to OBS facies models is viewed here to reflect 9b). Although it the
can be speculated that a greater number of equiprobable realizations might have resulted in some
too low degree of borehole conditioning attained with the current simulation set up (see Section 4.2).
improvement of sand probability models, most of the uncertainty associated to OBS facies models is
Indeed, because different subsets of only about one half of the available input cells could be honored in
viewed here to reflect the too low degree of borehole conditioning attained with the current
each simulation
simulation set run up(see
(seeSection
Section 4.2),
4.2). OBS is left
Indeed, with different
because too much degree
subsets of of
onlyfreedom for drawing
about one half of thefacies
at each
available input cells could be honored in each simulation run (see Section 4.2), OBS is left performance
cell, resulting in significant variability across different realizations. The better with too
of themuch
tested pixel-based
degree of freedomalgorithms
for drawing(Figure
facies9c,e respectively)
at each overinOBS
cell, resulting is suggested
significant by most
variability acrossof the
high-end values
different of the sand
realizations. Theprobability map from
better performance of TGS and SIS
the tested clustering
pixel-based within the
algorithms outline
(Figure 9c,eof the
mainrespectively)
sandbody. over Also,OBS it isisnoteworthy
suggested by thatmostnoise
of the cleaning
high-end values
reducesof the sand probability
considerably map from in
the uncertainty
TGS probability
the sand and SIS clustering
modelwithin
from TGSthe outline
(Figure of 9d),
the main sandbody.
whereas it has Also,
very it is noteworthy
little impact onthat thatnoise
from SIS
cleaning reduces considerably the uncertainty in the sand probability
(Figure 9f). Similar conclusions can be drawn from cross-sectional views (Figures 10 and 11) model from TGS (Figure 9d),
which
whereas it has very little impact on that from SIS (Figure 9f). Similar conclusions can be drawn from
suggests OBS provides a fuzzy picture of sand distribution, whereas pixel-based algorithms were
cross-sectional views (Figures 10 and 11) which suggests OBS provides a fuzzy picture of sand
capable at replicating the topology of sandbodies proposed by [27], even in area with only a few
distribution, whereas pixel-based algorithms were capable at replicating the topology of sandbodies
conditioning
proposedboreholes
by [27], even (Figure
in area11b,c).
with only a few conditioning boreholes (Figure 11b,c).

Figure 10. Cross-section (see Figure 3 for location and comparison with the geological cross-section
Figure Cross-section
10. [24])
after through sand(see probability
Figure 3 formodels
location and comparison
calculated from the with the geological
Object-Based cross-section
Simulation with
after [24]) through
honoring sandassigned
priority probability
to (a)models
boreholecalculated
data andfrom the Object-Based
(b) object geometry, theSimulation
Truncatedwith honoring
Gaussian
priority assigned
Simulation to (c)
prior (a)and
borehole
after (d)data
noiseand (b) object
cleaning, geometry,
and the Sequentialthe Truncated
Indicator Gaussian
Simulation Simulation
prior (e) and
prior after
(c) and aftercleaning.
(f) noise (d) noise cleaning, and the Sequential Indicator Simulation prior (e) and after (f)
noise cleaning.
Water 2018, 10, 844 16 of 23

Water 2018, 10, x FOR PEER REVIEW 15 of 22

Figure 11. Fence diagrams (see Figure 3a for location) through (a) the hydrostratigraphic model from
Figure[27] Fence
11.and sand diagrams
probability (see
modelsFigure 3a for
from the location)Simulation
Object-Based through with
(a) the hydrostratigraphic
honoring priority assignedmodel
from [27] and sand probability models from the Object-Based Simulation with honoring
to either (b) borehole data or (c) object geometry, and (d) the Truncated Gaussian Simulation priority
and (e)
assigned
the to either (b)
Sequential borehole
Indicator data or after
Simulation (c) object geometry, and (d) the Truncated Gaussian Simulation
noise cleaning.
and (e) the Sequential Indicator Simulation after noise cleaning.
Yet, the TGS probability models depict a remarkably smoother and stratified subsurface
structure than expected in a channelized alluvial fan such as BRM (Figure 11d,e).
Yet, the TGS probability models depict a remarkably smoother and stratified subsurface structure
than expected in a channelized
4.4. Connectivity alluvial fan such as BRM (Figure 11d,e).
of Modelled Sands
Calculation of connected sand volumes indicates that, irrespective of the algorithm used for
4.4. Connectivity of Modelled Sands
facies simulation, there is one main sand geobody (the model aquifer, hereafter) with volumes in the
range of 4.3–5.5
Calculation × 109 m3, accompanied
of connected with several
sand volumes sandbodies
indicates smaller by four
that, irrespective of orders of magnitude
the algorithm used for
facies simulation, there is one main sand geobody (the model aquifer, hereafter) with volumesa in the
(Figure 12a). Variability of this estimate is generally low across equiprobable realizations, reaching
maximum of 6%9 in 3the TGS realization set (Figure 12b), whereas it can be relatively large across
range of 4.3–5.5 × 10 m , accompanied with several sandbodies smaller by four orders of magnitude
alternative facies modelling approaches. In fact, TGS returns model aquifer volumes smaller (by at
(Figure 12a). Variability of this estimate is generally low across equiprobable realizations, reaching
a maximum of 6% in the TGS realization set (Figure 12b), whereas it can be relatively large across
alternative facies modelling approaches. In fact, TGS returns model aquifer volumes smaller (by at
least 5.5 × 108 m3 ) than obtained with SIS (Figure 12b), which is not surprising provided that the
former algorithm systematically underestimates the sand fraction (see Sections 4.1 and 4.2).
Water 2018, 10, x FOR PEER REVIEW 16 of 22

least2018,
Water 5.5 ×10,10 8 m3)
844than obtained with SIS (Figure 12b), which is not surprising provided that
17 ofthe
23
former algorithm systematically underestimates the sand fraction (see Sections 4.1–4.2).

Figure 12. (a) Bar chart of volumes (logarithmic scale, base = 100) of the ten largest geobodies
Figure 12. (a) Bar chart of volumes (logarithmic scale, base = 100) of the ten largest geobodies
comprised of connected sands from the Object-based Simulation with honoring priority assigned to
comprised of connected sands from the Object-based Simulation with honoring priority assigned
either borehole data (OBS-b) or object geometry (OBS-g), the Truncated Gaussian Simulation (TGS),
to either borehole data (OBS-b) or object geometry (OBS-g), the Truncated Gaussian Simulation (TGS),
and the Sequential Indicator Simulation (SIS); Box plots of (b) volumes of the largest geobody of
and the Sequential Indicator Simulation (SIS); Box plots of (b) volumes of the largest geobody of
connected sands (i.e., the model aquifer) and (c) ratio of model aquifer volume to total sand volume
connected sands (i.e., the model aquifer) and (c) ratio of model aquifer volume to total sand volume
(i.e., sand interconnectedness, see text for explanation) calculated from the full sets of facies
(i.e., sand interconnectedness, see text for explanation) calculated from the full sets of facies realization.
realization.

Since OBS was unsuccessful at honoring input data and the least successful in predicting facies,
the connectivity structure resulting from using this method is considered here as highly unrealistic and
Water 2018, 10, 844 18 of 23

Water 2018, 10, x FOR PEER REVIEW 17 of 22

thus not discussed further. On the other hand, statistics of model aquifer volumes from pixel-based
Since OBS was unsuccessful at honoring input data and the least successful in predicting facies,
facies the
models are interesting
connectivity in that they
structure resulting fromillustrate
using this how
method estimates obtained
is considered simulating
here as hydrofacies
highly unrealistic
mightanddepend notdiscussed
thus not only on further.
fractionOnbutthealso onhand,
other degree of connectedness
statistics of simulated
of model aquifer volumes fromhigh-K
pixel- facies.
In fact, normalizing
based aquifer
facies models are volumes
interestingbyin percent
that theyofillustrate
total sand howinestimates
each simulation
obtained highlights
simulating how
hydrofacies
the lower might
estimate depend aquifer
of model not only volume
on fraction but TGS
from also on degreetoofaconnectedness
relates of simulated
noticeable difference of sand
high-K facies. In fact, normalizing aquifer volumes by percent of total sand
interconnectedness between TGS and SIS realizations (Figure 12c). Also, it can be seen how noise in each simulation
highlights how the lower estimate of model aquifer volume from TGS relates to a noticeable
cleaning results in an increase of both model aquifer volumes and sand interconnectedness, especially
difference of sand interconnectedness between TGS and SIS realizations (Figure 12c). Also, it can be
when applied to TGS realizations.
seen how noise cleaning results in an increase of both model aquifer volumes and sand
interconnectedness, especially when applied to TGS realizations.
4.5. Particle Tracking Results
4.5. Particle
Because of Tracking
the poor Results
results obtained with OBS, the particle tracking experiment was run on
sample hydrofacies realizations
Because of the poor resultsfrom TGS and
obtained with SIS
OBS,only, before tracking
the particle and after noise cleaning.
experiment was run onPlotting
sample
particle hydrofacies
injection realizations
vs. arrival layer at from
the TGS
fourand SIS only,
corners before
of the test and after (Figure
volume noise cleaning. Plotting how,
13) highlights
particle
especially injection
before noisevs.cleaning,
arrival layer
the at the example
TGS four corners of the test
is typified byvolume
tracked(Figure 13) that
arrivals highlights how,
tend to be more
especially before noise cleaning, the TGS example is typified by tracked arrivals that tend
evenly distributed vertically (i.e., by layer or depth) and correlated to layer of injection (left side ofto be more
evenly distributed vertically (i.e., by layer or depth) and correlated to layer of injection (left side of
Figure 13a) compared to the SIS example (left side of Figure 13b), indicating a greater flow channeling
Figure 13a) compared to the SIS example (left side of Figure 13b), indicating a greater flow channeling
in the in
latter.
the latter.

Figure 13. Scatter plots of injection vs. arrival layer at the four cardinal corners (N, E, S, and W,
13. Scatter
Figureclockwise fromplots of of
top left injection
each pane;vs. see
arrival
Figurelayer
5a) ofatthe
thetest
four cardinal
volume from corners (N,tracking
the particle E, S, and W,
clockwise from top
experiment leftsample
run on of each pane;
facies see Figure
realization 5a)the
from (a) ofTruncated
the test volume
Gaussianfrom the particle
Simulation tracking
and (b) the
Sequential Indicator Simulation prior (left) and after (right) noise cleaning. The
experiment run on sample facies realization from (a) the Truncated Gaussian Simulation and (b) thebackground color (see
legendIndicator
Sequential at the bottom left) indicates
Simulation priorthe(left)
cumulative number
and after of arrivals
(right) noise averaged
cleaning.overThefive layers. R2 is color
background
the Pearson correlation coefficient for linear regression fits passing through the origin.
(see legend at the bottom left) indicates the cumulative number of arrivals averaged over five layers.
2
R is the Pearson correlation coefficient for linear regression fits passing through the origin.
Water 2018, 10, 844 19 of 23
Water 2018, 10, x FOR PEER REVIEW 18 of 22

Reiterating the experiment after noise cleaning (right-hand side of Figure 13a,b) results in turn
more similar
in a more similar flow
flow behavior
behavior across
across the
the two
two alternative
alternative facies models, though a greater flow
anisotropy is still apparent in the SIS example. These results are interpreted here to reflect the highly
disordered spatial variability of K resulting from widespread noise in raw TGS hydrofacies models,
which ultimately
which ultimatelytranslates
translates into
into a more
a more isotropic
isotropic flow flow than expected
than expected in an alluvial
in an alluvial aquifer
aquifer with with
laterally
laterally extensive
extensive channel-fillchannel-fill
sandbody. sandbody. An appreciable
An appreciable anisotropyanisotropy of the connectivity
of the connectivity structurestructure of
of the SIS
the SISismodel
model is also highlighted
also highlighted by time ofbyparticle
time ofarrivals
particlewhich
arrivals whichpreferential
indicate indicate preferential and along
and faster flow faster
flowN-S
the along the N-S
direction direction
(Figure 14b), (Figure
roughly14b), roughly
parallel parallel
to local to local
orientation orientation
of the of thenetwork.
BRM channel BRM channel
Even
network.
after Evenhydrofacies
cleaning after cleaning hydrofacies
realization from realization
noise, time from noise,arrivals
of particle time ofindicate
particle SIS
arrivals indicate
and TGS SIS
models
andassociated
are TGS models withare associated
alternative with alternative
connectivity connectivity
structures (right side structures (right
of Figure 14), side of that
suggesting Figure
even14),
at
suggesting
site that evenconditioning
with numerous at site with numerous conditioning
boreholes the boreholes might
choice of algorithm the choice of algorithm
be critical might be
to groundwater
critical
flow to groundwater flow assessment.
assessment.

Figure 14. Number vs. time of arrival of tracked particles at the four cardinal corners of the facies
Figure
models 14. Number
obtained vs. (a)time
using the of arrival ofGaussian
Truncated tracked particles at and
Simulation the four cardinal
the (b) corners
Sequential of the
Indicator
facies models obtained using (a) the Truncated Gaussian
Simulation, prior to (left) and after noise cleaning (right). Simulation and the (b) Sequential Indicator
Simulation, prior to (left) and after noise cleaning (right).
5. Discussion
5. Discussion
5.1. Which Modelling Algorithm Does Better with Lithology from Dense Borehole Data?
5.1. Which Modelling Algorithm Does Better with Lithology from Dense Borehole Data?
Results of this study confirm some of the well-known limitations of classical geostatistical
Results of this study confirm some of the well-known limitations of classical geostatistical algorithms
algorithms in replicating the lithofacies heterogeneity of channelized alluvial sediments [1,2,11–
in replicating the lithofacies heterogeneity of channelized alluvial sediments [1,2,11–14,37,38]. However, it
14,37,38]. However, it was possible to rank the tested algorithms based on how closely they honored
was possible to rank the tested algorithms based on how closely they honored input data and predicted
input data and predicted facies at validation boreholes [11], as well as they replicated the current
facies at validation boreholes [11], as well as they replicated the current hydrostratigraphic model of
hydrostratigraphic model of the study area [26]. Even though resulting in geologically realistic facies
the study area [26]. Even though resulting in geologically realistic facies distributions, the Object-Based
distributions, the Object-Based Simulation (OBS) of this study failed to achieve full conditioning to
Simulation (OBS) of this study failed to achieve full conditioning to boreholes, as well as to reproduce
boreholes, as well as to reproduce global and layer-by-layer facies proportions. This reflects the
global and layer-by-layer facies proportions. This reflects the unease of OBS at honoring dense input data
unease of OBS at honoring dense input data by inserting user-defined objects and translated into
by inserting user-defined objects and translated into much variability across equiprobable realizations
much variability across equiprobable realizations and, ultimately, poor facies prediction. While it
and, ultimately, poor facies prediction. While it must be acknowledged that using a coarser grid layering
must be acknowledged that using a coarser grid layering (e.g., 0.5 m, corresponding to less than a
half of the conditioning cells used here) might have allowed to fully honor input boreholes (albeit at
Water 2018, 10, 844 20 of 23

(e.g., 0.5 m, corresponding to less than a half of the conditioning cells used here) might have allowed to
fully honor input boreholes (albeit at the cost of less vertical detail), it must be concluded that OBS is not
viable for capturing small-scale heterogeneity in densely investigated sites. By contrast, due to their fully
conditional nature, facies models from the Truncated Gaussian Simulation (TGS) and the Sequential
Indicator Simulation (SIS) do fully honor borehole data but fail to capture the curvilinear shapes typical
of channelized alluvial sediments, thus resulting in a generally poor geological realism, especially of
plan geometries. This has been widely documented in the literature and reflects the inadequateness
of variogram ranges as a metric of spatial correlation of facies belonging to geobodies with complex
3D geometry [37]. However, SIS was able to better reproduce facies proportions and capture the
WNW-ESE anisotropy of the BRM channel-network compared to TGS, as well as was less affected by
presence of spurious cells assigned to outlier facies (i.e., noise). On the other hand, the widespread
noise in TGS models is viewed here to reflect the implementation of a rule of lateral facies transition
which, due to repeated fluvial channeling, is only rarely preserved in BRM and results therefore in
erratic and presumably uncorrected facies estimation. Despite the poor geological realism of individual
facies realizations, the sand probability models computed from TGS and SIS show a good match with
the current knowledge of the BRM shallow aquifer. Similar conclusions can be drawn by closeness of
facies prediction results, which suggests that the geostatistic estimate from the pixel-based methods
does better than simply guessing facies based on their abundance in borehole cores. One concluding
remark should address the borehole lithology data used in this study, which required to group the
observed variety of soil types into a few operative facies of practical use for geostatistical modelling.
Because detail and consistency of core descriptions were highly variable, and boreholes were too
many to allow for a case-by-case sedimentological interpretation, no other criteria but dominant grain
size could be used to define operative facies. As a result, the link between facies and component
depositional elements of BRM can be strong (e.g., channel fills are in most cases represented by sands)
but by no means exclusive (e.g., sands can occur in levees as well, whereas silts and clays can be locally
intercalated within channel-fill sands), potentially undermining all those modelling approaches that
rely on fixed facies transition rules and identity between facies and depositional objects, including
OBS and TGS, as well as the nowadays highly popular Pluri-Gaussian [17] and the Multiple-Point
Statistics [38]. It is therefore concluded that, especially when working with large borehole datasets
where definition of operative facies can be weak, using SIS might still represent a pragmatic (it does not
require any assumption on spatial relationship among facies) yet decent hydrofacies modelling choice.

5.2. Likely Implications for Aquifer and Groundwater Flow Assessment

Given the complexity of the groundwater flow in study site [27–30], a validation of hydrofacies
models against direct hydrogeological observations was not undertaken in this work. Nonetheless,
in Section 4.3 it has been shown that both TGS and SIS were able to replicate with good confidence
the current hydrostratigraphic model of BRM [27], suggesting these methods can be used in similar
contexts for expeditious appraisal of aquifer depth even at sites with boreholes spacing in order of a
few to several hundred meters. In addition, in Sections 4.4 and 4.5, it is shown that hydrofacies models
from TGS and SIS are associated with different degree of sand connectedness and simulated flow
behavior chiefly due to widespread noise in TGS simulation results. Such noise has in fact a twofold
impact on connectivity in that it can both act as a ‘thief’ to the model aquifer volumes, for example
when a consistent fraction of simulated sand occurs as isolated cells in a low-K background, and
considerably increase flow path tortuosity, for example when the model aquifer is speckled with
numerous cells populated with low-K facies. The latter aspect is particularly well expressed by particle
tracking results of Section 4.5, where the widespread noise in the TGS hydrofacies realization translates
into a spatially disordered distribution of K and, ultimately, into a very diffuse flow (i.e., there is low
variability in the number of particles reaching both the different layers and corners of the test volume)
strikingly contrasting with presence in BRM of high-K channel-fill sandbodies. Although it has been
shown that noise cleaning may dampen differences between alternative hydrofacies models, it is
Water 2018, 10, 844 21 of 23

advised that the admissible degree of noise cleaning should be judged based on a sound estimation of
hydrofacies abundance proportions, so as to avoid aquifer volume and connectedness misestimations.

6. Conclusions
Three geostatistical algorithms commonly employed for sedimentary facies and hydrofacies
modelling, namely the Object-Based Simulation (OBS), the Sequential Indicator Simulation (SIS), and
the Truncated Gaussian Simulation (TGS), were tested using a large borehole lithology dataset from
the channelized alluvial deposits of the Brenta River Megafan (BRM, Upper Pleistocene), NE Italy.
The specific objective was to explore algorithm strengths and weaknesses, with special reference to
hydrogeological applications at sites with dense borehole information. Key findings are as follows:
• Though compromising with geological realism of facies clusters shapes, TGS and SIS are better
suited in place of OBS for their ease of conditioning to closely spaced boreholes.
• Pixel-based facies models of this study, especially those from TGS, suffer for ‘noise’ in form
of unwanted isolated cells taking outlier hydrofacies values, which require cleaning for better
assessment of aquifer facies distribution.
• SIS provides better facies prediction and renders a less noisy picture of subsurface geology
than TGS, without requiring assumptions about spatial relationship among operative facies,
which makes it the best suited for use with large borehole lithology datasets lacking detail and
quality consistency.
• Statistics of connected sands and results of the particle tracking simulation indicate TGS and SIS
hydrofacies models are associated with significantly different aquifer connectivity scenarios, e.g.,
a relatively poorly connected aquifer (TGS) typified by diffuse, nearly isotropic flow as opposed
to a better-connected aquifer (SIS) featuring preferential flow paths hosted within the sandy
channel-fills of BRM.
• Differences between the two alternative pixel-based aquifer models are due to widespread noise
in TGS realizations, suggesting that noise cleaning should be considered and implemented with
care before simulating groundwater flow.

Author Contributions: Conceptualization, M.M., F.F., G.P.B.; Methodology, M.M and J.T..; Software, M.M. and J.T.;
Validation, G.P.B., Y.Y. and Z.Z.; Formal Analysis, M.M. and J.T.; Investigation, M.M. and J.T.; Resources, M.M., F.F.,
G.P.B and J.T.; Data Curation, M.M., G.P.B. and J.T.; Writing-Original Draft Preparation, M.M.; Writing-Review &
Editing, M.M., F.F., G.P.B.; Visualization, M.M. and J.T.; Supervision, M.M., F.F., G.P.B.; Project Administration, M.M.
Funding: This research received no external funding.
Acknowledgments: The two anonymous reviewers are thanked for their useful comments. Provincia di Venezia is
warmly acknowledged for providing the dataset used to make this study, along with additional GIS, cartographic,
and bibliographic material. We are also grateful to Schlumberger™ which has supplied an academic license of
Petrel 2014™ to the University of Milan.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Dell’arciprete, D.; Felletti, F.; Bersezio, R. Simulation of fine-scale heterogeneity of meandering river aquifer
analogues: Comparing different approaches. In geoENV VII—Geostatistics for Environmental Applications,
Quantitative Geology and Geostatistics; Atkinson, P., Lloyd, C., Eds.; Springer: Dordrecht, The Netherlands,
2010; Volume 16, pp. 127–137. [CrossRef]
2. Dell’arciprete, D.; Bersezio, R.; Felletti, F.; Giudici, M.; Comunian, A.; Renard, P. Comparison of three geostatistical
methods for hydrofacies simulation: A test on alluvial sediments. Hydrogeol. J. 2012, 20, 299–311. [CrossRef]
3. Colombera, L.; Felletti, F.; Mountney, N.P.; McCaffrey, W.D. A database approach for constraining stochastic
simulations of the sedimentary heterogeneity of fluvial reservoirs. Am. Assoc. Pet. Geol. Bull. 2012, 96, 2143–2166.
[CrossRef]
4. Colombera, L.; Mountney, N.P.; Felletti, F.; McCaffrey, W.D. Models for guiding and ranking well-to-well
correlations of channel bodies in fluvial reservoirs. Am. Assoc. Pet. Geol. Bull. 2014, 98, 1943–1965. [CrossRef]
Water 2018, 10, 844 22 of 23

5. Comunian, A.; De Micheli, L.; Lazzati, C.; Felletti, F.; Giacobbo, F.; Giudici, M.; Bersezio, R. Hierarchical
simulation of aquifer heterogeneity: Implications of different simulation settings on solute-transport
modelling. Hydrogeol. J. 2016, 24, 319–334. [CrossRef]
6. Fogg, G.E.; Noyes, C.D.; Carle, S.F. Geologically based model of heterogeneous hydraulic conductivity in an
alluvial setting. Hydrogeol. J. 1998, 6, 131–143. [CrossRef]
7. Anderson, M.P.; Aiken, J.S.; Webb, E.K.; Mickelson, D.M. Sedimentology and hydrogeology of two braided
stream deposits. Sediment. Geol. 1999, 129, 187–199. [CrossRef]
8. De Marsily, G.; Delay, F.; Gonçalvès, J.; Renard, P.; Teles, V.; Violette, S. Dealing with spatial heterogeneity.
Hydrogeol. J. 2005, 13, 161–183. [CrossRef]
9. Bianchi, M.; Zheng, C. A lithofacies approach for modelling non-Fickian solute transport in a heterogeneous
alluvial aquifer. Water Resour. Res. 2016, 52, 552–565. [CrossRef]
10. Matheron, G. Random Functions and their Application in Geology. In Geostatistics Computer Applications in
the Earth Sciences; Merriam, D.F., Ed.; Springer: Boston, MA, USA, 1970; pp. 79–87.
11. Pyrcz, M.J.; Deutsch, C.V. Geostatistical Reservoir Modelling, 2nd ed.; Oxford University Press: Oxford, UK,
2014; p. 348. ISBN 9780199731442.
12. Seifert, D.; Jensen, J.L. Object and pixel-based reservoir modelling of a braided fluvial reservoir. Math. Geol.
2000, 32, 581–603. [CrossRef]
13. Deutsch, C.V.; Tran, T.T. FLUVSIM: A program for object-based stochastic modelling of fluvial depositional
systems. Comput. Geosci. 2002, 28, 525–535. [CrossRef]
14. Keogh, K.J.; Martinius, A.W.; Osland, R. The development of fluvial stochastic modelling in the Norwegian
oil industry: A historical review, subsurface implementation and future directions. Sediment. Geol. 2007, 202,
249–268. [CrossRef]
15. Pyrcz, M.J.; Boisvert, J.B.; Deutsch, C.V. A library of training images for fluvial and deepwater reservoirs and
associated code. Comput. Geosci. 2008, 34, 542–560. [CrossRef]
16. Deutsch, C.V.; Journel, A.G. GSLIB: Geostatistical Software Library and User's Guide, 2nd ed.; Oxford University
Press: New York, NY, USA, 1998; p. 369.
17. Beucher, H.; Renard, D. Truncated Gaussian and derived methods. C. R. Geosci. 2016, 348, 510–519. [CrossRef]
18. Dethlefsen, F.; Ebert, M.; Dahmke, A. A geological database for parameterization in numerical modelling of
subsurface storage in northern Germany. Environ. Earth Sci. 2014, 71, 2227–2244. [CrossRef]
19. Vázquez-Suñé, E.; Marazuela, M.Á.; Velasco, V.; Diviu, M.; Pérez-Estaún, A.; Álvarez-Marrón, J. A geological
model for the management of subsurface data in the urban environment of Barcelona and surrounding area.
Solid Earth 2017, 7, 1317. [CrossRef]
20. Pham, H.V.; Tsai, F.T.C. Modelling complex aquifer systems: A case study in Baton Rouge, Louisiana (USA).
Hydrogeol. J. 2017, 25, 601–615. [CrossRef]
21. Deutsch, C.V. Cleaning categorical variable (lithofacies) realizations with maximum a-posteriori selection.
Comput. Geosci. 1998, 24, 551–562. [CrossRef]
22. Hong, S.; Deutsch, C.V. Another Look at Realization Cleaning; Centre for Computational Geostatistics Report
12; University of Alberta: Edmonton, AB, Canada, 2010; p. 118.
23. Vitturi, A.; Bondesan, A.; Fontana, A.; Mozzi, P.; Primon, S.; Bassan, V. Geologia. In Atlante Geologico
Della Provincia di Venezia—Note Illustrative, Quarto d’Altino (Venezia); Vitturi, A., Bassan, V., Mazzuccato, A.,
Primon, S., Bondesan, A., Ronchese, F., Zangheri, P., Eds.; Arti Grafiche Venete: Venezia, Italia, 2011;
pp. 333–357. ISBN 9788890720703.
24. Bondesan, A.; Primon, S.; Bassan, V.; Vitturi, A. (Eds.) Le Unità Geologiche Della Provincia di Venezia; Provincia
di Venezia, Servizio Geologico e Difesa del Suolo: Venezia, Italia, 2008.
25. Fontana, A.; Mozzi, P.; Marchetti, M. Alluvial fans and megafans along the southern side of the Alps.
Sediment. Geol. 2014, 301, 150–171. [CrossRef]
26. Donnici, S.; Serandrei-Barbero, R.; Bini, C.; Bonardi, M.; Lezziero, A. The caranto paleosol and its role in the
early urbanization of Venice. Geoarchaeology 2011, 26, 514–543. [CrossRef]
27. Fabri, P.; Zàngheri, P.; Bassan, V.; Fagarazzi, E.; Mazzuccato, A.; Primon, S.; Zogno, C. Sistemi Idrogeologici
Della Provincia di Venezia: Acquiferi Superficiali; Provincia di Venezia, Servizio Geologico, Difesa del Suolo e
Tutela del Territorio: Venezia, Italia, 2013.
28. Carbognin, L.; Rizzetto, F.; Tosi, L.; Teatini, P.; Gasparetto-Stori, G. The salt intrusion in the Venetian lagoon
area: II. The southern basin. Geol. Appl. 2005, 2, 124–229.
 Water 2018, 10, 844 23 of 23

29. Da Lio, C.; Tosi, L.; Zambon, G.; Vianello, A.; Baldin, G.; Lorenzetti, G.; Manfè, G.; Teatini, P. Long-term
groundwater dynamics in the coastal confined aquifers of Venice (Italy). Estuar. Coast. Shelf Sci. 2013, 135,
248–259. [CrossRef]
30. Beretta, G.P.; Terrenghi, J. Groundwater flow in the Venice lagoon and remediation of the Porto Marghera
industrial area (Italy). Hydrogeol. J. 2017, 25, 847–861. [CrossRef]
31. Associazione Geotecnica Italiana. Raccomandazioni Sulla Programmazione ed Esecuzione Delle Indagini
Geotecniche; AGI; SGE: Padova, Italy, 1977.
32. Folk, R.L. Petrology of Sedimentary Rocks; Hemphill Publishing Company: Austin, TX, USA, 1980.
33. Daly, C.; Quental, S.; Novak, D. A faster, more accurate Gaussian simulation. In Proceedings of the
GeoCanada Conference, Calgary, AB, Canada, 10–14 May 2010.
34. Deutsch, C.V. A Short Note on Cross Validation of Facies Simulation Methods; Centre for Computational
Geostatistics Annual Report, Report 1; University of Alberta: Edmonton, AB, Canada, 1998; p. 109.
35. Harbaugh, A.W. MODFLOW-2005, the US Geological Survey Modular Ground-Water Model: The Ground-Water
Flow Process; Department of the Interior, US Geological Survey: Reston, VA, USA, 2005.
36. Pollock, D.W. Semianalytical Computation of Path Lines for Finite-Difference Models. Groundwater 1998, 26,
743–750. [CrossRef]
37. Pollock, D.W. User Guide for MODPATH Version 6—A Particle-Tracking Model for MODFLOW. US Geological Survey
Techniques and Methods 6-A41; Department of the Interior, US Geological Survey: Reston, VA, USA, 2012.
38. Jha, S.K.; Mariethoz, G.; Mathews, G.; Vial, J.; Kelly, B.F. Influence of Alluvial Morphology on Upscaled
Hydraulic Conductivity. Groundwater 2016, 54, 384–393. [CrossRef] [PubMed]

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

View publication stats