Uncertainty Quantification and Reduction in The Characterization of Subsurface Stratigraphy Using Limited Geotechnical Investigation Data PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

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

net/publication/330270504

Uncertainty Quantification and Reduction in the Characterization of


Subsurface Stratigraphy Using Limited Geotechnical Investigation Data

Article  in  Underground Space · January 2019


DOI: 10.1016/j.undsp.2018.10.008

CITATIONS READS

0 112

1 author:

Xiangrong Wang
University of Dayton
15 PUBLICATIONS   67 CITATIONS   

SEE PROFILE

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

External corrosion and integrity assessment of underground pipeline systems based on in-line inspection data, in-situ soil survey and environmental data View project

Bayesian Machine Learning for Geological Modeling and Geophysical Segmentation View project

All content following this page was uploaded by Xiangrong Wang on 25 January 2019.

The user has requested enhancement of the downloaded file.


Available online at www.sciencedirect.com

ScienceDirect
Underground Space xxx (xxxx) xxx
www.elsevier.com/locate/undsp

Uncertainty quantification and reduction in the characterization


of subsurface stratigraphy using limited geotechnical investigation data
Xiangrong Wang
Department of Civil and Environmental Engineering and Engineering Mechanics, The University of Dayton, Dayton, OH 45469-0243, USA

Received 25 June 2018; received in revised form 20 September 2018; accepted 5 October 2018

Abstract

Subsurface stratigraphy is critical to the design, construction, and subsequent performance of geotechnical structures. However, in
practice it is impossible to identify the stratigraphy of a subsurface geological domain with absolute certainty, due to the limitations
imposed by geotechnical investigation techniques and project budgets. This paper presents a subsurface stratigraphic modeling and
uncertainty quantification approach, which is established based on an improved and extended geological modeling technique previously
established by the author and others, for simulating the stratigraphy of both two-dimensional (2D) and three-dimensional (3D) cases
with more complex geological features. Furthermore, this approach provides quantitative evaluation of the amount of stratigraphic
uncertainty in the current interpretation and enables the systematic reduction of stratigraphic uncertainty through the investigation
of additional targeted borehole drilling locations. Illustrative examples, including artificial cases as well as two real cases from existing
geotechnical projects, are presented in this study to demonstrate the use of the proposed analysis approach.
Ó 2019 Tongji University and Tongji University Press. Production and hosting by Elsevier B.V. on behalf of Owner. This is an open access article
under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Site investigation; Stratigraphy; Stochastic modeling; Uncertainty quantification; Borehole drilling

1 Introduction Moreover, it is widely recognized that there is a tradeoff


between the cost invested in geotechnical site investigation
Obtaining an adequate and accurate understanding of and the additional benefits derived from the resulting geo-
subsurface stratigraphy is an essential first task in solving logical information that can be used in subsequent phases
many geotechnical engineering problems. However, of geotechnical design and construction. Therefore, design-
regardless of the extent of field geotechnical investigations ers and site engineers face a challenge in obtaining a suffi-
and the resulting information from borehole logs, in situ ciently accurate description of a site’s stratigraphy at the
testing, and geophysical surveys, it is typically impossible minimum cost (Jaksa et al., 2005). To address this chal-
to make direct and continuous observations across an lenge, there is a need for novel interpretation approaches
entire subsurface domain. Therefore, the inference of sub- that can not only generate possible stratigraphic configura-
surface stratigraphy unavoidably involves various degrees tions but also provide quantification of uncertainty accord-
of uncertainty (Juang, Jiang, & Christopher, 2001; Li, Qi, ing to field observations and suggest further drilling and
et al., 2016; Liao, Druss, Neff, & Brenner, 1996; Wang, sampling locations that can reduce the uncertainty to the
Wang, Liang, & Liu, 2018; Wang, Wang, Liang, Zhu, & greatest extent.
Di, 2018; Wang, Wellmann, Li, Wang, & Liang, 2017). To interpret soil stratification in a subsurface geological
domain with sparsely distributed observational data, sev-
eral approaches have been developed that are based on
E-mail address: xwang5@udayton.edu interpolation methods (Auerbach & Schaeben, 1990;

https://doi.org/10.1016/j.undsp.2018.10.008
2467-9674/Ó 2019 Tongji University and Tongji University Press. Production and hosting by Elsevier B.V. on behalf of Owner.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
2 X. Wang / Underground Space xxx (xxxx) xxx

Blanchin & Chilès, 1993; Calcagno, Chilès, Courrioux, & realizations and simulating common stratigraphic struc-
Guillen, 2008; Chilès, Aug, Guillen, & Lees, 2004; Zhu & tures. In addition, examples are presented to illustrate the
Zhang, 2013). However, most of these methods can only capability of the workflow in generating reasonable realiza-
provide point estimation as the maximum likelihood esti- tions using available borehole logs and geological knowl-
mate (MLE), without taking the uncertainty of modeling edge, as well as guiding geotechnical site investigation
outcomes into consideration. In other words, although processes through uncertainty quantification.
the most likely deterministic stratigraphic configuration This paper is organized in the following manner. In Sec-
can be obtained, its probability and variability are not esti- tion 2, an introduction of the workflow is provided. In Sec-
mated; thus, the credibility or confidence in the accuracy of tion 3, the procedure for generating stratigraphic
such interpretations is not known. Several existing stochas- realizations and quantitatively measuring the associated
tic modeling techniques—including the Markov chain stratigraphic uncertainty is provided. In Section 4, artificial
model (Elfeki & Dekking, 2001; Li, Qi, et al., 2016), the examples are used to illustrate the generation of common
conditional Gaussian random field (Li, Zhang, & Li, stratigraphic structures in accordance with available geo-
2015) and multiple-point geostatistics (Caers & Zhang, logical knowledge. In Section 5, two real-world projects
2004; Hu & Chugunova, 2008; Toftaker & Tjelmeland, are used to demonstrate the capability of the approach
2013)—can provide multiple possible geological modeling for assessing the confidence level or associated risk of the
results. However, few of these techniques manage to quan- interpretation of subsurface stratigraphy, as well as for
tify the uncertainty associated with intrinsic nonstationary guiding geotechnical site investigations. In Section 6, the
stochastic subsurface conditions. In fact, compared with conclusions are presented.
simulating a deterministic geological configuration, quanti-
fying modeling uncertainty imposes higher demands on the 2 Stochastic geological modeling framework
adopted modeling method since such quantification of
variability is meaningful only if each of the stochastic real- 2.1 Overview of the modeling workflow
izations generated by the modeling method is in accor-
dance with the inherent characteristics of the subsurface Unlike common segmentation or recognition problems
geological structures. However, some implicit assumptions that focus on extracting the latent field from a complete
(or strong prior knowledge in some contexts) in the existing observation at all spatial points (e.g., all pixels in an image)
modeling methods, such as the fixed correlation structures in a random field, available information is sparsely located
in Gaussian random fields, the stationary transition proba- in the test field in the geotechnical site investigation pro-
bility matrixes in Markov chain models, and the predefined cess. Therefore, modeling stratigraphy is essentially a task
data templates in training image methods, may impose cer- to reasonably infer the subsurface characteristics (i.e., soil
tain restrictions on the ability to reproduce the inherent labels) of the unobserved regions, according to the avail-
anisotropic and non-stationary nature of geological able site investigation data and the geological knowledge
structures. of the stratigraphic structures. An inferred possible stratig-
Recently, a geological modeling method based on the raphy can be regarded as realistic only if both the trend or
Markov random field (MRF) for simulating complex large-scale characteristics (i.e., the general shape of the ani-
stratigraphic structures has been developed (Li, Wang, sotropic and nonstationary stratigraphic profile) and the
Wang, & Liang, 2016). Specifically, the MRF-based mod- noise or local characteristics (i.e., small patches and fluctu-
eling method and the Markov chain model share the same ations of the local boundary between different soil types)
underlying assumption that the subsurface stratigraphic are reasonably reflected. Specifically, the former is gener-
structures involve Markovian possesses and can be ally derived from the available observations, such as bore-
described using Markov priors. It is worth noting that this hole logs, outcrops and known (or inferenced from
assumption has been widely recognized and persistently geophysical data) stratigraphic dips, whereas the latter is
utilized for modeling subsurface stratigraphy and sedimen- based on the assumption that the stratigraphic or sedimen-
tary layering structures (Carle & Fogg, 1996, 1997, Dai, tary processes are formed by material transitions in space,
Ritzi, & Dominic, 2005, Li, Li, & Shi, 1999; Wang, and a probabilistic description for this phenomenon is trea-
Wang, & Liang, 2017; Wang, Wang, Liang, Zhu, et al., ted as the MRF prior (Wang, Wellmann, et al., 2017).
2018). The difference is that the framework of the MRF Compared with other interpolation methods, random
model provides a more flexible and intuitive way to specify field-based models are more intuitive in providing physical
the Markovian spatial correlation, which enables the reflec- meaning. Since the local and the large-scale characteristics
tion and reproduction of the heterogeneous and nonsta- closely interact with each other, it is difficult to implement
tionary characteristics of subsurface geological structures. a modeling procedure in which both aspects are simultane-
The developed modeling method is illustrated by applying ously satisfied. Therefore, we decompose the stochastic
it to both artificial and real-world examples. For the sake simulation process into two stages. Details are described
of clarity and conciseness, in this work, the previously in the following sections.
developed modeling method will be briefly introduced, with In the first stage, a Monte Carlo simulation (MCS) using
an emphasis on some new extensions for generating 3D the single-side Markov property is employed to generate

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 3

initial stratigraphic configurations, using the information by R. Each element r 2 R and any of its adjacent elements
obtained from borehole logs and known stratigraphic ori- s 2 R are regarded as neighbors, denoted as
entation. To be more specific, the simulation is quite simi- s 2 @r () r 2 @s. Next, a label xr 2 L (representing the
lar to the common interpretation process, where set of all soil/rock types in the modeled geological domain)
boundaries of different soil formations are formed via is associated with each element r 2 R, and we let
sequential random field sampling at one borehole location x ¼ ðxr jr 2 RÞ 2 X represent a subsurface stratigraphic
to start, and are then extended to other borehole locations configuration, where X is the set of all possible configura-
based on prior knowledge of stratigraphic orientation. The tions. Under these definitions, x is said to be an MRF with
only difference is that the stratigraphic uncertainty is spon- respect to a given neighborhood system if the prior joint
taneously considered in the MCS process due to its probability P ðxÞ > 0 for all x 2 X, and the conditional
stochastic nature. During this stage, large-scale characteris- probability are fulfilled with the following term:
tics (i.e., anisotropic layered structures indicated by bore-
P ðxr jxr Þ ¼ P ðxr jx@r Þ ð1Þ
hole logs and the nonstationary stratigraphic orientation
indicated by geophysical test results or knowledge of local We further denote the likelihood of measuring the sim-
geology) are completely considered, whereas the local char- ilarity between an initial configuration xi and an updated
acteristics governed by MRF prior are only partially con- configuration x as P ðxi jxÞ. A simple way to understand
sidered, as the prior energy in an initial configuration of the likelihood function is that an updated configuration
stratigraphy is not optimized. may have a higher chance to occur if it agrees more with
In the second stage, spatial points in the generated initial the large-scale characteristics; the posterior probability
configuration, other than those at the borehole locations, for stratigraphic configuration x, given the initial estimate
are optimized using a Markov Chain Monte Carlo xi , is denoted as P ðxjxi Þ. According to the Bayes theorem,
(MCMC) algorithm to satisfy the local characteristics gov- there exists a relationship among these probabilities as
erned by the MRF prior. It is worth mentioning that over- follows:
correction by the MRF prior can damage the large-scale
P ðxjxi Þ / P ðxÞP ðxi jxÞ ð2Þ
characteristics in the initial configuration, as the large-
scale trend is formed by complicated physical processes Due to the equivalence of MRF and Gibbs functions
and hence will not perfectly follow a mathematical random (Clifford, 1990), P ðxÞ and P ðxi jxÞ can be further specified
field model. One of the known negative impacts of overcor- by means of the following energy functions:
rection is oversmoothing (Tolpekin & Stein, 2009). To
1
avoid such overcorrection, the initial stratigraphic configu- P ðxÞ ¼ expfU ðxÞg ð3Þ
ration in stage #1 should be considered as a constraint dur- Zp
ing the subsequent optimization process in stage #2. 1
P ðxi jxÞ ¼ expfU ðxi jxÞg ð4Þ
Therefore, a likelihood function (see Section 2.4) is used Zl
as a similarity measure between the initial configuration
and the optimized configuration in terms of energy. Com- in which Z p and Z l are normalizing constants, and U ðxÞ
bining the two aforementioned modeling stages, the gener- and U ðxi jxÞ are referred to as the prior energy and likeli-
ated stratigraphic realizations can reasonably reflect both hood energy, respectively.
the large-scale and local characteristics according to both
available observations and knowledge of local geology, 2.3 Prior energy
and thereby can be regarded as qualified modeling results
of the real stratigraphy. A flow chart showing the above- We model the prior energy as the sum of pairwise inter-
mentioned modeling procedure is provided in Fig. 1, and actions between elements as follows:
the final step ‘‘Stratigraphic uncertainty quantification‘‘ X
U ðxÞ ¼ V c ðxÞ ð5Þ
will be explained in Section 3. c2C

2.2 Markov random field and equivalent Gibbs distribution where c represents a clique containing soil labels (xr , xs ) of a
pair of neighboring elements (r, s), C is the set of all cliques
A complete introduction to MRFs and the equivalence in the MRF, and V c is referred to as the prior potential for
between an MRF and the corresponding Gibbs distribu- each clique c. A classical Potts model (Wu, 1982) is then
tion can be found in the existing literature (Geman & employed for V c as follows:
Geman, 1984; Li, 2001). In the following sections, only 
qðr; sÞ if xr ¼ xs
the basic concepts necessary for understanding the V c ðx r ; xs Þ ¼ ð6Þ
0 if xr –xs
employed modeling technique in the present work are
introduced. In the case where the subsurface geological It is worth mentioning that Eqs. (1)–(6) actually show a
domain of interest is discretized by a certain mesh scheme commonly used procedure to define an MRF. However,
(square lattices for 2D sections and cubic lattices for 3D the specific spatial correlation in the defined MRF,
bodies in this study), the finite set of elements is denoted which directly determines the structure of the modeled

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
4 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 1. Flowchart of the modeling and uncertainty quantification approach.

stratigraphy, is highly dependent on the formation of the angle hs!r is defined as the strength of the spatial correla-
prior potential function qðr; sÞ. Therefore, encoding our tion between the soil label of element s and the soil label
prior knowledge of spatial correlation by specifying of the neighboring element r, as shown in Fig. 3(a). The
qðr; sÞ is the core of the established model, since it is aimed anisotropy of the developed spatial correlation model is
at reflecting the inherent characteristics of stratigraphic reflected by the unequal lengths of the axes of the ellipse.
structures. Three main characteristics of stratigraphic The direction of the minor axis indicates the normal direc-
structures are taken into consideration in the employed tion of the interfaces between soil formations, and the
model: (1) heterogeneity, which indicates that the modeled direction of the major axis indicates the tangential direc-
region consists of multiple soil/rock formations; (2) tion of the interface between soil formations (i.e., the local
anisotropy, which indicates that the strength of the spatial stratigraphic dip). Typically, the length of the minor axis
correlation is directionally dependent; and (3) nonstation- rmin is set to 1.0; the length of the major axis r1 , denoted
arity, which indicates that the direction of the preponder- as the correlation strength parameter, determines the
ant correlation strength (i.e., the stratigraphic dip) differs degree of anisotropy in the model. Hence, qðr; sÞ can be cal-
from one location to another. To incorporate these desired culated using the following equation:
characteristics into the proposed model, the spatial correla- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
tion between soil/rock labels of a pair of neighboring qðr; sÞ ¼ r1 = cos2 ðhs!r  Hr Þ þ r21 sin2 ðhs!r  Hr Þ ð7Þ
elements (r; s) is designed to be related to two features:
(1) the label of the neighboring element xs and (2) the rela- Furthermore, the nonstationarity of the model is repre-
tive direction of element s to element r. For a 2D model, sented by the various polar angles Hr of each element,
the relative direction is represented by a polar angle hs!r which are referred to as orientation parameters and cause
from the centroid of element r to the centroid of element the local stratigraphic dips to vary from one location to
s in a given global polar coordinate system O-XY, as shown another. In this model, the orientation parameters are
in Fig. 2(a). Similarly, two polar angles hs!r , us!r in a given regarded as deterministic inputs rather than variable model
global spherical coordinate system O-XYZ, as illustrated in parameters. In practice, such orientation information can
Fig. 2(b), can uniquely determine the relative direction of be obtained or interpreted using geological maps, geophys-
neighboring elements in a 3D model. ical measurements and local experience. Typically, such
First, we consider only the spatial correlation of soil measurements or interpretation can be regarded as soft
labels between a pair of neighboring elements (r; s). For a data, which can only provide a general trend of the strati-
2D model, an ellipse centered at the centroid of element r graphic dips in the subsurface region but may not accu-
is used to describe the strength of the spatial correlation rately specify the exact dips at certain local points.
between the soil label of element r and the soil labels of Different from hard data, such as borehole logs, which
neighboring elements in different relative directions. The need to be satisfied completely, soft data could be partially
major axis of the ellipse has a polar angle Hr 2 ð p2 ; p2 in satisfied with respect to the uncertainty associated with the
the defined global polar coordinate system O-XY. Under measurement or estimation. However, it has been reported
this condition, the radius length qðr; sÞ along the polar that incorporating such soft data can significantly improve

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 5

Fig. 2. (a) Geometric condition of neighboring element pairs in the 2D model; (b) geometric condition of neighboring element pairs in the 3D model.

Fig. 3. (a) Elliptical spatial correlation model for the 2D model; (b) ellipsoidal spatial correlation model for the 3D model.

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
geostatistical simulation processes and outcomes (Poeter & 2
qðr; sÞ ¼ r1 r2 = r1 2 sin2 hsin2 u þ r2 sin2 hcos2 u þ r1 2 r2 2 cos2 h
McKenna, 1995; Wang, Wang, et al., 2017; Wang,
Wellmann, et al., 2017). As will be shown in the following ð8Þ
sections, such partial incorporation of soft data is automat-
ically implemented in the proposed stochastic modeling in which h ¼ hs!r  Hr þ p=2, u ¼ us!r  Ur .
process since strata orientation parameters only impact
the probability of the label assignment instead of affecting 2.4 Likelihood energy
the assigned label directly.
The spatial correlation model for a 3D model is similar As mentioned in the previous section, unrestrained use
to the 2D case, except that the strength of the spatial cor- of MRF prior may lead to overcorrection of the initial con-
relation is represented by the radius length qðr; sÞ of an figurations xi , thereby deterioration of the large-scale char-
ellipsoid, as shown in Fig. 3(b). The only differences are acteristics of the modeled stratigraphy. Therefore, we
(1) the ellipsoid has two major axes and corresponding cor- introduce the likelihood energy U ðxi jxÞ to prevent such
relation strength parameters (r1 and r2 ) and (2) two orien- overcorrection. As a measure of similarity between the ini-
tation parameters (Hr and Ur ) are needed to determine the tial configuration xi and its corresponding updated realiza-
major extension direction of the soil layer at the location of tion, U ðxi jxÞ can be represented as the accumulation of the
element r. Given these parameters, the correlation strength likelihood potential V r relating to the soil label assignment
qðr; sÞ in the 3D model can be calculated using the follow- of each element r in the current configuration x and the ini-
ing equation: tial estimate xi as follows:

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
6 X. Wang / Underground Space xxx (xxxx) xxx
X
U ðxi jxÞ ¼ V r ðx; xi Þ ð9Þ 2016). A total of 1 000 initial stratigraphic configurations
r2R were generated by the MCS process; their corresponding
optimized stratigraphic realizations were further obtained
Since the soil labels are discrete variables, the following
through the MCMC process. As a demonstration, one of
form of V r is employed in the proposed model:
 the initial stratigraphic configurations and its correspond-
  l if xr ¼ xir ing optimized realization are shown in Figs. 4(b) and (c).
V r xr ; xir ¼ ð10Þ
0 if xr –xir It can be noted that the large-scale characteristics of the ini-
tial configuration are maintained while the smoothing of
in which l is referred to as a smoothing factor. In this local boundaries is implemented.
study, l is set as 0.2, and a detailed parametric study of
l can be found in (Wang, Li, Wang, Rong, & Liang,
2016). It can be noted from Eqs. (3)–(4) and (9)–(10) that 3.2 Stratigraphic uncertainty quantification
the amendment of soil label assignments in the initial con-
figuration xi that results in an increase of prior probability After a sufficient number of stratigraphic realizations
P ðxÞ accordingly leads to a decrease of likelihood probabil- are generated, the possible soil labels for each element
ity P ðxi jxÞ. In other words, during the MCMC process for and their corresponding probabilities of being assigned
maximizing the posterior probability P ðxjxi Þ in Eq. (2) in can be obtained. For instance, pl ðrÞ (l 2 L, r 2 R) can be
the second modeling stage, a balance is achieved between calculated according to the frequency of assigning element
the prior part and the likelihood part. Such balance allows r with soil label l. According to the computed probabilities,
local amendments to increase prior probability with rea- we can mark those elements that have a probability of
sonably smoothed local boundaries but rejects global being assigned with one soil formation greater than a given
changes that contradict the overall shape of the stratigra- confidence level (e.g., 95%) as elements with confident label
phy determined by borehole logs and local geologic assignments. The ratio of the number of these elements to
information. the number of total elements in the domain is then defined
as the confidence ratio. For those elements with unconfi-
3 Uncertainty quantification procedure dent soil labels, we further employ the concept of informa-
tion entropy and compute its value for each element in the
3.1 Generating stratigraphy realizations modeled geological domain according to the following
form:
Given the definition of posterior probability/energy, a X
commonly used Gibbs sampler is employed to conduct H r ¼  l2L pi ðrÞLnðpi ðrÞÞ ð11Þ
the conditional sampling process in the first modeling stage
as well as the conditional updating process in the second A detailed introduction to the application of informa-
modeling stage. To avoid redundancy, detailed modeling tion entropy for measuring geological uncertainty can be
procedures, including issues regarding the algorithm of found in Wellmann and Regenauer-Lieb (2012). As shown
the Gibbs sampler, the sampling sequence, the convergence in Fig. 4(d), the map of information entropy can provide us
criteria, and the optimal number of generated stratigraphic with a clear view of the uncertainties associated with each
realizations, are not discussed in this paper; interested read- element in the modeled subsurface region. It can be noted
ers are referred to the authors’ previous publications, such that the stratigraphic uncertainty of the elements located
as (Li, Wang, et al., 2016; Wang et al., 2016). Herein, we close to the observational data (i.e., ground surface or
provide an artificial example to illustrate the aforemen- boreholes) is negligible since the observational data has a
tioned two-stage modeling procedure. strong influence on the soil label assignment of the nearby
The subsurface domain of interest in the artificial exam- elements. In contrast, for those elements far from the bore-
ple is a 2D section with a length of 40 meters and a height holes, the locations of contact surfaces between soil forma-
of 20 meters. The entire region is discretized with square tions become more uncertain. Such results are in
lattices into 12 800 elements (the element size is accordance with our intuition and common sense since
0.25 m  0.25 m). Three synthetic vertical boreholes are borehole information can only provide a limited constraint
located at positions with horizontal distances of 5 m, on the label assignment within a certain distance from the
20 m and 35 m from the left boundary. It can be noted borehole location.
from these borehole logs that three soil formations exist
in the modeled region. The dimensions of the modeled sub-
surface domain, as well as the adopted synthetic borehole 4 Examples for modeling common stratigraphic structures
logs and surface information, are shown in Fig. 4(a). We
first set the orientation parameters hr for each element to In this section, using the example provided in the previous
0 and the strength parameter r1 to 3.0, which represents section as a reference case, we further provide a series of 2D
the case of a moderately layered soil deposit based on artificial examples to demonstrate and validate the capabil-
our existing studies (Li, Wang, et al., 2016; Wang et al., ity of the developed modeling workflow to reproduce some

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 7

common stratigraphic structures according to available The fault structure can be modeled by adjusting the local
measurements or local geologic information. neighborhood system at the location of the faults using
the developed modeling workflow. Herein, we use the same
4.1 Modeling of faults input data from borehole logs in the reference case but
place a cut in the neighborhood to divide adjacent elements
Herein, we look at a special case where a fault structure so that they will be located on opposite sides of the mod-
with a known location and orientation exists in the subsur- eled fault. The location of the fault interface is marked
face domain of interest. The fault structure can be regarded on one of the generated realizations as well as on the infor-
as embedding two parallel free boundaries into an other- mation entropy map, as shown in Fig. 5. From both the
wise continuous domain so that the spatial correlation of information entropy map and the illustrative stratigraphic
soil labels on the opposite sides of the fault is disconnected. realization, it can be noted that the assigned soil labels are

Fig. 4. (a) Available borehole logs; (b) one of the initial configurations for the reference case; (c) one of the stratigraphic realizations for the reference case;
(d) information entropy map for the reference case.

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
8 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 4. (continued)

continuous on the same side of the fault interface, but they resent the stratigraphic dips in the modeled subsurface
exhibit strong discontinuity between the two different sides domain. The orientation parameter hr for each element is
of the fault. Therefore, the presented case demonstrates specified as the polar angle of the local tangent of these
that the developed modeling workflow can reproduce the sinusoids. Following the same procedure as described in
fault structure and reflect the corresponding uncertainty, the previous section, a sufficient number of stratigraphic
given sufficient knowledge of the location of the fault realizations are generated according to the new orientation
interface. parameters. A typical realization and the information
entropy map are illustrated in Figs. 6(b) and (c), respec-
4.2 Modeling of folding and uplift tively. From these figures, it can be noted that the gener-
ated stratigraphic realizations exhibit a similar trend to
On the basis of the reference case described above, we the predefined stratigraphic dip that is seen in Fig. 6(a).
now attempt to incorporate the prior knowledge of the Such agreement in the trend of the stratigraphy can
stratigraphic dips by setting the orientation parameters be attributed to the elliptical and ellipsoidal spatial
hr . A group of sinusoids shown in Fig. 6(a) is used to rep- correlation models (introduced in Section 2), which result

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 9

Fig. 5. (a) One of the stratigraphic realizations for the case considering fault structure; (b) information entropy map for the case considering fault
structure.

in anisotropic correlation strength in different directions. parameters (r1 = 1.0, 3.0, 5.0) are discussed in this section.
Such anisotropic correlation enables representing the The information entropy map for each modeling case is
known stratigraphic dips and generate stratigraphic real- shown in Fig. 7. According to the definition of the correla-
izations accordingly. Therefore, the developed workflow tion strength parameters, a parameter r1 equal to 1.0 indi-
possesses the capability to model stratigraphy under com- cates isotropic correlation strength in all directions.
plex geological conditions, including uplift and folding. Without a dominant extension direction in the soil layer,
beyond a certain distance of the borehole locations, the lay-
4.3 Modeling of different layered structures ered structure may not be maintained, and the locations of
contact surfaces between soil formations will become
To demonstrate the effect of the correlation strength highly uncertain. Therefore, in a case where the correlation
parameter r1 in the 2D case (or r1 , r2 in the 3D case) on strength parameter r1 is equal to 1.0, the regions in between
the generated subsurface stratigraphic realizations, a group boreholes will exhibit a higher degree of uncertainty than
of modeling cases using the same borehole information as the regions close to the borehole locations. On the other
the reference case but with different correlation strength hand, parameter r1 >1.0 provides anisotropic correlation

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
10 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 6. (a) The adopted synthetic orientation information; (b) one of the stratigraphic realizations for the case considering orientation information; (c)
information entropy map for the case considering orientation information.

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 11

Fig. 7. (a) Information entropy map usingr1 = 1.0; (b) information entropy map usingr1 = 3.0; (c) information entropy map usingr1 = 5.0.

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
12 X. Wang / Underground Space xxx (xxxx) xxx

strength; the larger the parameter r1 becomes, the stronger responding borehole logs, and the created 3D model for the
the anisotropy will be. Hence, a large parameter r1 can subsurface domain in this project can be found in Fig. 9.
exert a strong extension tendency of the soil layer along Since the depth of the investigated region in both the 2D
the tangential direction of the interfaces between the two and the 3D projects is shallow, and the relief is low to mod-
contacting soil formations. Consequently, in a case where erate, the orientation parameters for both cases are set to
the correlation strength parameter r1 is equal to 5.0, a very zero to represent horizontal sedimentary conditions. Fur-
stable layered structure of soil formation can be preserved. thermore, the correlation strength parameters are set to
The possible locations of the interfaces between the con- 3.0 to represent a moderately layered aggregation style
tacting strata are limited to a much smaller area. under conditions of normal soil deposition.
Adjusting correlation strength parameters allows for
modeling stratigraphy with various aggregation character- 5.2 Interpreting the interface between soil and bedrock (top
istics, from a clustered structure to a highly layered one. of rock elevation)
For a specific case, a Bayesian inferential framework can
be implemented to determine the appropriate values for In this section, we focus on estimating the location of
correlation strength parameters, which make the generated the interface between the soil and rock for the investigated
subsurface stratigraphic realizations compatible with the subsurface regions since such interpretation results are crit-
available observations. The discussion of this aspect of ical for designing an adequate foundation to support the
the framework is outside the scope of this paper; the details structural loads. Hence, we divide the subsurface region
can be found in Li, Wang, et al. (2016) and Wang, into two main categories: overburden soil and bedrock.
Wellmann, et al. (2017). The ground surface of both projects, according to the field
observations, contains no outcrops.
5 Uncertainty quantification and reduction in site For the 2D case, we present modeling results for three
characterization for real projects scenarios with different combinations of borehole log data
as input:
In this section, we apply the proposed stratigraphic
modeling method and uncertainty quantification frame- Case #1: Borehole #1 + Borehole #5
work to two real-world highway construction projects to Case #2: Borehole #1 + Borehole #3 + Borehole #5
illustrate the use of the proposed methodology in engineer- Case #3: Borehole #1 + Borehole #2 + Borehole #3
ing practice. + Borehole #4 + Borehole #5.

Figure 10(a) shows the confidence ratio of these three


5.1 Geology and observations of the projects cases under a confidence level of 90%. It can be noted
that with the increase in the number of borehole logs used
5.1.1 2D case as input, the confidence ratio values exhibit a significant
The 2D case is a project to replace an existing four-span upward trend from 46.16% (Case #1) to 89.19% (Case
bridge over Interstate 90 in Sheffield, Ohio. The construc- #3). Such changes indicate that uncertainty in the loca-
tion area is located within the Erie Lake Plain, which is tion of the soil-rock interface in the modeled domain
characterized by low relief terrain. A drilling report consist- decreases with an increase in the number of available
ing of five borehole logs indicates the composition of over- borehole logs. We can draw a similar conclusion by com-
burden soils as a mixture of lacustrine sand, silt, and clay, paring the confident label assignments and the informa-
and the overburden soils are underlain by shale and sand- tion entropy maps of the three cases shown in Fig. 11.
stone. The location of the available boreholes is marked in It can be noted from the results of Case #1 that its area
Fig. 8(a). Corresponding borehole logs can be found in of the uncertain region, as well as the area associated with
Fig. 8(b). The stratigraphic profile connecting the available high information entropy values, are the largest. This is
boreholes is converted into a 2D domain of interest, as because two borehole logs located far away from each
shown in Fig. 8(c). other cannot provide enough constraint to ascertain the
location of the interface in the region between them.
5.1.2 3D case However, the uncertain region in Case #3 is limited to
The 3D case is a project to reconstruct the Interstate 77/ a much smaller area. This result indicates that additional
Interstate 76 interchange in Akron, Ohio. Soil samples boreholes drilled in highly uncertain regions can eliminate
obtained from drilling operations show that subsurface the uncertainty in the soil-rock interface near the borehole
soils consist of approximately 2.6–8.5 m of overburden locations. Meanwhile, the distance between boreholes is
material underlain by shale. The overburden material is closer, so a stronger constraint can be exerted to prevent
variable—it contains sand, clay, silt, gravel, and mixtures the soil/rock labels from varying greatly. More impor-
of these materials. A 3D subsurface region of the project tantly, by comparing the modeling results of Case #1,
containing nine boreholes is studied using the developed Case #2 and logs of Borehole #2–4, it can be found that
analysis approach. The locations of the boreholes, the cor- the soil-rock interfaces detected by the newly added

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 13

Fig. 8. (a) Available boreholes in the 2D project; (b) borehole logs for the 2D project; (c) MRF created for the 2D project.

borehole logs exactly locate at the regions with high infor- In the modeling of the 3D case, we also use different
mation entropy values in the previous modeling cases, groups of borehole logs to establish three modeling cases:
which are the most likely locations of the interface indi-
cated by the developed approach. Such results can be Case #4: Borehole #6 through Borehole #9 (a total of
regarded as a validation process, which demonstrates that four boreholes)
the generated stratigraphic realizations using the Case #5: Borehole #6 through Borehole #13 (a total of
approach developed herein are realistic. eight boreholes)

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
14 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 10. (a) Confidence ratios of different modeling cases for interpreting
the soil-rock interface in the 2D project; (b) confidence ratios of different
modeling cases for interpreting the soil-rock interface in the 3D project.

to the 2D modeling cases, the results show that more con-


fident interpretation results can be obtained by inputting
Fig. 9. (a) Available boreholes in the 3D project; (b) borehole logs for the additional borehole logs.
3D project; (c) MRF created for the 3D project.
The above examples are intended to show how strati-
graphic uncertainty can be evaluated using the established
Case #6: Borehole #6 through Borehole #14 (a total of stratigraphic modeling and uncertainty quantification
nine boreholes) workflow. For situations where existing borehole logs are
available, the developed approach can provide a quantita-
From the analysis results shown in Fig. 10(b), the uncer- tive measurement of the stratigraphic uncertainty, in terms
tainty in the location of the soil-rock interface decreases of confidence ratios, as well as its spatial distribution, in
with the increasing number of borehole logs used in the terms of information entropy maps. Such quantitative
model, manifested as a rise in the confidence ratio under measurements enable the assessment of the confidence of
a confidence level of 90% from 42.16% to 69.84%. The con- interpretations of the stratigraphy, for the entire subsurface
fident label assignment for each case is illustrated in region or for local zones, in accordance with the demands
Fig. 12. Furthermore, Fig. 13 shows the contour of infor- of individual projects. In a case where additional geotech-
mation entropy on the cross section of y = 15 m. Similar nical site investigation (i.e., new borehole drillings) is

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 15

Fig. 11. (a) Information entropy map for interpreting the soil-rock interface in the 2D project; (b) map of confident label assignments for interpreting the
soil-rock interface in the 2D project.

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
16 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 12. Maps of confident label assignments for interpreting the soil-rock interface in the 3D project using (a) four boreholes; (b) eight boreholes; and (c)
nine boreholes.

needed to obtain a more confident interpretation, the a confidence level of 75% and the resulting information
developed workflow can also provide suggestions regarding entropy map are shown in Figs. 14(b) and (c), respectively.
the locations the additional borehole drilling should be The computed confidence ratio for the modeled case is
conducted—namely, the borehole drilled at the location 72.23%.
with the thickest uncertain region or the highest informa- Compared to the location of the soil-rock interface,
tion entropy value should be able to provide data that given the same observation information, reproducing
would reduce the stratigraphic uncertainty to the greatest stratigraphy consisting of multiple soil categories and inter-
extent. With the aid of the established workflow, the newly faces is much more uncertain, as indicated by the lower
obtained borehole logs can be used as input for updating confidence ratio even under a less strict confidence level.
the model and the stratigraphic uncertainty quantification, However, this does not affect the application of the devel-
and the cycle of model interpretation and new borehole site oped approach to generate reasonable stratigraphic realiza-
suggestion can be continued until the desired confidence tions and to estimate the stratigraphic uncertainty,
ratio is achieved. conditional on the available borehole logs. As shown in
the analysis results, to obtain a more confident understand-
5.3 Interpreting the stratigraphic structure ing of this complicated stratigraphic structure, additional
boreholes need to be drilled. Furthermore, it is worth men-
In this section, the samples of overburden soil obtained tioning that, in such a complex case, the drilling locations
from boreholes in the 2D project are classified into multiple for additional boreholes suggested by the developed
categories according to the American Association of State approach (i.e., the most uncertain region between two
Highway and Transportation Officials (AASHTO) soil existing boreholes) may no longer be located at the mid-
classification system, as shown in the borehole logs in point of existing boreholes, as shown by the uncertain area
Fig. 8(b). Thus, the stratigraphic structure of the geological between Borehole #1 and Borehole #2 in Figs. 14(b) and
domain in this project can be updated for a more compli- (c). Therefore, it seems that the commonly used site inves-
cated case to verify the application of the proposed work- tigation strategy that uses equal spacing between drilling
flow with practical problems requiring a refined locations may not always be the most optimized.
classification of soil formations. To obtain the best result
possible, all five available borehole logs in this project are 6 Summary and concluding remarks
used for modeling the subsurface stratigraphy. Following
the same analysis procedure, stratigraphic realizations The tradeoff between the cost of conducting a thorough
can be generated, and one of them is illustrated in geotechnical site investigation and the acquired geological
Fig. 14(a). The map of confident label assignment under knowledge for implementing a cost-effective yet reliable

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 17

Fig. 13. Contours of information entropy along the cross section of y = 15 m for interpreting the soil-rock interface in the 3D project using (a) four
boreholes; (b) eight boreholes; and (c) nine boreholes.

design and construction is an issue that almost every common subsurface stratigraphic structures conditioned
geotechnical engineer must confront. To address this prob- on borehole logs and prior knowledge. Such capability
lem, in this paper, we have presented an extended and more makes the established workflow a flexible and versatile tool
versatile geological modeling approach and workflow, for interpreting complex stratigraphy in engineering
which can assist in the uncertainty quantification of inter- practice.
preted stratigraphy as well as guide additional site investi- Furthermore, the developed analysis approach is cap-
gation in an objective and quantitative manner. able of providing stochastic realizations of possible subsur-
The modeling technique and procedure, including a new face stratigraphy with corresponding uncertainty
extension for modeling 3D structures and a two-stage sam- quantification. The ‘‘confidence ratio” is adopted as a
pling process, were briefly introduced in this paper. Major direct indicator of the confidence concerning the inter-
effort was spent on illustrating the significance and effec- preted stratigraphy. The information entropy map further
tiveness of the established workflow using both artificial provides an intuitive visual description of the stratigraphic
and real-world examples. uncertainty.
As demonstrated in the artificial cases shown in Sec- Studies of two real geotechnical projects demonstrate
tion 4, the developed modeling approach is able to simulate that the developed analysis approach is adequate in both

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
18 X. Wang / Underground Space xxx (xxxx) xxx

Fig. 14. (a) One of the stratigraphic realizations for the case considering multiple soil formations; (b) Information entropy map for interpreting the
stratigraphic structure in the 2D project; (c) Map of confident label assignments for interpreting the stratigraphic structure in the 2D project.

the 2D and 3D domains. A reasonable assessment of the Blanchin, R., & Chilès, J. P. (1993). The Channel Tunnel: Geostatistical
prediction of the geological conditions and its validation by the reality.
confidence of the current interpretation of the subsurface Mathematical Geology, 25(7), 963–974.
stratigraphy can be obtained using the developed approach Caers, J., & Zhang, T. (2004). Multiple-point geostatistics: A quantitative
based on existing data from borehole logs and local geo- vehicle for integrating geologic analogs into multiple reservoir models,
integration of outcrop and modern analog data in reservoir models.
logic information. Furthermore, based on the uncertainty AAPG Memoir, 80, 383–394.
quantification results, the developed approach can be used Calcagno, P., Chilès, J. P., Courrioux, G., & Guillen, A. (2008).
to optimize the locations of additional boreholes to reduce Geological modelling from field data and geological knowledge: Part
I. Modelling method coupling 3D potential-field interpolation and
uncertainties. geological rules. Physics of the Earth and Planetary Interiors, 171(1–4),
147–157.
Conflict of interest Carle, S. F., & Fogg, G. E. (1996). Transition probability-based indicator
geostatistics. Mathematical Geology, 28(4), 453–476.
Carle, S. F., & Fogg, G. E. (1997). Modeling spatial variability with one
The author confirms there is no conflict of interest. and multidimensional continuous-lag Markov chains. Mathematical
Geology, 29(7), 891–918.
Chilès, J. P., Aug, C., Guillen, A., & Lees, T. (2004). Modelling the
References geometry of geological units and its uncertainty in 3D from structural
data: The potential-field method. In Proceedings of international
Auerbach, S., & Schaeben, H. (1990). Computer-aided geometric design of symposium on orebody modelling and strategic mine planning
geologic surfaces and bodies. Mathematical Geology, 22(8), 957–987. (pp. 22–24).

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
X. Wang / Underground Space xxx (xxxx) xxx 19

Clifford, P. (1990). Markov random fields in statistics. Disorder in physical environment: From theory to practice. New York: ASCE (pp. 119–133).
systems: A volume in honour of John M. Hammersley. USA: Oxford New York: ASCE.
University Press (pp. 19–32). USA: Oxford University Press. Poeter, E. P., & McKenna, S. A. (1995). Reducing uncertainty associated
Dai, Z., Ritzi, R. W., & Dominic, D. F. (2005). Improving permeability with ground-water flow and transport predictions. Groundwater, 33(6),
semivariograms with transition probability models of hierarchical 899–904.
sedimentary architecture derived from outcrop analog studies. Water Toftaker, H., & Tjelmeland, H. (2013). Construction of binary multi-grid
Resources Research, 41(7), W07032. Markov random field prior models from training images. Mathemat-
Elfeki, A., & Dekking, M. (2001). A Markov chain model for subsurface ical Geosciences, 45(4), 383–409.
characterization: Theory and applications. Mathematical Geology, 33 Tolpekin, V. A., & Stein, A. (2009). Quantification of the effects of land-
(5), 569–589. cover-class spectral separability on the accuracy of Markov-random-
Geman, S., & Geman, D. (1984). Stochastic relaxation, Gibbs distribu- field-based superresolution mapping. IEEE Transactions on Geoscience
tions, and the Bayesian restoration of images. IEEE Transactions on and Remote Sensing, 47(9), 3283–3297.
Pattern Analysis and Machine Intelligence, PAMI-6(6), 721–741. Wang, X., Li, Z., Wang, H., Rong, Q., & Liang, R. Y. (2016).
Hu, L., & Chugunova, T. (2008). Multiple-point geostatistics for modeling Probabilistic analysis of shield-driven tunnel in multiple strata consid-
subsurface heterogeneity: A comprehensive review. Water Resources ering stratigraphic uncertainty. Structural Safety, 62, 88–100.
Research, 44(11), W11413. Wang, X., Wang, H., & Liang, R. Y. (2017). A method for slope stability
Jaksa, M. B., Goldsworthy, J. S., Fenton, G. A., Kaggwa, W. S., Griffiths, analysis considering subsurface stratigraphic uncertainty. Landslides,
D. V., Kuo, Y. L., & Poulos, H. G. (2005). Towards reliable and 15(5), 925–936.
effective site investigations. Geotechnique, 55, 109–121. Wang, X., Wang, H., Liang, R. Y., & Liu, Y. (2018). A semi-supervised
Juang, C., Jiang, T., & Christopher, R. (2001). Three-dimensional site clustering-based approach for stratification identification using bore-
characterisation: Neural network approach. Geotechnique, 51(9), hole and cone penetration test data. Engineering Geology, 248,
799–809. 102–116.
Li, S. Z. (2001). Markov random field modeling in computer vision (2nd ed.). Wang, X., Wang, H., Liang, R. Y., Zhu, H., & Di, H. (2018). A hidden
Berlin: Spinger-Verlag. Markov random field model based approach for probabilistic site
Li, W., Li, B., & Shi, Y. (1999). Markov-chain simulation of soil textural characterization using multiple cone penetration test data. Structural
profiles. Geoderma, 92(1–2), 37–53. Safety, 70, 128–138.
Li, D. Q., Qi, X. H., Cao, Z. J., Tang, X. S., Phoon, K. K., & Zhou, C. B. Wang, H., Wellmann, J. F., Li, Z., Wang, X., & Liang, R. Y. (2017). A
(2016). Evaluating slope stability uncertainty using coupled Markov segmentation approach for stochastic geological modeling using
chain. Computers and Geotechnics, 73, 72–82. hidden Markov random fields. Mathematical Geosciences, 49, 145–177.
Li, Z., Wang, X., Wang, H., & Liang, R. Y. (2016). Quantifying Wellmann, J. F., & Regenauer-Lieb, K. (2012). Uncertainties have a
stratigraphic uncertainties by stochastic simulation techniques based meaning: Information entropy as a quality measure for 3-D geological
on Markov random field. Engineering Geology, 201, 106–122. models. Tectonophysics, 526–529, 207–216.
Li, X., Zhang, L. M., & Li, J. (2015). Using conditioned random field to Wu, F. Y. (1982). The potts model. Reviews of Modern Physics, 54(1),
characterize the variability of geologic profiles. Journal of Geotechnical 235–268.
and Geoenvironmental Engineering, 142(4), 04015096. Zhu, H., & Zhang, L. M. (2013). Characterizing geotechnical anisotropic
Liao, S. S., Druss, D. L., Neff, T. L., & Brenner, B. R. (1996). Just One spatial variations using random field theory. Canadian Geotechnical
More Boring, and We’ll Know for Sure! Uncertainty in the geologic Journal, 50(7), 723–734.

Please cite this article as: X. Wang, Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical
investigation data, Underground Space, https://doi.org/10.1016/j.undsp.2018.10.008
View publication stats

You might also like