Research Article

Transportation Research Record

1–24
Ó The Author(s) 2024
Machine Learning Methods for
Article reuse guidelines:
Geotechnical Site Characterization sagepub.com/journals-permissions
DOI: 10.1177/03611981241257512
and Scour Assessment journals.sagepub.com/home/trr

Negin Yousefpour1 , Zhongqiang Liu2 , and Chao Zhao3

Abstract
Reliable geotechnical site characterization and geohazard assessment are critical for bridge foundation design and manage-
ment. This paper explores existing and emerging artificial intelligence-machine learning methods (AI-ML) transforming geo-
technical site characterization and scour assessment for bridge foundation design and maintenance. The prevalent ML
techniques applied for subsurface characterization are reviewed, and step-by-step methodologies for stratigraphy classifica-
tion, borehole interpretation, geomaterial characterization, and ground modeling are provided. The ML techniques for maxi-
mum scour depth prediction are reviewed, and a simple ML methodology is proposed to provide a more reliable tool for
scour depth estimation for implementation in practice. Also, a novel deep learning approach, with a detailed implementation
description, is recommended for real-time scour monitoring and assessment of existing bridges. The challenges with database
design and data processing for ML modeling, model optimization, training and validation, and uncertainty assessments are dis-
cussed, and innovative techniques for addressing them are reviewed.

Keywords
Scour, Geotechnical Site Charactrization, Geohazards, Foundations of bridges, Geotechnical Asset Management, Machine
Learning (Artificial Intelligence), Neural Networks

Reliable geotechnical site characterization and geoha- geotechnics, Zhang et al.’s (5) on DL applications in geo-
zard assessment are critical elements in bridge infrastruc- technical engineering and ML modeling of soil properties
ture design and management. Scour, a major geohazard (6), Dikshit et al.’s (7) on geohazard modeling, Tehrani
affecting bridges over waterways, is the main cause of et al.’s (8) landslide studies and recently Phoon and
failure in the United States and worldwide (1–3). In this Zhang’s (9) review on the future of ML in geotechnics.
paper, we review emerging artificial intelligence (AI) and The objective of this discussion is to highlight the
machine learning (ML) technologies that are transform- recent advances and underscore the importance of con-
ing geotechnical site characterization and scour assess- tinued research and innovation in harnessing the full
ments for bridge foundation design and maintenance. potential of AI-ML for safer, more resilient, and cost-
AI, a versatile field encompassing ML, data analytics, effective bridge foundations. The paper is organized as
and computational modeling, has catalyzed transformative follows. The next section introduces the synopsis on AI
change across various industries. Geotechnical engineering and ML. We then present subsurface characterization
is no exception, benefiting from AI’s capabilities to extract and scour assessment using ML. We also provide recom-
valuable insights, enhance predictive accuracy, and optimize mendations with step-by-step implementation methods
decision-making processes, considering geohazards. This
fusion of geotechnical engineering and AI has far-reaching 1
Department of Infrastructure Engineering, The University of Melbourne,
implications, promising safer and more reliable infrastruc- Parkville, Victoria, Australia
2
ture, cost savings, and minimized environmental impact. Norwegian Geotechnical Institute, Oslo, Norway
3
Faculty of Engineering, China University of Geosciences, Wuhan, China
There are state of the art reviews on ML/deep learning
(DL) applications in geotechnical engineering, including Corresponding Author:
Yousefpour and Fallah’s (4) on ML applications in Negin Yousefpour, [email protected]
2 Transportation Research Record 00(0)

for applications in practice. We end the paper with a

conclusion section.

Synopsis on Artificial Intelligence

and Machine Learning
AI and ML represent two dynamic and interrelated fields
at the forefront of modern technology. AI is a broader
concept, encompassing the development of intelligent
systems capable of tasks traditionally requiring human
intelligence, such as problem solving, language under-
standing, and decision making. ML, on the other hand,
is a subset of AI focused on the development of algo-
rithms that enable computers to learn from and make
predictions or decisions based on data.
AI and ML have seen unprecedented growth and
Figure 1. Machine learning survey: What are the machine
application across various industries. In the domain of learning applications that can transform value for our industry in
AI, natural language processing (NLP) has enabled the next 10 years? (Select all that apply).
conversational AI and language translation (10).
Computer vision has transformed image and video
analysis (11), while reinforcement learning has empow- construction automation and management (11, 26),
ered robots and autonomous systems (12). Meanwhile, transport planning (27), and in various aspects of geo-
ML techniques, including supervised learning, unsuper- technical engineering (4, 5).
vised learning, and reinforcement learning, have revolu-
tionized data-driven decision-making processes in fields
ranging from health care and finance to engineering Subsurface Characterization Using ML
applications (13, 14). The synergy between AI and ML
is evident in their iterative and data-centric approach. Subsurface characterization, including stratigraphic con-
ML algorithms learn patterns from data and improve figuration and the associated geotechnical properties, has
their performance over time, aligning with the AI goal long been a challenge in geotechnical practice. Because
of creating intelligent, adaptive systems. Additionally, of the geological processes leading to the compounded
the availability of vast data sets and increased comput- deposition and subsequent changes of geomaterials,
ing power has fueled the advancement of AI and ML including layering, each stratum exhibits variation from
applications. point to point within a volume (28). Nevertheless, the
Deep learning employs artificial neural networks identification of subsurface stratification and the charac-
inspired by the human brain to solve complex problems terization of spatially varying soil and rock properties
(15). DL has revolutionized AI by enabling break- with the limited availability of site investigation informa-
throughs in computer vision, natural language process- tion are indispensable for site characterization and fol-
ing, and speech recognition (16). Convolutional neural lowing design analyses.
networks (CNNs) (17) have enhanced image recognition, The TC309 (International Society for Soil Mechanics
while recurrent neural networks (RNNs) (18) and trans- and Geotechnical Engineering [ISSMGE]’s Technical
former models (19) have elevated language understand- Committee for Machine Learning and Big Data) carried
ing and generation to unprecedented levels. Generative out an online survey in 2019, receiving 114 responses.
AI is the most recent advancement in this field, creating The survey highlighted that machine learning is expected
systems capable of generating content, such as images, to have a significant impact on geotechnical site charac-
text, music, and more, often with remarkable creativity. terization among others. Site characterization is one of
At its core, generative AI seeks to enable machines to the leading areas where machine learning can transform
produce novel, human-like output. The driving force practice over the next 10 years (Figure 1).
behind many generative AI models is the use of neural
networks, particularly generative adversarial networks
Stratigraphy Classification
(GANs) (20).
In civil engineering, AI-ML has been used mostly in Stratigraphy classification is essential in site investigation,
structural health and performance monitoring (21–23), as it provides the necessary knowledge on the geological
design optimization (24), predictive maintenance (25), body (29). For example, the significance of the stratigraphy
Yousefpour et al 3

classification might be observed in the failure of a slope

Note: DL = deep learning; ML = machine learning; CNN = convolutional neural network; SVM = support vector machine; RF = random forest; GNN = graph neural network; RNN = recurrent neural
consisting of multiple strata, in which the failure is often

Iso-value of potential field

dominated by the spatial distribution of the weak stratum

Rock type of iron ore

(30). However, the subsurface stratigraphic configuration

Target

Mineralization type
Sedimentary facies
Sedimentary facies
at a site is hard to characterize accurately because of the

Scalar field, rock

complexity and inherent spatial variability of the strata

Stratum type

Stratum type

Stratum type
and the limited availability of exploration data.

Rock types
Lithofacies
The essence of stratigraphic modeling is a process of

Strata
interpolating and predicting strata of the whole area of
interest from a limited amount of known data (31). It gen-
erally treats geological units as discrete variables, such as
stratigraphic classes or rock classes, and thus formulates

Graph adjacency matrix, graph

the modeling task as a classification problem correspond-

Position, mineral components

Position, residual density, and

ing to discrete variables. DL and ML are the methods

Strata, rock, gravity, and

magnetic susceptibility
that learn a model with statistical characteristics from

Coordinates, elevation
Position, orientation
Position, orientation
known data and use the model to make judgments and

Input
predictions about new scenarios. In this way, the princi-

Log parameters

node matrix
ples of stratigraphic modeling fit with DL and ML (32).

magnetic
Pixel value
In recent years, there has been a boom in research on

Position
the stratigraphy classification based on DL or ML. As

Strata

Strata
listed in Table 1 (33), the possibility of various machine
learning algorithms in handling stratigraphy classifica-
tion tasks has been verified, from shallow classification
algorithms, such as convolutional neural networks

Boreholes, sections, geological

(CNN), eXtreme gradient boosting (XGBoost), support

Wells, geology map, sections

Borehole and 3-D inversion

Training image of geological
vector machines (SVM), decision tree (DT), random for-
est (RF), and maximum likelihood, to variants of neural

Orientation measures
Training image of flow

Borehole, orientation
networks, such deep feed-forward neural networks
Data

cross sections
(DFNN), RNNs, graph neural networks (GNN), and
GAN. The modeling ideas of these methods may be Drill holes
Boreholes

Boreholes
Table 1. Summary of DL and ML Algorithms Used in Stratigraphy Classification

Well-logs
divided into stratigraphy classification methods based on

model
the image analysis and those based on the borehole map
interpretation. Wells
The above studies have established the foundation for
the application of machine learning to stratigraphy classi-
fication. In addition, they also reveal the characteristics
of different approaches. Traditional machine learning
Multiple-point geostatistics
DL or ML algorithm

classifiers are suitable for learning tasks with small sam-

ples, while DL algorithms dominated by neural networks
Maximum likelihood

generally require large amounts of training data to obtain

Gaussian process
Stacking method

better learning results. However, it is always troublesome

Decision tree

to acquire and label geological data. The samples are too

XGBoost

sparse to represent the feature space effectively, thus lim-

GNN
CNN

RNN
SVM
SVM

iting to some extent the performance of DL models in

RF

geological reconstruction tasks.

Mariethoz and Caers (34)

Image Analysis
Bai and Tahmasebi (35)

xalves et al. (41)

xalves et al. (42)
Smirnoff et al. (37)

The depositional stratigraphic relationships can be quan-

Shi and Wang (36)

Hillier et al. (44)

Wang et al. (38)

Xiang et al. (40)

Zhou et al. (45)

Adeli et al. (39)

titatively reflected in a training image (TI), which can be

Jia et al. (43)

defined as a conceptual representation of the expected

subsurface heterogeneities in the area of interest.
network.
Study

Gonc
Gonc

Stochastic conditional simulation methods using TIs,

for example, multiple-point statistics (34–46) and
4 Transportation Research Record 00(0)

iterative convolution XGBoost (47), have been devel- likelihood information). There is a balance between the
oped to depict stratigraphic connectivity between soil prior geological model and site-specific measurements.
deposits. These image-based stochastic simulation Prior geological knowledge governs the posterior subsur-
algorithms have been successfully applied to tackle face system when available measurements are limited.
practical geotechnical problems, for example, reclama- However, the influence of a training image weakens as the
tion and slope stability (48). number of site-specific measurements increases (or likeli-
A TI is an ensemble of prior geological knowledge, hood information strengthens), and the final predicted geo-
which enables quantitative incorporation of subjective logical cross section will be mainly dominated by site-
geological interpretation of a studied domain, and it can specific data when many measurements are taken from a
be directly obtained from a nearby site or previous proj- specific site. The basic structure of this proposed frame-
ects with similar geological settings. The idea of TI is work can be summarized with the following steps.
appealing to geological and geotechnical practitioners as
it can effectively leverage on prior geological knowledge Step 1. Collecting Training Images. The first step for estab-
in a quantitative manner and combat the problem of lishing a training image database is to collect many geo-
scarce geological data often encountered in the develop- logical cross sections from different sources and digitalize
ment of subsurface geological cross sections. It should be them using a consistent format. The potential training
noted that the performance of stochastic simulation images can be obtained from four different sources: 1)
methods for subsurface geological modeling can be the delineation of subsurface geological cross sections is a
greatly influenced by TI particularly when site-specific must for every geotechnical project, and it is, therefore,
measurements are limited, and an improper TI may lead natural to collect geological cross sections developed
to geological realizations incompatible with observed from previous projects and use them as training images;
data or even false interpretation of geological processes 2) conceptual geological models developed by engineer-
(49, 50). This underscores the importance of selecting a ing practitioners, or even hand drawn by experienced
proper TI for stochastic simulation methods. geologists, can also be used as training images; 3) a train-
Image-based stochastic simulations learn stratigraphic ing image can also be simulated using different generative
features from a single training image and leverage on the models such as object-based, process-based, and process-
extracted patterns to yield an alternative representation mimicking models, for example, Mariethoz and Caers
of subsurface stratigraphy while conditioning on available (34); and 4) training images can snapshots taken directly
site-specific data (e.g., slope outcrops or borehole logs). from experiments using small-scale models.
A training image can be viewed as a prior ensemble of
local geological knowledge and experience (e.g., inter-
relationships between soil types and orientations of soil Step 2. Categorizing Training Images. To facilitate the subse-
layer boundaries) with the required spatial scale at the site quent selection of training images for subsurface strati-
of interest. More specifically, a qualified training image is graphy, all the collected training images may be further
a numerical representation of believed stratigraphic het- classified into different categories in accordance with
erogeneities. Although it does not necessarily enclose all their geological origin, location, and application scenar-
the detailed stratigraphic features at a target site, it should ios. After determining the appropriate mode of origin or
exhaust major repetitive stratigraphic relationships and deposit type, a collected training image can be further
structures (34). Training images essentially serve as effec- categorized to different subgroups based on locations.
tive supplements to overcome challenges associated with Training images nearby are deemed to share similar local
data sparsity, which is an intrinsic issue in geotechnical depositional environments. Finally, training images col-
site investigation and geological modeling. lected from similar application scenarios (e.g., slope sta-
Shi and Wang (36) proposed a framework for the condi- bility analysis) should be grouped together as different
tional simulation of subsurface stratigraphy, based on the application scenarios might focus on the accurate deli-
typical cross sections for weathered granite and tuff slopes neation of different stratigraphic patterns.
in Hong Kong. Subsequently, the proposed framework was
applied to delineate subsurface stratigraphy and quantify Step 3. Selecting Training Images. In the absence of training
associated stratigraphic uncertainty using real slope cross image databases, a candidate training image to be obtained
sections in Hong Kong. In the context of this framework, from nearby sites or projects with similar geological set-
the inputs are the images of the geological cross sections. tings can be adopted, which has achieved preliminary suc-
The geological cross section predicted by this framework is cess in site planning and the appraisal of subsurface
dependent on both the training image adopted (i.e., a prior stratigraphy (47). However, the procedure for identifying a
geological model) and site-specific measurements (i.e., qualified training image can be greatly simplified when a
Yousefpour et al 5

suitable training image database is available. For a specific experience (51). Although linear interpolation may be
site with only slope outcrops available, a potential training conventionally used to develop subsurface geological
image may be obtained directly from a training image cate- cross sections from limited scattered data, a follow-up
gory, which shares a similar geological origin, location, design or analysis based on this deterministic interpreta-
and application scenario with the concerned slope. tion method might be considered a poor decision (e.g.,
Therefore, a compatible training image category can read- Scheidt et al. [52]), particularly when the stratigraphic
ily be set up by incorporating cross sections that are col- uncertainty is prevailing. To overcome the shortcomings
lected from nearby sites. Note that geological processes are of the linear interpolation, many techniques and methods
invaluable when attempting to compile a database of train- have been developed to describe, simulate, and model
ing images and the geological process has been included in strata, such as the octree model (53), B-rep model (54),
geological origins. geochron concepts (55), and tri-prism model (56).
However, these methods rely on the guidance of expert
Step 4. Ensemble Learning Subsurface Stratigraphy Using knowledge and experience in the selection of assump-
Training Image Database. Each candidate training image tions, parameters, and data interpolation methods,
can be viewed as an eigen-pattern set of the subsurface which are subjective and limited (57). Assumptions about
system and only represents a specific geological config- the borehole data distribution must be made, and it is
uration under a given geological origin and application difficult to evaluate the stratum simulation results
scenario. The combination of multiple training images effectively.
can be considered an ‘‘orthogonal decomposition’’ of the The recent advance in emerging machine learning
subsurface system and enables a comprehensive apprai- methods provides a fresh perspective on the development
sal of subsurface geological patterns and stratigraphic of subsurface geological cross sections. For example,
uncertainty. Ensemble learning bypasses the selection of Porwal et al. (58) used radial function and neural net-
a single best prior geological model or training image for work to evaluate potential maps in mineral exploration.
subsurface stratigraphy but combines diverse strati- Zhang et al. (59) predicted karst collapse based on the
graphic patterns from multiple prior geological models Gaussian process. Rodriguez-Galiano et al. (60) con-
for a unit characterization of stratigraphic uncertainty. ducted a study on mineral exploration based on a deci-
This is particularly important for developing subsurface sion tree. Gaurav (61) combined machine learning,
geological cross sections when only limited site-specific pattern recognition, and multivariate geostatistics to esti-
data are available, and it further emphasizes the necessity mate the final recoverable shale gas volume. Sha et al.
of establishing a training image database for subsurface (62) used a convolutional neural network to characterize
stratigraphy. unfavorable geological bodies and surface issues. It is
noted that although the spatial distribution of strata can
be characterized effectively with these approaches, the
Step 5. Simulating Subsurface Stratigraphy and Quantifying
stratigraphic uncertainty is often ignored.
Stratigraphic Uncertainty. The subsurface stratigraphic con-
To this end, various machine learning methods, such
figuration can be simulated using the image-based sto-
as the Bayesian compressive sampling (63), the artificial
chastic conditional methods, such as the multiple-point
statistics (34) and iterative convolutional XGBoost algo- neural network (64) and the multilayer perceptron neural
rithm (47). Essentially, the image-based stochastic condi- network (65), have been advanced. These machine learn-
tional method relies on a flexible data event template to ing methods are appealing to practical engineers as they
retrieve compatible stratigraphic patterns from a limited can effectively combine limited site-specific measure-
set of training images for establishing cumulative distri- ments and prior geological knowledge. Note that the
bution function (CDF) curves, which are subsequently stratigraphy’s interpolation accuracy mainly depends on
used for sampling and determination of soil types at the number of measurements collected. In addition, to
unsampled locations. Then, a geological cross section or characterize the stratigraphic configuration and associ-
realization is completed. Based on the simulated strati- ated uncertainty, the geostatistical methods, Markov-
graphic configurations, the stratigraphic uncertainty can based simulation methods (e.g., Markov random field
be quantified using the information entropy. and coupled Markov chain) (66–69) and conditional ran-
dom field-based simulation methods (29, 70), have been
developed to derive probabilistic stratigraphic relation-
Borehole Interpretation ships between observed data for spatial interpolation of
The subsurface stratigraphic configuration at a project soil boundaries. The successful applications of those
site is usually obtained through spatial interpolation of methods rely heavily on the accurate estimation of tran-
the site-specific measurements (e.g., boreholes or cone sition probabilities or spatial correlation. Directly esti-
penetration tests), coupled with local geological mating transition probabilities or spatial correlation
6 Transportation Research Record 00(0)

from site-specific measurements can be complex as mea- the initial fields can be sampled independently on each
surements are usually sparse and limited. element solely via the DANN-KHMD classifier.
Note that the stratigraphic modeling with multi-source
data fusion is expected to reduce the influence of the
measurement error and then improve the simulation Step 3. Conducting Gibbs Sampling and Updating Model
accuracy of the stratigraphy. Xiao et al. (71) proposed a Parameters. 1) Define a prior distribution of model para-
coupled machine learning method to integrate the bore- meters (i.e., the contextual constraint, b) via a multivari-
hole and CPTU data under a rigorous Bayesian frame- ate Gaussian distribution with a mean vector and a
work and to identify and separate the noisy CPTU data diagonal covariate matrix. 2) Provide an initial guess of
without subjective judgment, which contributes to more model parameters b, and an initial stratigraphic config-
reliable soil classification and property evaluation. uration. 3) Given the current model parameters b and
Wei and Wang (72) developed a novel stratigraphic the current stratigraphic configuration, calculate the con-
uncertainty quantification approach by integrating the ditional probability of all unknown elements. 4)
Markov random field theory and the discriminant adap- Generate an updated stratigraphic configuration via the
tive nearest neighbor–based k-harmonic mean distance chromatic sampler according to the conditional prob-
classifier into a Bayesian framework. The inputs of this ability acquired. 5) Update the model parameters b using
approach are the stratigraphies collected at borehole prior distribution of b and the likelihood function. 6)
locations. And the number of the required boreholes Iterate 3) to 5) until the specific convergence criterion is
may be dependent on the stratigraphic structure. For met. This is a single simulation of the stratigraphic
example, more boreholes may be required for the strati- configuration.
graphic modeling at the site with the complicated strati-
graphic structure (e.g., the fold stratigraphic structure) Step 4. Quantifying Stratigraphic Uncertainty. Generate multi-
than that with the simple stratigraphic structure (e.g., the ple initial stratigraphic configurations from Step 2 and
horizontally layered stratigraphic structure). This new execute Step 3. Based on the simulated stratigraphic con-
approach has the following advantages: 1) inferring stra- figurations, the stratigraphic uncertainty can be quanti-
tigraphic profile and associated uncertainty in an auto- fied using the information entropy.
matic and fully unsupervised manner; 2) reasonable
initial stratigraphic configurations can be sampled and
therefore lower the computational cost; 3) both strati- Characterization of Geomaterials
graphic uncertainty and model uncertainty are taken into
consideration throughout the inferential process; 4) rely- Many parameters characterizing the properties of geo-
ing on no training stratigraphy images. The main proce- materials, for example, index and strength parameters of
dures of the proposed method may be summarized as soils, are intercorrelated. Many correlations of geomater-
follows (this method has been implemented in Python ials were established in the early decades of soil and rock
3.7.) Interested audiences may contact the corresponding mechanics. They had been verified by sufficient research-
author of Wei and Wang (72) for the in-house developed ers with wide practical geotechnical and rock engineering
Python package ‘‘PyMRF.’’ applications. Geomaterials are rarely homogeneous by
nature and may vary spatially because of complex geolo-
gical processes, which motivates geotechnical and rock
Step 1. Collecting Borehole Data and Classifying the engineers to update empirical correlations once more
Stratigraphy. The first step is to collect the stratum infor- data are collected. An example is the CPTU correlations
mation, which can be revealed through borehole explora- for Norwegian clays established by Karlsrud et al. (73)
tion or directly observable from the ground surface, which have been updated by Paniagua et al. (74) using
outcrops, or both. Then, the borehole stratigraphy more advanced multiple regression methods based on a
should be classified based on the borehole data collected. database of 61 block sample data points and CPTU mea-
surements. There is, therefore, always a need for better
Step 2. Sampling an Initial Field Using the DANN-KHMD understanding of the behavior of soils and rocks to
Classifier. For generating reasonable initial fields, the dis- improve geotechnical design.
criminant adaptive nearest neighbor–based k-harmonic Unlike statistical analyses, ML algorithms are able to
mean distance (DANN-KHMD) classifier is developed learn the association between geotechnical design para-
to label the unknown (non-borehole) elements using meters (e.g., undrained shear strength) and index para-
long-range spatial patterns learned from known (bore- meters without necessarily assuming a structural model
hole) elements. It is essentially an approach to roughly in the data. Given that a large quantity of data has been
‘‘guess’’ possible labels of the unknown elements given collected and stored by the rapid advancement in digital
known elements in a probabilistic manner. Accordingly, technology over recent years, ML has been widely used
Yousefpour et al 7

Table 2. Case Studies Recently Published by Some of the Most Prolific Authors Adopting ML for Site- or Region-Specific Analyses,
Comparing the Performance of Different ML Algorithms

Study ML algorithms Predicted parameters Region

Pham et al. (76) PANFIS, GANFIS, SVR, ANN Undrained shear strength of soft Nhat Tan and Cua Dai bridges in
soils Viet Nam
Nguyen et al. (77) BPNN Friction angle Da Nang–Quang Ngai
expressway project, Vietnam
Mittal et al. (78) MLR, ANN, SVR, ANFIS Coefficient of consolidation Ha Noi–Hai Phong highway
project, Vietnam
Saedi and Mohammadi (79) ANN Unconfined Compression Sanandaj–Sirjan, Takab,
Strength (UCS) and Hamedan, Borujerd, and
Stiffness (E) Neyriz sites in Iran
Li et al. (80) RF, ANN Permeability, soil-water Singapore
Characteristic Curve (SWCC)
Lin et al. (81) ANN Shear strength and Zhuhai, China
compressibility
Liu et al. (28) ANN, RF, SVM Modulus of compression and Shen Zhen, China
coefficient of compressibility
Niyogi et al. (82) SVM, RF, DNN Lateritic soil shear strength The Ratnagiri–Sangameshwar
section of National Highway
66 in Maharashtra, India

Note: ML = machine learning; PANFIS = particle swarm optimization—adaptive network based fuzzy inference system; GANFIS = genetic algorithm—
adaptive network based fuzzy inference system; SVR = support vector regression; ANN = artificial neural network; BPNN = backpropagation neural
network; MLR = multiple linear regression; ANFIS = adaptive network based fuzzy inference system; RF = random forest; SVM = support vector machine;
DNN = deep neural network.

to characterize complex behaviors of geomaterials (ANN) for predicting the undrained shear strength of
because of its strong nonlinear fitting capability. soft soils, collected from two bridge projects in Vietnam.
A recent overview of the application of ML algo- Nguyen et al. (77) developed a backpropagation neural
rithms to the prediction of soil properties in the past 10 network (BPNN) machine learning model to predict the
years was presented by Zhang et al. (75). The implemen- internal friction angle of the soil based on 145 soil sam-
tation of ML techniques has shown exponential growth ples collected from Da Nang-Quang Ngai expressway
since 2018. Six classical ML algorithms were compared project in Vietnam. With soil samples from another
in their study, namely genetic programming (GP), evo- region in Vietnam, Mittal et al. (78) developed ML mod-
loutionary polynomial regression (EPR), support vector els, that is, multiple linear regression (MLR), ANN,
regression (SVR), RF, feed-forward neural network SVR, and adaptive network based fuzzy inference system
(FFNN) and Monte Carlo dropout-based artificial (ANFIS), to predict the coefficient of consolidation in
neural network (ANN_MCD). However, important the soil. Saedi and Mohammadi (79) investigated the
challenges still remain, such as site uniqueness, sparse relation between UCS (Unconfined Compression
and incomplete site-specific data, lack of a benchmark Strength) and E (Stiffness) of migmatites and microstruc-
data set, and the generalization of ML models. tural characteristics using image processing and ANN
techniques. Li et al. (80) developed a database containing
both saturated and unsaturated hydraulic and mechani-
Site- and Region-Specific Data cal soil properties in Singapore to predict unknown para-
A literature review is presented in this section, mainly meters, for example, permeability and soil-water
discussing the most recent studies, within which a com- characteristic curve (SWCC), using RF and ANN. Lin
parison among different ML techniques has been per- et al. (81) developed an ANN model to map the shear
formed. Table 2 shows, for each article referenced, the strength and compressibility of soft soils based on a
list of ML algorithms adopted, the predicted parameters, database consisting of over 2,000 sets of physical and
and the location and area of the case studies. Pham et al. mechanical properties for soft soils in Zhuhai city,
(76) compared the performance of four machine learning China. Liu et al. (28) developed three ML models—
methods: particle swarm optimization—adaptive net- ANN, RF, and SVM—to map the two compressibility
work based fuzzy inference system (PANFIS), genetic indices based on a database consisting of 743 sets of
algorithm—adaptive network based fuzzy inference sys- physical properties and corresponding compression
tem (GANFIS), SVR, and artificial neural networks indices for soft soils in Shenzhen, China. Niyogi et al.
8 Transportation Research Record 00(0)

(82) assessed three machine learning–based approaches, et al. (76) concluded that out of four models the PANFIS
namely SVM, RF, and deep neural network (DNN), for emerges as a promising technique for prediction of the
predicting the performance of lateritic soil shear strength strength of soft soils. The ANN was the best model in
based on soil samples collected along the Ratnagiri– Liu et al. (28), as it provided a simple analytical form
Sangameshwar section of National Highway 66 in with no hidden dependency between the bias and pre-
Maharashtra, India. They concluded that the DNN dicted indices. While building and testing the ML mod-
model has the highest prediction accuracy for the resi- els, the entire data set is usually split into training and
dual soil shear strength among the three distinct pro- testing data sets. The training data set could be divided
posed ML models. further into training and validating data sets, or the
Different authors have employed different operational cross-validation techniques could be applied to a limited
procedures to move from the construction of the data- data sample, without further splitting of the training data
base needed to feed the ML algorithms, to the prediction set.
of the properties of geomaterials, and to the performance
evaluation of the computational model. Three main
common phases of analysis may be recognized in each Step 3. Testing. The ML models established in Step 2 need
procedure: 1) data preprocessing; 2) model building and further evaluation of their performance against unseen
validation; and 3) testing. data sets. The common performance metrics that are
typically adopted in the literature include:
Step 1. Data Preprocessing. Data preprocessing facilitates
 expressions quantifying the error of the analysis
the training process by appropriately transforming the
entire training data set to remove outliers, produce the by means of an objective function (OF), the coeffi-
optimal set of input variables (features), and normalize cient of determination (R2), the mean absolute
different features to an equivalent range, which can be error (MAE), and the root mean square error
used to build an ML model. It is not surprising that out- (RMSE);
lier detection is not very often carried out as only a lim-  the model bias method using bias mean, bias coef-
ited number of data points are available in most of the ficient of variation (COV), and bias probability
studies. Moreover, outliers can be hard to define for geo- distribution (81).
properties. Among the few studies, Li and Misra (83)
used isolation forest to remove outliers of compressional
and shear travel time logs (DTC and DTS) acquired Generic and Benchmark Data Sets
using sonic logging tools. Correlation analysis and fea- Site-specific data in geotechnical engineering are gener-
ture selection is another important step in data prepro- ally limited (from several hundred to a thousand at
cessing. It is vital to remove highly correlated input most). Therefore, data samples collected from various
features and irrelevant features for the prediction of sources or places were used to develop ML models. It is
physical and mechanical properties. Correlation coeffi- worth mentioning that the database compilation by
cients are widely used to measure the correlation between ISSMGE TC304 (304 dB) (http://140.112.12.21/issmge/
features, for example, the Pearson correlation coefficient tc304.htm) could be a suitable platform for using open
and Spearman correlation coefficient. Various feature data sets. Asghari et al. (85) and Zhang et al. (6) used the
selection techniques can be used for removing irrelevant 304 dB to investigate the application of ML methods for
features, such as least absolute shrinkage and selection the prediction of the undrained shear strength of soft
operator algorithm (LASSO), random forests—recursive soils and other complex correlations in engineering
feature elimination (RF-RFE), and mutual information metrics. Ma et al. (86) developed hybrid GA-SVM and
have been applied by Mittal et al. (78). SHapley Additive PSO-SVM models for the prediction of permeability of
exPlanations (SHAP) (84) was used by Li et al. (80) to cracked rock based on a database developed from exist-
investigate the impact of each input variable on different ing literature. Zhang et al. (87) investigated the perfor-
output soil properties. These indices could be used to mance of five commonly used ML algorithms, that is,
validate the results, enabling researchers to explore the backpropagation neural network (BPNN), extreme
algorithm’s logic and verify its reliability. learning machine (ELM), SVM, RF, and evolutionary
polynomial regression (EPR) in predicting compression
Step 2. Model Building and Validation. As indicated in Table index Cc based on a global database consisting of 311
2, most studies used conventional ML techniques, for data points of initial void ratio e0, liquid limit water con-
example, ANN, SVM and tree-based methods. There is tent wL, and plasticity index Ip. Chen and Xue (88) col-
no consensus on a specific ‘‘optimal’’ ML algorithm for lected a total of 151 data sets from the literature that
predicting physical properties of geomaterials. Pham were used to construct the ML models.
Yousefpour et al 9

integrate the geotechnical and geophysical data collected

from a site and provide a three-dimensional map of the
stratigraphy and the geo-properties.
In geotechnical projects, the quantitative integrated
ground model is a requirement for cost-optimal site char-
acterization. The ground model refers mainly to the strati-
graphic configuration and the associated geotechnical
properties in this study. Extensive studies focusing on the
stratigraphy classification or the geomechanical characteri-
zation have been reported in the geotechnical literature
(e.g., Mariethoz and Caers [34], Shi and Wang [47], and
Wei and Wang [72]). However, relatively limited studies
dealt with both the stratigraphic configuration and the
geomechanical parameters in a specific model (92). Note
that both the strata and the geo-properties at a site are a
Figure 2. Predicted versus measured cone penetration test product of the same deposit histories, tectonic, and human
(CPT) undrained shear strength (in kPa) over the test data set. activities; their spatial distributions are expected to share
The vertical line shows the estimated standard deviation. The similar features. Thus, the study on the site characteriza-
black dashed line is the 1:1 line. tion that considers both the stratigraphy classification and
the geomechanical characterization, especially the spatial
variabilities of the stratigraphic configuration and the
Uncertainty Quantification associated geotechnical properties, is warranted.
Uncertainties associated with geomaterials (soils, rocks), In the conventional engineering approach, the sub-
geologic processes, and possible subsequent treatments surface stratigraphic configuration at a site is usually
are usually large and complex. Current ML modeling is constructed through spatial interpolation of the strati-
always deterministic. As such, only high-quality data can graphy collected at borehole locations, coupled with
be used because there is no capacity to address uncertain- local experience (51). The geo-properties of each stra-
ties. Zhang et al. (89) pointed out that applying the pre- tum are taken as fixed values (e.g., the mean of the
dicted result without any reliability evaluation using ML- geo-properties derived at borehole locations) (93). The
based models may induce high risk. A Bayesian neural ground model constructed with the conventional
network (BNN) integrated with variational inference (VI) approach simplifies the actual geological body.
and Monte Carlo dropout (MCD) was used in their study Although this conventional approach has been widely
to predict the compression index Cc and undrained shear adopted in practice, no scientific rationale is available
strength su of clays and to evaluate the reliability of the to support this simplification (92).
ML model. An ANN-based model that takes into account Geotechnical analyses and interpretations often rely
uncertainty estimates was developed in (90). The database on isolated 1-D boreholes. On the other hand, geophysi-
of CPTU and triaxial tests used contained 241 laboratory cal data are collected in 2-D lines and 3-D volumes.
triaxial tests. All test specimens were of high quality, Geophysical data therefore provide the natural link to
including 180 undisturbed specimens taken with a 72-mm repopulate geotechnical properties found in the 1-D
diameter fixed piston sampler and 61 undisturbed speci- boreholes onto a larger area and thereby build a consis-
mens taken with 40-cm diameter block samples. The tent and robust ground model (94). Thus, it is current
ANN model requires the water pressure and effective practice to conduct both geophysical and geotechnical
stresses in addition to the CPTU data. The standard investigations for large geotechnical projects on land and
deviation estimate was performed based on the ‘‘dropout offshore developments. Ground models integrate the
method’’ described in Gal (91). The predicted and mea- geotechnical and geophysical data collected from a site
sured undrained shear strength in triaxial compression are and provide a 3-D map of the stratigraphy and the geo-
compared in Figure 2. An estimated standard deviation of properties. Many approaches, such as the geometrical
the predicted undrained shear strength is also shown. approach (94), the geostatistical approach (95), and the
ANNs (96) have been reported to map the dynamic prop-
erties from the seismic data (stratigraphic information, P-
3-D Integrated Ground Model
wave velocities, amplitudes, and their attributes) into the
For large geotechnical projects on land and offshore geotechnical or geomechanical properties. Sauvin et al.
developments, it is current practice to conduct both geo- (94) showed that the machine learning approach provides
physical and geotechnical investigations. Ground models the most accurate prediction of the ground model from
10 Transportation Research Record 00(0)

the geophysical data, over the geometrical geostatistical developing 2-D subsurface geological cross sections from
approach and the geostatistical approach. training images and limited site-specific measurements.
The implementation procedures of the machine learn- A training image reflects prior geological knowledge at
ing approach presented in Mittal et al. (78) can be sum- the area of interest and serves as an effective supplement
marized as follows. to limited site-specific data (99).

Step 1. Derive a range of quantitative attributes from

the seismic reflection data, particularly acoustic impe-
Step 3. Modeling Soil Property Spatial Variability from
dance using a genetic algorithm for optimization. Limited Measurements
Step 2. Convert the seismic attributes from the time Bayesian compressive sampling (BCS) is a non-
domain into the depth domain. parametric machine learning method developed for inter-
Step 3. Down-sample the piezocone penetration test polating spatially varying geo-properties (e.g., cone pres-
(CPTU) tip resistance (qc) to a sample interval match- sure) from sparse measurements. Under compressive
ing that of the depth-converted seismic attributes. sensing/sampling, a complete signal (e.g., 2-D spatially
Step 4. Use an ANN to perform multi-attribute regres- varying soil property profiles) can be approximated as a
sion between the range of quantitative seismic attri- weighted summation of a limited number of pre-specified
butes and CPTU qc by training at several calibration basis functions (e.g., discrete cosine basis functions).
sites.

It is worth noting that above approaches focused on Step 4. Sequential Modeling of Soil Property Spatial
deriving the most-probable ground model; however, the Variability for Each Soil Type
spatial variabilities of the stratigraphic configuration and Once subsurface geological cross sections or realizations
the associated geotechnical properties were generally are developed using IC-XGBoost2D, the best estimate of
ignored. To overcome this obstacle, Shi and Wang (97) cone pressure, sleeve friction, and pore pressure within
proposed a stochastic framework for modeling the strati- the 2-D cross section with high spatial resolution can be
graphic uncertainty and spatial variability of soil proper- obtained using BCS and CPT measurements.
ties by machine learning and random field simulation
from limited site investigation data. This framework
could effectively generate multiple realizations of geolo- Scour Assessment Using ML
gical cross-section and random field samples of geotech-
Various recent studies have looked into applications of
nical properties from limited measurements (obtained
AI-ML in geohazard assessment and management for
from the CPT), through which the uncertainties associ-
geotechnical systems (100, 101). In this paper, we focus
ated with the ground model can be characterized. We
on scour as one of the most critical mechanisms affecting
recommend the proposed framework for characterizing
bridge foundations and how ML techniques can be
the ground model and associated uncertainties.
implemented step by step to provide more reliable bridge
scour estimation and real-time risk assessment for design
Step 1. CPT-Based Soil Classification and Interpretation and maintenance purposes, respectively.
of Consolidation Parameters In the past two decades, numerous studies have
explored local scour prediction around bridge piers
The cone penetration test (CPT) is a commonly used in using ML. SVM, genetic algorithms (102, 103), and
situ testing method for soil classification and characteri- artificial neural networks (ANNs), particularly FFNN
zation of subsurface geotechnical property profiles. CPT or multilayer perceptrons (MLP) (104–107) are among
provides direct continuous vertical line measurements of the most commonly used techniques. These ML meth-
cone pressure, sleeve friction, and pore pressure. Apart ods have shown superior performance over traditional
from soil classification, CPT data can also be used to empirical equations in the accuracy of maximum scour
estimate the soil properties of fine-grained materials via depth predictions for a given training data set; however,
empirical correlations established in the literature (e.g., the generalization ability of these types of predictive
Robertson and Cabal [98]). models decays significantly outside the convex hull of
the training data set, which can lead to poor generaliza-
tion to unseen data (108). Therefore, to use these mod-
Step 2. Stratigraphic Uncertainty Modeling by IC-
els for maximum scour depth prediction, the bridge
XGBoost2D characteristics must be within the training database
The iterative convolution extreme gradient boosting (IC- range. A good review of the literature can be found in
XGBoost2D) is a stochastic simulation algorithm for Sharafati et al. (109).
Yousefpour et al 11

Given the limitations of the current scour prediction levels can result in different reported scour depth
models, the complexity of scour phenomenon, uncer- in field measurement data. For more information,
tainty in flow (flood levels), riverbed and geomorpholo- refer to Landers and Mueller (1).
gical conditions, real-time monitoring, and forecast to 4) ML knowledge and implementation complexities:
manage the scour risk is evolving as a promising tool. Lack of ML (and coding) knowledge and experi-
Yousefpour et al. (107, 108) have pioneered this ence in practice engineers, along with a lack of
approach by using historical monitoring data from guidelines, hinder AI-ML application for bridge
bridge piers, including timeseries of scour depth and river scour design.
flow depth variation. In their approach, they have used
both DL and Bayesian inference methods that have The following methodology is proposed to develop a
shown reasonable accuracy in providing real-time assess- simple ML model to estimate the maximum scour depth
ment of scour depth. In their most recent study, they with a confidence bound for a new bridge. For more
have developed long short-term memory (LSTM) models details on the approach, readers are referred to
that can provide estimates of scour depth a week in Yousefpour et al. (107). Engineering due diligence and
advance for case study bridges in Alaska (108). quality control checks must be taken when applying this
Based on a critical review of the existing methods, method for design:
the following techniques are recommended for maxi-
mum (design) and real-time scour depth prediction
Step 1. Database Compilation. Using the USGS National
(maintenance).
Bridge Scour Database, a ‘‘global’’ ML model can be
trained first. This model needs to be later retrained
Maximum Scour Depth Prediction Using Feed-Forward (transfer learning as explained in step 2) using a ‘‘local’’
or site-specific database with a well-designed statistical
Neural Networks
distribution. The local database should include data
Fully connected FFNNs or MLPs are powerful in repre- from several bridges with relatively similar riverbeds,
senting nonlinear high dimensional processes influenced flow, and structural characteristics. The key input/target
by multiple physical factors. MLPs are universal approx- parameters for the scour ML model are listed below, as
imators theoretically capable of approximating any func- identified in many earlier studies. These parameters for
tion, even with only one hidden layer with enough the new bridge should lie in the convex hull of the local
nonlinear computational units (neurons) (110, 111). database for a reliable scour prediction:
The application of FFNNs in developing scour pre- Input parameters (features)
diction models has been explored by numerous studies,
as referenced in the previous section. Many used the US  Sediment transport (scour type): This can be
National Scour Study database developed by the US divided into live-bed or clear-bed. In clear-water
Geological Survey (USGS) in collaboration with the conditions, there is limited transport of bed mate-
Federal Highway Administration (FHWA) (1, 2). rial into the channel from upstream, whereas, in
However, these models can hardly be used in mainstream live-bed conditions, materials are transported
scour design. The major obstacles in upscaling such ML through upstream flow and deposited downstream
models to practice are: (filling after scour). For more details on the defini-
tion, refer to Landers and Mueller (1) and Mueller
1) Poor generalization/extrapolation capacity: This and Wagner (2).
is a result of the dependency of ML models on the  Flow depth: This should be measured using
range of training data, meaning ML models need gauges, stage sensors, or other techniques. For
to be trained with a database that is statistically best practice in the measurement of such para-
representative of a particular bridge. meters, readers are referred to Mueller and
2) Database deficiencies and poor statistical design: Wagner (2).
Databases are built with stitching scattered data  Flow velocity: Acoustic doppler velocimeters
sets without much statistical design, involving (ADV) can measure average velocity around piers.
many variables that show sparse range across the For best practice in velocity measurements, read-
various data sets. ers are referred to Landers and Mueller (1).
3) Errors and subjectivity in measurements: This  Pier/pile geometry (shape, skew, width, and
issue with scour depth measurement requires a length): This is available in national databases
judgment on the reference surface. Different judg- (US National Bridge Inventory [NBI]) and can be
ments/assumptions on reference levels such as measured on site or extracted from bridge plans/
ambient bathymetric, as-built, or maximum bed drawings.
12 Transportation Research Record 00(0)

 Bed condition: This determines the cohesive or

non-cohesive nature of sediments around the
bridge.
 Sediment size: Median grain size (D50) can be mea-
sured from sampling at the site or from bridge
design site investigation reports. The common
sampling methods are explained in Landers and
Mueller (1).

Target parameter (output)

 Scour depth: This should be measured after a

flood event—within 24 to 48 h in a live-bed scour
situation (mostly observed in non-cohesive river
beds); this allows reaching to equilibrium scour
depth. For clear-bed scour the time-dependency
of scour needs to be established to determine the Figure 3. Simple architecture for MLP/FNN (3).
suitable time of scour depth measurements. Note: MLP = multilayer perceptron; FNN = feed-forward neural network.

The key point in this step is to develop a statistically

well-designed database that can be representative of the stopping based on a ‘‘patience number,’’ which is the
population of bridges in a region of interest. The impor- maximum consecutive number of times to allow the error
tance of the statistical design of databases is seemingly to increase over a validation data set during training
overlooked in the current literature, especially given the (113, 114).
surge of featureless DL approaches. By referring to core The generalization of ML models is assessed based on
statistical knowledge, we know that the skewness of a a test data set, which represents the model performance
database (especially for smaller data sets) can lead to the on unseen data. Several metrics can be used for this pur-
model favoring specific parameters and trends in the pose, including RMSE, coefficient of determination (R2),
data (bias), which results in unreliable predictions. To MSE, and MAE. MAE and RMSE allow error measure-
have a well-designed database, the statistical distribution ment in the same unit as the target parameter.
of all parameters in the database should be analyzed to Transfer learning has been proven to improve the per-
ensure that there is a relatively uniform distribution formance of ML models significantly (115). In this
across the range of variability. The data should be nor- method, instead of training an ML model from scratch,
malized before training the ML model. that is, starting from a random point in the space of
weights and biases, the model is initialized using a set of
parameters from a previously trained model. In the con-
Step 2. Developing a ‘‘Simple Enough’’ ML Model. A simple text of image data and pattern recognition, this can have
MLP (FFNN) network is shown in Figure 3. A three- a tremendous effect; therefore, using pre-trained models
layer architecture, with one hidden layer applying a sig- is a must. In the context of scour prediction, this can
moid activation function such as the ones explained in mean ‘‘transferring the learning’’ from a model trained
Yousefpour et al. (112) can be followed. For regression using a global database of scour to a site-specific ML
problems, using more than one layer and too many units model that needs to be trained with a local database.
often does not lead to more accurate predictions. For For this purpose, first a ‘‘global ML’’ model can be
theoretical details of MLP algorithms, readers are trained according to the recommendations in step 1, then
encouraged to refer to Hornik (110). Training in MLPs a ‘‘local ML’’ can be trained, initializing from the set of
means adjusting the weights and biases to reach the uni- weights and biases of the trained global model.
versal minimum of a loss function, which is usually
defined as the mean of squared prediction errors (MSE)
across the training database. However, more often than Step 3. Validation and Fine-Tuning. Using an ensemble of
not, optimization techniques find a local minimum models provides more accuracy in predictions and
instead of the global minimum. Several techniques can enables uncertainty assessment. An ensemble of FFNN
be implemented to avoid being trapped in a local mini- can be generated by repeating (random initiation of net-
mum, such as regularization, cross-validation, and early work weights and biases) the training over many times
stopping, among others. A practical method is early ( + 50) to find a group (ensemble) of best models. The
Yousefpour et al 13

Figure 4. Cross-validation methods: (a) K-Fold; and (b) Monte Carlo.

ensemble’s performance can be reported by using first- Step 4. Assessment of ML Predictions. The ensemble of the
order statistics of the resulting distributions. Readers are best FFNNs generates a distribution of predictions on
referred to Yousefpour (3), Bateni et al. (104), and Sagi scour depth for a given bridge. Having a larger ensemble
and Rokach (116) for more details on ensemble methods. ensures the uncertainty is better captured. Mean, stan-
One of the most effective cross-validation techniques dard deviation, and 95% confidence bound can be
that also can be integrated with the ensemble method, is reported from the generated distribution. We recom-
to use K-fold or Monte Carlo cross-validation (117) (as mend scatter-plotting the scour depth versus velocity and
opposed to the hold-out method) to enable making the flow depth for all the records in the local database and
most out of a small database. In this technique, the net- locating where the ML prediction lies within the local
works in the ensemble use various parts of the data for and global database. The ultimate judgment on the scour
training, validation, and testing, as shown in Figure 4. depth estimates for a given bridge at a target flood level
Fine-tuning and hyperparameter optimization can (e.g., 200-year return period), needs to be based on both
lead to superior network configurations. The main hyper- the ML prediction and the historic local trends observed
parameters for an FFNN model include the number of from past flood events.
hidden layers, the number of units in each hidden layer,
the type of activation function in units, input feature
selection/combination, optimization, and training para-
meters such as learning rate, maximum number of Real-Time Scour Assessment Using Deep Learning
epochs, and early stopping criteria (patience number, Regular inspection and monitoring of the scour in exist-
minimum error, etc.). The most commonly used hyper- ing bridges has been mandatory in the United States and
parameter search techniques are grid-search and random some other countries because of frequent catastrophic
search, which can be applied to speed up the optimization events in the past few decades caused by scour failure of
process (108, 118). Several available codes and libraries bridges. Scour is the most common cause of failure for
in Python can be readily applied, such as Optuna and bridges, with approximately 260,000 scour critical bridges
Talos (119). across the country (120).
14 Transportation Research Record 00(0)

Figure 5. Sensors mounted on bridge piers in California and Alaska for continuous measurement of river water and bed elevation: (a)
solar-powered stage; and (b) sonar.
Source: Photo courtesy of Alaska USGS and Renesys.

Given the shortcomings in scour depth estimation with Step 1. Collecting Site-Specific Historic Scour Data: Continuous
existing methods, many bridges are highly vulnerable to Measurement of Bed and Flow Level (and Velocity) Variation.. A
scour. The US NBI has identified the scour risk of all review of existing scour field monitoring data can be
bridges in the United States with a ‘‘scour-vulnerability’’ found in Hunt (123). Despite the challenges of sensor
index (121). This index is assessed based on regular scour readings, which may result in missing data across a year,
inspection and monitoring mandated by FHWA. such as sensor damage resulting from flooding, accumu-
Real-time monitoring for bridges exposed to a higher lation of trash and debris, and vandalism, they can pro-
risk of scour can enable reliable risk management and a vide valuable information about the scour process for a
data-driven understanding of scour process. Several US particular bridge. The generated timeseries data of flow
Departments of Transportation (DOT) in partnership depth and scour depth enable site-specific scour assess-
with USGS have started rigorous scour monitoring pro- ment, which can be powerful in addressing the ‘‘extrapo-
grams for their most vulnerable bridges. States such as lation’’ problem of ML models discussed in the previous
Alaska, with most bridges exposed to seasonal flooding, section.
started this program in the early 2000s (available at Any type of sensor that can continuously measure
Alaska Science Center). Other states such as Idaho, flow and bed levels can be instrumental. The measure-
Colorado, and Oregon, have followed similar approaches. ment of bed elevation measurement using sonar sensors
This monitoring program allows for real-time measure- is well explained in Henneberg (124). Current methods
ment of bed elevation and flow level at a bridge pier using for measurement of flow level or stage are explained in
sonic/echosounder devices (see Figure 5) that measure the Sauer and Turnipseed (125). Flow velocity measurements
distance based on the sonic wave travel time within a spe- are also recommended, which can be measured by acous-
cific material (water or air). For more details, readers are tic doppler and velocimeters as described in Sauer and
referred to USGS published reports for each state (122). Turnipseed (126).
Yousefpour et al. (107, 108) studied the historic Alaska
monitoring data in their recent research to develop scour
forecast and early warning models using DL methods. Step 2. Data Processing. Data processing is a vital step
They have proposed an AI-based methodology for real- before ML development. Although DL models have
time scour assessment to enable more reliable risk assess- been proven to be powerful in addressing noise and out-
ment for critical bridges. The DL models can also provide liers in training data (108, 118, 127), performing the fol-
insights into estimating the maximum scour depth for lowing processing can enable faster learning and more
new bridges within the same region with similar flow, reliable performance:
structure, and characteristics. The proposed methodology
can be implemented for the assessment of scour in existing 1) Basic cleaning and synchronization of timeseries:
bridges as per the following steps: For scour sensor data, that could mean ensuring
Yousefpour et al 15

Figure 6. Outlier detection in sonar (scour) data, comparing

various denoising and filtering techniques.

Figure 7. Comparing various filtering/smoothing methods for

sonar timeseries data.
random errors in data are removed and ensuring
a reasonable resolution (reading intervals) for all
timeseries data and synchronizing the readings
as the Gaussian process can enable more realistic
among various sensors, such as bed elevation,
data representations (129).
flow elevation, and velocity. Down-sampling and
4) Smoothing and filtering: Moving average and
up-sampling can be performed to ensure the data
other auto-regressive methods can be quite pow-
are captured with reliable resolution. For scour,
erful for timeseries smoothing and denoising (130,
half-hourly or hourly readings are recommended
131). They have specifically been found useful for
as scour and flooding can develop within a few
sonar and stage sensor data (171). Signal process-
hours.
ing methods such as the low-pass filtering tech-
2) Outlier detection: Detecting outliers is essential as
nique can enable removing extreme frequency
it can mislead the model with seemingly correct
ranges from the data. Figure 7 shows how these
but erroneous readings. Figure 6 shows an exam-
different methods compare for sonar data from a
ple of how accurate outlier detection can affect
bridge in Alaska.
identifying the true trend of the scour data. One
of the common methods for outlier rejection is
the median absolute deviation (MAD) method. Step 3. Developing DL Models: LSTMs and CNNs. In RNNs,
Data is first normalized (standard normal) and the output of each neuron goes back to itself, and the
points with absolute values greater than 3.5 are units in the subsequent layer at each time step. This
defined as outliers within a selected time window. enables these algorithms to keep a ‘‘memory’’ of past
The length of this window is critical and needs to ‘‘activations.’’ LSTM networks are particular types of
be adjusted depending on reading intervals and RNN that use special ‘‘gates’’—input, forget, and output
expected time variations in the data. For instance, gates—to update the memory state at each step, as shown
variation in scour depth (in case of live-bed scour) in Figure 8 (132). This enables LSTMs to maintain a
can happen within a few hours to a few days. short and long-term memory of temporal patterns within
MAD is more robust than traditional standard timeseries data and incorporate them into future predic-
deviation methods and less sensitive to outliers tions. LSTMs have shown superiority compared with
and extremes. other RNNs for timeseries forecasting in various fields,
3) Missing data imputation: This is very common owing this success to their ability to overcome the prob-
for scour data resulting from blockage of sensors lem of vanishing gradients (133).
by debris accumulation or flood-related distur- For scour forecasting, Yousefpour and Correa (108)
bance, among other reasons. Therefore, applying showed that LSTM can provide far more accurate pre-
emerging imputation techniques in cases where dictions of scour compared with empirical equations. In
the available data is sparse but nevertheless valu- their case studies on Alaskan bridges, they developed
able in understanding the scour trend is essential. three variants of LSTM algorithms that can provide
Moritz and Bartz-Beyerstein (128) provide an scour prediction seven days in advance with an average
overview of univariate timeseries imputation tech- error of 0.2 to 0.35m. In a more recent study, they
niques. One of the simplest methods is linear explored applying CNNs as a more computationally effi-
interpolation and more advanced methods, such cient DL algorithm for real-time forecast of scour, and
16 Transportation Research Record 00(0)

Figure 8. Long short-term memory (LSTM) unit-showing details of input, output, and forget gates; xt = input feature at time step t; ct =
cell memory state; ht = hidden state or output at time step t; s = sigmoid activation function; tanh = hyperbolic tangent function (108).

Figure 9. Architectures of networks for scour forecast: (a) LSTM; and (b) CNN: iw = input width; lw = label (output) width of a data
slice (118).
Note: LSTM = long short-term memory; CNN = convolutional neural network.

compared CNN and LSTM performance on data from CNN architectures used in these studies for sonar time-
Alaskan bridges (118). Results showed competitive per- series prediction.
formance by temporal CNNs with significantly lower The training of DL models is done by timeseries
computational cost. Figure 9 shows the LSTM and sequences generated by slicing the whole timeseries data
Yousefpour et al 17

Figure 10. Data slicing for scour timeseries forecasting with deep learning (DL): (a) example of a data slice (sequence) and how it is fed
into a DL model; and (b) slicing the timeseries data using a sliding window, including input plus label sequence.

into smaller sequences (data slices) using a sliding win- Cross-validation methods similar to FFNNs can be
dow. This process is explained in detail in Yousefpour implemented to improve the performance of the DL
and Correa (108). The sliding window moves over the models and reduce the chances of overfitting. One crucial
database one timestep at a time. Each sequence (data difference is the temporal dependencies in timeseries
slice) has an ‘‘input’’ and a ‘‘target’’ or ‘‘label’’ segment, data, which can make data division for training a bit
as shown in Figure 10. The training is done in batches absurd. We recommend ensuring that training, valida-
containing several data slices. The network’s weights and tion, and testing are consecutive unless time is intro-
biases are adjusted like MLPs by minimizing the error of duced as an additional feature to the models. Methods
prediction over the label length of data slices. such as K-fold cross-validation can be implemented
sequentially, as shown in Figure 11. In this method, the
model is retrained at each step using transfer learning;
Step 4. Fine-Tuning and Validation of DL. To obtain the opti- the test data set in each step needs to be free of overlaps
mal configuration of these DL algorithms, hyperpara- with the previous step data to avoid evaluating the model
meter tuning must be performed, similar to what was over ‘‘seen’’ data. A hold-out test data set can be kept
recommended for FFNNs in the previous sections. One aside to evaluate the final model at the end of sequential
important hyperparameter in timeseries forecasting is the training. This metric can be compared and reported with
input and label width and their ratio. The impact of the average performance of models over test data sets
selecting the most optimal sequence length to ensure the across the sequential training steps.
temporal variations in the data are well explained in The prediction of the DL model over data slices that
Yousefpour and Correa (108) and Yousefpour and have continuous overlaps results in overlap in predic-
Pouragha (134). The label width is the length of forecast tions which will provide an uncertainty assessment
window and should be selected to provide enough time opportunity. In addition, repeating the training several
for implementing proper risk countermeasures. In case times provides a range of predictions that can be
of bridge scour, a minimum of seven days was deemed included in the set of predictions. Based on that, confi-
necessary. dence intervals and the meaning of predictions can be
18 Transportation Research Record 00(0)

Figure 11. Sequential K-fold cross-validation with transfer learning: the arrow shows transfer learning between training steps.

Figure 12. Scour forecast (seven days in advance) over a test subset of data (June to Oct, 2021) from the top five best LSTM and CNN
models (118).
Note: LSTM = long short-term memory; CNN = convolutional neural network.

established. Figure 12 shows an example of an LSTM metric for the scour forecast model performance (read
and CNN model prediction over a test subset for an more in Yousefpour and Correa [108]).
Alaskan bridge (118).
In addition to metrics of average error (MSE, MAE, Step5. Assessment of Upcoming Scour. The proposed real-
or RMSE), the DL model predictions should be evalu- time DL forecast models can provide a tool for risk
ated for trend detection, that is, peaks and troughs assessments of an upcoming scour event. The uncertainty
through fill and scour episodes. The error in prediction and reliability of the forecast must be assessed using
of peaks and troughs can provide another performance engineering judgment by also evaluating meteorological
Yousefpour et al 19

and flood forecasts and the available historical trends for reviewed the results and approved the final version of the
a particular bridge and/or similar bridges in a region. manuscript.
Lower bounds of scour depth predictions are recom-
mended to minimize risk in detected major scours. Using
Declaration of Conflicting Interests
real-time cameras or aerial photography (e.g., drones)
can provide further inputs for more reliable decision The author(s) declared no potential conflicts of interest with
making and taking the most appropriate plans of action respect to the research, authorship, and/or publication of this
article.
to manage scour failure risks.

Funding
Conclusions The author(s) disclosed receipt of the following financial sup-
port for the research, authorship, and/or publication of this
This paper discussed emerging AI-ML techniques for
article: Funding for this publication is provided by the
subsurface site characterization and geohazard assess- Norwegian Geotechnical Institute, Oslo and the University of
ments. Our review revealed that ML can be a promising Melbourne.
solution for characterizing subsurface conditions and
geomaterials. ML algorithms make more accurate pre-
dictions than traditional empirical correlations, while the ORCID iDs
interpretation and explanation of ML models need to be Negin Yousefpour https://orcid.org/0000-0002-6634-2658
regarded as a secondary objective. There is no consensus Zhongqiang Liu https://orcid.org/0000-0002-1693-5746
on an ‘‘optimal’’ ML/DL algorithm for subsurface site
characterization, even when looking at the results of the References
most recent comparative studies in the prediction of
1. Landers, M. N., and D. S. Mueller. Channel Scour at
physical and mechanical properties of geomaterials. The
Bridges in the United States. Office of Engineering
choice between adopting conventional ML (ANN, RF, Research and Development, Federal Highway Administra-
SVM, etc.) or DL algorithms primarily depends on the tion, Washington, D.C., 1996.
type and quantity of available data and if explicit expres- 2. Mueller, D. S., and C. R. Wagner. Field Observations and
sions are preferred. The uncertainty associated with sub- Evaluations of Streambed Scour at Bridges. Office of Engi-
surface features (layer boundaries, discontinuities, voids, neering Research and Development, Federal Highway
anomalies, etc.), spatial variability of geo-properties and Administration, McLean, VA, 2005.
ML modeling is another issue that needs further explora- 3. Yousefpour, N. Comparative Deterministic and Probabilis-
tion. In the assessment of geohazards, bridge scour was tic Modeling in Geotechnics: Applications to Stabilization of
the focus of this paper; nevertheless, the reviewed and Organic Soils, Determination of Unknown Foundations for
recommended techniques can be applied to other types Bridge Scour, and One-Dimensional Diffusion Processes.
PhD dissertation. Texas A&M University, 2013.
of geohazards. A step-by-step methodology for imple-
4. Yousefpour, N., and S. Fallah. Applications of Machine
menting both traditional ML and emerging DL methods Learning in Geotechnics. Proc., Civil Engineering Research
was discussed in application to prediction and real-time in Ireland Conference, Cork, 2018.
forecasting of scour for bridge foundations. The guide- 5. Zhang, W., H. Li, Y. Li, H. Liu, Y. Chen, and X. Ding.
lines and provided references can pave the way for imple- Application of Deep Learning Algorithms in Geotechnical
menting similar ML methods in practice. Some of the Engineering: A Short Critical Review. Artificial Intelligence
obstacles and challenges of applying these methods were Review, Vol. 54, No. 8, 2021, pp. 5633–5673.
also highlighted. 6. Zhang, W., C. Wu, H. Zhong, Y. Li, and L. Wang. Predic-
tion of Undrained Shear Strength Using Extreme Gradient
Boosting and Random Forest Based on Bayesian Optimi-
Acknowledgment zation. Geoscience Frontiers, Vol. 12, No. 1, 2021,
We acknowledge the contribution of Fatemeh Safari Honar in pp. 469–477.
preparing the figures for this paper. 7. Dikshit, A., B. Pradhan, and A. M. Alamri. Pathways and
Challenges of the Application of Artificial Intelligence to
Geohazards Modelling. Gondwana Research, Vol. 100,
Author Contributions 2021, pp. 290–301.
The authors confirm contribution to the paper as follows: study 8. Tehrani, F. S., M. Calvello, Z. Liu, L. Zhang, and S.
conception and design: Negin Yousefpour, Zhongqiang Liu; Lacasse. Machine Learning and Landslide Studies: Recent
data collection: Negin Yousefpour, Zhongqiang Liu, Chao Advances and Applications. Natural Hazards, Vol. 114,
Zhao; analysis and interpretation of results: Negin Yousefpour, No. 2, 2022, pp. 1197–1245.
Zhongqiang Liu, Chao Zhao; draft manuscript preparation: 9. Phoon, K. K., and W. Zhang. Future of Machine Learn-
Negin Yousefpour, Zhongqiang Liu, Chao Zhao. All authors ing in Geotechnics. Georisk: Assessment and Management
20 Transportation Research Record 00(0)

of Risk for Engineered Systems and Geohazards, Vol. 17, 24. Memarzadeh, M., and M. Pozzi. Model-Free Reinforce-
No. 1, 2023, pp. 7–22. ment Learning with Model-Based Safe Exploration: Opti-
10. Gao, J., M. Galley, and L. Li. Neural Approaches to Con- mizing Adaptive Recovery Process of Infrastructure
versational AI. Proc., 41st International ACM SIGIR Con- Systems. Structural Safety, Vol. 80, 2019, pp. 46–55.
ference on Research and Development in Information 25. Liu, C., D. Tang, H. Zhu, and Q. Nie. A Novel Predictive
Retrieval, Ann Arbor, MI, 2018. Maintenance Method based on Deep Adversarial Learning
11. Fang, W., L. Ding, P. E. Love, H. Luo, H. Li, F. Pena- in the Intelligent Manufacturing System. IEEE Access, Vol.
Mora, B. Zhong, and C. Zhou. Computer Vision Applica- 9, 2021, pp. 49557–49575.
tions in Construction Safety Assurance. Automation in Con- 26. Akinosho, T. D., L. O. Oyedele, M. Bilal, A. O. Ajayi, M.
struction, Vol. 110, 2020, p. 103013. D. Delgado, O. O. Akinade, and A. A. Ahmed. Deep
12. Brunke, L., M. Greeff, A. W. Hall, Z. Yuan, S. Zhou, J. Learning in the Construction Industry: A Review of Pres-
Panerati, and A. P. Schoellig. Safe Learning in Robotics: ent Status and Future Innovations. Journal of Building
From Learning-Based Control to Safe Reinforcement Engineering, Vol. 32, 2020, p. 101827.
Learning. Annual Review of Control, Robotics, and Autono- 27. Karami, Z., and R. Kashef. Smart Transportation Plan-
mous Systems, Vol. 5, 2022, pp. 411–444. ning: Data, Models, and Algorithms. Transportation Engi-
13. Bashar, A., M. R. Rabbani, S. Khan, and M. A. Moh’d neering, Vol. 2, 2020, p. 100013.
Ali. Data Driven Finance: A Bibliometric Review and Sci- 28. Liu, H., P. Lin, and J. Wang. Machine Learning
entific Mapping. Proc., International Conference on Data Approaches to Estimation of the Compressibility of Soft
Analytics for Business and Industry, Sakheer, Bahrain, Soils. Frontiers in Earth Science, Vol. 11, 2023, p. 1147825.
2021. 29. Gong, W., C. Zhao, C. H. Juang, H. Tang, H. Wang, and
14. Ghahramani, M., Y. Qiao, M. C. Zhou, A. O. Hagan, and J. X. Hu. Stratigraphic Uncertainty Modelling with Random
Sweeney. AI-Based Modeling and Data-Driven Evaluation Field Approach. Computers and Geotechnics, Vol. 125,
for Smart Manufacturing Processes. IEEE/CAA Journal of 2020, p. 103681.
Automatica Sinica, Vol. 7, No. 4, 2020, pp. 1026–1037. 30. L’Heureux, J. S., O. Longva, A. Steiner, L. Hansen, M. E.
15. Sarker, I. H. Deep Learning: A Comprehensive Overview Vardy, M. Vanneste, H. Haflidason, et al. Identification of
on Techniques, Taxonomy, Applications and Research Weak Layers and Their Role for the Stability of Slopes at
Directions. SN Computer Science, Vol. 2, No. 6, 2021, p. Finneidfjord, Northern Norway. Proc., Submarine Mass
420. Movements and Their Consequences: 5th International Sym-
16. Shrestha, A., and A. Mahmood. Review of Deep Learning posium, Springer, Netherlands, 2012.
Algorithms and Architectures. IEEE Access, Vol. 7, 2019, 31. Frank, T., A. L. Tertois, and J. L. Mallet. 3D-Reconstruc-
pp. 53040–53065. tion of Complex Geological Interfaces from Irregularly
17. Rawat, W., and Z. Wang. Deep Convolutional Neural Net- Distributed and Noisy Point Data. Computers and Geos-
works for Image Classification: A Comprehensive Review. ciences, Vol. 33, No. 7, 2007, pp. 932–943.
Neural Computation, Vol. 29, No. 9, 2017, pp. 2352–2449. 32. Karpatne, A., I. Ebert-Uphoff, S. Ravela, H. A. Babaie, and
18. Schmidhuber, J. Deep Learning in Neural Networks: An V. Kumar. Machine Learning for the Geosciences: Chal-
Overview. Neural Networks, Vol. 61, 2015, pp. 85–117. lenges and Opportunities. IEEE Transactions on Knowledge
19. Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. and Data Engineering, Vol. 31, No. 8, 2018, pp. 1544–1554.
Jones, A. N. Gomez, q. Kaiser, and I. Polosukhin. Atten- 33. Li, H., B. Wan, D. Chu, R. Wang, G. Ma, J. Fu, and Z.
tion is All You Need. Advances in Neural Information Pro- Xiao. Progressive Geological Modeling and Uncertainty
cessing Systems, Vol. 30, 2017. Analysis Using Machine Learning. ISPRS International
20. Reed, S., Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and Journal of Geo-Information, Vol. 12, No. 3, 2023, p. 97.
H. Lee. Generative Adversarial Text to Image Synthesis. 34. Mariethoz, G., and J. Caers. Multiple-Point Geostatistics:
Proc., 33rd International Conference on Machine Learning, Stochastic Modeling with Training Images. John Wiley &
New York, NY, 2016. Sons, UK, 2014.
21. Avci, O., O. Abdeljaber, S. Kiranyaz, M. Hussein, M. 35. Bai, T., and P. Tahmasebi. Hybrid Geological Modeling:
Gabbouj, and D. J. Inman. A Review of Vibration-Based Combining Machine Learning and Multiple-Point Statis-
Damage Detection in Civil Structures: From Traditional tics. Computers and Geosciences, Vol. 142, 2020, p. 104519.
Methods to Machine Learning and Deep Learning Appli- 36. Shi, C., and Y. Wang. Development of Training Image
cations. Mechanical Systems and Signal Processing, Vol. Database for Subsurface Stratigraphy. Georisk: Assessment
147, 2021, p. 107077. and Management of Risk for Engineered Systems and Geo-
22. Azimi, M., A. D. Eslamlou, and G. Pekcan. Data-Driven hazards, Vol. 17, No. 1, 2023, pp. 23–40.
Structural Health Monitoring and Damage Detection 37. Smirnoff, A., E. Boisvert, and S. J. Paradis. Support Vector
Through Deep Learning: State-of-the-Art Review. Sensors, Machine for 3D Modelling from Sparse Geological Infor-
Vol. 20, No. 10, 2020, p. 2778. mation of Various Origins. Computers and Geosciences,
23. Bao, Y., Z. Chen, S. Wei, Y. Xu, Z. Tang, and H. Li. The Vol. 34, No. 2, 2008, pp. 127–143.
State of the Art of Data Science and Engineering in Struc- 38. Wang, G., T. R. Carr, Y. Ju, and C. Li. Identifying
tural Health Monitoring. Engineering, Vol. 5, No. 2, 2019, Organic-Rich Marcellus Shale Lithofacies by Support Vec-
pp. 234–242. tor Machine Classifier in the Appalachian Basin. Comput-
ers and Geosciences, Vol. 64, 2014, pp. 52–60.
Yousefpour et al 21

39. Adeli, A., X. Emery, and P. Dowd. Geological Modelling 54. Molenaar, M. A Topology for 3-D Vector Maps. Interna-
and Validation of Geological Interpretations Via Simula- tional Journal of Applied Earth Observation and Geoinfor-
tion and Classification of Quantitative Covariates. Miner- mation, 1992, pp. 25–34.
als, Vol. 8, No. 1, 2017, p. 7. 55. Caumon, G., and J. L. Mallet. 3D Stratigraphic Models:
40. Xiang, J., K. Xiao, E. J. M. Carranza, J. Chen, and S. Li. Representation and Stochastic Modelling. Proc., Interna-
3D Mineral Prospectivity Mapping with Random Forests: tional Association for Mathematical Geology–XIth Interna-
A Case Study of Tongling, Anhui, China. Natural tional Congres, Liège, Belgium, 2006.
Resources Research, Vol. 29, 2020, pp. 395–414. 56. Houlding, S. 3D Geoscience Modeling: Computer Tech-
41. Gonc xalves, Í. G., S. Kumaira, and F. Guadagnin. A niques for Geological Characterization. Springer Science &
Machine Learning Approach to the Potential-Field Business Media, Berline, 2012.
Method for Implicit Modeling of Geological Structures. 57. Randle, C. H., C. E. Bond, R. M. Lark, and A. A. Mona-
Computers and Geosciences, Vol. 103, 2017, pp. 173–182. ghan. Uncertainty in Geological Interpretations: Effective-
42. Goncxalves, Í. G., F. Guadagnin, S. Kumaira, and S. L. Da ness of Expert Elicitations. Geosphere, Vol. 15, No. 1, 2019,
Silva. A Machine Learning Model for Structural Trend pp. 108–118.
Fields. Computers and Geosciences, Vol. 149, 2021, p. 104715. 58. Porwal, A., E. J. M. Carranza, and M. Hale. Artificial
43. Jia, R., Y. Lv, G. Wang, E. Carranza, Y. Chen, C. Wei, Neural Networks for Mineral-potential Mapping: A Case
and Z. Zhang. A Stacking Methodology of Machine Study from Aravalli Province, Western India. Natural
Learning for 3D Geological Modeling with Geological- Resources Research, Vol. 12, 2003, pp. 155–171.
Geophysical Datasets, Laochang Sn camp, Gejiu (China). 59. Zhang, Y., G. Su, and L. Yan. Gaussian Process Machine
Computers and Geosciences, Vol. 151, 2021, p. 104754. Learning Model for Forecasting of Karstic Collapse. Proc.,
44. Hillier, M., F. Wellmann, B. Brodaric, E. de Kemp, and E. Applied Informatics and Communication: International Con-
Schetselaar. Three-Dimensional Structural Geological ference, Xi’an, China, 2011.
Modeling Using Graph Neural Networks. Mathematical 60. Rodriguez-Galiano, V., M. Sanchez-Castillo, M. Chica-Olmo,
Geosciences, Vol. 53, No. 8, 2021, pp. 1725–1749. , and M. Chica-Rivas. Machine Learning Predictive Models
45. Zhou, C., J. Ouyang, W. Ming, G. Zhang, Z. Du, and Z. for Mineral Prospectivity: An Evaluation of Neural Net-
Liu. A Stratigraphic Prediction Method Based on Machine works, Random Forest, Regression Trees and Support Vector
Learning. Applied Sciences, Vol. 9, No. 17, 2019, p. 3553. Machines. Ore Geology Reviews, Vol. 71, 2015, pp. 804–818.
46. Shi, C., and Y. Wang. Nonparametric and Data-Driven 61. Gaurav, A. Horizontal Shale Well EUR Determination
Interpolation of Subsurface Soil Stratigraphy from Limited Integrating Geology, Machine Learning, Pattern Recogni-
Data Using Multiple Point Statistics. Canadian Geotechni- tion and Multivariate Statistics Focused on the Permian
cal Journal, Vol. 58, No. 2, 2021, pp. 261–280. Basin. Proc., SPE Liquids-Rich Basins Conference-North
47. Shi, C., and Y. Wang. Development of Subsurface Geolo- America, Odessa, Ukraine, 2017.
gical Cross-Section from Limited Site-Specific Boreholes 62. Sha, A. M., Z. Tong, and J. Gao. Recognition and Mea-
and Prior Geological Knowledge Using Iterative Convolu- surement of Pavement Disasters based on Convolutional
tion XGBoost. Journal of Geotechnical and Geoenvironmen- Neural Networks. China Journal of Highway and Transport,
tal Engineering, Vol. 147, No. 9, 2021, p. 04021082. Vol. 31, No. 1, 2018, pp. 1–10.
48. Wang, Y., C. Shi, and X. Li. Machine Learning of Geologi- 63. Wang, Y., Y. Hu, and T. Zhao. CPT-Based Subsurface Soil
cal Details from Borehole Logs for Development of High- Classification and Zonation in a 2D Vertical Cross-Section
Resolution Subsurface Geological Cross-Section and Geo- Using Bayesian Compressive Sampling. 2019. https://doi.
technical Analysis. Georisk: Assessment and Management of org/10.1139/cgj-2019-0131.
Risk for Engineered Systems and Geohazards, Vol. 16, No. 64. Maiti, S., and R. K. Tiwari. Neural Network Modeling
1, 2021, pp. 2–20. and An Uncertainty Analysis in Bayesian Framework: A
49. Ortiz, J. M., and C. V. Deutsch. Indicator Simulation Case Study from the KTB Borehole Site. Journal of Geo-
Accounting for Multiple-Point Statistics. Mathematical physical Research: Solid Earth, Vol. 115, No. B10, 2010.
Geosciences, Vol. 36, No. 5, 2004, pp. 545–565. 65. Fuentes, I., J. Padarian, T. Iwanaga, and R. W. Vervoort.
50. Pyrcz, M. J., J. B. Boisvert, and C. V. Deutsch. A Library 3D Lithological Mapping of Borehole Descriptions Using
of Training Images for Fluvial and Deepwater Reservoirs Word Embeddings. Computers & Geosciences, Vol. 141,
and Associated Code. Computers and Geosciences, Vol. 34, 2020, p. 104516.
No. 5, 2008, pp. 542–560. 66. Li, D. Q., X. H. Qi, Z. J. Cao, X. S. Tang, K. K. Phoon,
51. Nobre, M. M., and J. F. Sykes. Application of Bayesian and C. B. Zhou. Evaluating Slope Stability Uncertainty
Kriging to Subsurface Characterization. Canadian Geotech- Using Coupled Markov Chain. Computers and Geotech-
nical Journal, Vol. 29, No. 4, 1992, pp. 589–598. nics, Vol. 73, 2016, pp. 72–82.
52. Scheidt, C., L. Li, and J. Caers. Quantifying Uncertainty in 67. Li, Z., X. Wang, H. Wang, and R.Y. Liang. Quantifying
Subsurface Systems. John Wiley & Sons, Hoboken, NJ, Stratigraphic Uncertainties by Stochastic Simulation Tech-
2018. niques Based on Markov Random Field. Engineering Geol-
53. Chen, H. H., and T. S. Huang. A Survey of Construction ogy, Vol. 201, 2016, pp. 106–122.
and Manipulation of Octrees. Computer Vision, Graphics, 68. Qi, X. H., D. Q. Li, K. K. Phoon, Z. J. Cao, and X. S.
and Image Processing, Vol. 43, No. 3, 1988, pp. 409–431. Tang. Simulation of Geologic Uncertainty Using Coupled
22 Transportation Research Record 00(0)

Markov Chain. Engineering Geology, Vol. 207, 2016, 83. Li, H., and S. Misra. Robust Machine-Learning Workflow
pp. 129–140. for Subsurface Geomechanical Characterization and Com-
69. Gong, W., H. Tang, H. Wang, X. Wang, and C. H. Juang. parison Against Popular Empirical Correlations. Expert
Probabilistic Analysis and Design of Stabilizing Piles in Systems with Applications, Vol. 177, 2021, p. 14942.
Slope Considering Stratigraphic Uncertainty. Engineering 84. Lundberg, S. M., and S. I. Lee. A Unified Approach to
Geology, Vol. 259, 2019, p. 105162. Interpreting Model Predictions. Advances in Neural Infor-
70. Zhao, C., W. Gong, T. Li, C. H. Juang, H. Tang, and H. mation Processing Systems, Vol. 30, 2017.
Wang. Probabilistic Characterization of Subsurface Strati- 85. Asghari, V., Y. F. Leung, and S. C. Hsu. Deep Neural Net-
graphic Configuration with Modified Random Field work Based Framework for Complex Correlations in Engi-
Approach. Engineering Geology, Vol. 288, 2021, p. 106138. neering Metrics. Advanced Engineering Informatics, Vol. 44,
71. Xiao, T., H. F. Zou, K. S. Yin, Y. Du, and L. M. Zhang. 2020, p. 101058.
Machine Learning-Enhanced Soil Classification by Inte- 86. Ma, G., Z. Chao, and K. He. Predictive Models for Perme-
grating Borehole and CPTU Data with Noise Filtering. ability of Cracked Rock Masses Based on Support Vector
Bulletin of Engineering Geology and the Environment, Vol. Machine Techniques. Geotechnical and Geological Engi-
80, 2021, pp. 9157–9171. neering, Vol. 39, 2021, pp. 1023–1031.
72. Wei, X., and H. Wang. Stochastic Stratigraphic Modeling 87. Zhang, P., Z. Y. Yin, Y. F. Jin, T. H. Chan, and F. P.
Using Bayesian Machine Learning. Engineering Geology, Gao. Intelligent Modelling of Clay Compressibility Using
Vol. 307, 2022, p. 106789. Hybrid Meta-heuristic and Machine Learning Algorithms.
73. Karlsrud, K., T. Lunne, D. A. Kort, and S. Strandvik. Geoscience Frontiers, Vol. 12, No. 1, 2021, pp. 441–452.
CPTU Correlations for Clays. Proc., 16th International 88. Chen, C., and X. Xue. A Novel Hybrid Intelligent Model
Conference on Soil Mechanics and Geotechnical Engineer- for the Prediction of Creep Coefficients based on Random
ing, Osaka, Japan, 2005. Forest and Support Vector Machine. Ocean Engineering,
74. Paniagua, P., M. D’Ignazio, J.-S. L’Heureux, T. Lunne, Vol. 266, 2022, 113191.
and K. Karlsrud. CPTU Correlations for Norwegian 89. Zhang, P., Z. Y. Yin, and Y. F. Jin. Bayesian Neural Net-
Clays: An Update. AIMS Geosciences, Vol. 5, No. 2, 2019, work-Based Uncertainty Modelling: Application to Soil
pp. 82–103. Compressibility and Undrained Shear Strength Prediction.
75. Zhang, P., Z. Y. Yin, and Y. F. Jin. State-of-the-Art Canadian Geotechnical Journal, Vol. 59, No. 4, 2022,
Review of Machine Learning Applications in Constitutive pp. 546–557.
Modeling of Soils. Archives of Computational Methods in 90. NGI. CPTU Prediction Model. Techncial Note 20190685-
Engineering, Vol. 28, 2021, pp. 3661–3686. 01-TN. NGI, Norway, 2020.
76. Pham, B. T., T. A. Hoang, D. M. Nguyen, and D. T. Bui. 91. Gal, Y. Uncertainty in Deep Learning, Doctoral Disserta-
Prediction of Shear Strength of Soft Soil Using Machine tion. University of Cambridge, UK, 2016.
Learning Methods. Catena, Vol. 166, 2018, pp. 181–191. 92. Gong, W., C. Zhao, C. H. Juang, Y. Zhang, H. Tang,
77. Nguyen, T. A., H. B. Ly, and B. T. Pham. Backpropaga- and Y. Lu. Coupled Characterization of Stratigraphic
tion Neural Network-Based Machine Learning Model for and Geo-Properties Uncertainties–A Conditional Ran-
Prediction of Soil Friction Angle. Mathematical Problems dom Field Approach. Engineering Geology, Vol. 294,
in Engineering, Vol. 2020, 2020, pp. 1–11. 2021, p. 106348.
78. Mittal, M., S. C. Satapathy, V. Pal, B. Agarwal, L. M. 93. Griffiths, D. V., and G. A. Fenton. Probabilistic Slope Sta-
Goyal, and P. Parwekar. Prediction of Coefficient of Con- bility Analysis by Finite Elements. Journal of Geotechnical
solidation in Soil Using Machine Learning Techniques. and Geoenvironmental Engineering, Vol. 130, No. 5, 2004,
Microprocessors and Microsystems, Vol. 82, 2021, p. pp. 507–518.
103830. 94. Sauvin, G., M. Vanneste, M. E. Vardy, R. T. Klinkvort,
79. Saedi, B., and S. D. Mohammadi. Prediction of Uniaxial and F. Carl Fredrik. Machine Learning and Quantitative
Compressive Strength and Elastic Modulus of Migmatites Ground Models for Improving Offshore Wind Site Char-
by Microstructural Characteristics Using Artificial Neural acterization. Proc., Offshore Technology Conference, OTC,
Networks. Rock Mechanics and Rock Engineering, Vol. 54, Houston, TX, 2019.
2021, pp. 5617–5637. 95. Forsberg, C. F., T. Lunne, M. Vanneste, L. James, T. I.
80. Li, Y., H. Rahardjo, A. Satyanaga, S. Rangarajan, and D. Tjelta, A. Barwise, and C. Duffy. Synthetic CPTs from
T. T. Lee. Soil Database Development with the Applica- Intelligent Ground Models Based on the Integration of
tion of Machine Learning Methods in Soil Properties Pre- Geology, Geotechnics and Geophysics as A Tool for Con-
diction. Engineering Geology, Vol. 306, 2022, p. 106769. ceptual Foundation Design and Soil Investigation Plan-
81. Lin, P., X. Chen, M. Jiang, X. Song, M. Xu, and S. Huang. ning. Proc., Offshore Site Investigation Geotechnics 8th
Mapping Shear Strength and Compressibility of Soft Soils International Conference Proceeding, London, 2017.
with Artificial Neural Networks. Engineering Geology, Vol. 96. Vardy, M. E., M. A. Clare, M. Vanneste, C. F. Forsberg,
300, 2022, p. 106585. and J. K. Dix. Seismic Inversion for Site Characterization:
82. Niyogi, A., T. A. Ansari, S. K. Sathapathy, K. Sarkar, and When, Where and Why should We Use It? Proc., Offshore
T. N. Singh. Machine Learning Algorithm for the Shear Technology Conference, Houston, TX, 2018.
Strength Prediction of Basalt-Driven Lateritic Soil. Earth 97. Shi, C., and Y. Wang. Assessment of Reclamation-induced
Science Informatics, Vol. 16, No. 1, 2023, pp. 899–917. Consolidation Settlement Considering Stratigraphic
Yousefpour et al 23

Uncertainty and Spatial Variability of Soil Properties. 111. Kim, T., and T. Adali. Approximation by Fully Complex
Canadian Geotechnical Journal, Vol. 59, No. 7, 2022, Multilayer Perceptrons. Neural Computation, Vol. 15, No.
pp. 1215–1230. 7, 2003, pp. 1641–1666.
98. Robertson, P. K., and K. L. Cabal. Guide to Cone Penetra- 112. Yousefpour, N., Z. Medina-Cetina, F. G. Hernandez-
tion Testing for Geotechnical Engineering. Gregg Drilling Martinez, and A. Al-Tabbaa. Stiffness and Strength of
and Testing, Signal Hill, CA, 2012. Stabilized Organic Soils—Part II: Parametric Analysis
99. Tahmasebi, P.Multiple Point Statistics: A Review. In and Modeling with Machine Learning. Geosciences, Vol.
Handbook of Mathematical Geosciences: Fifty Years of 11, No. 5, 2021, p. 218.
IAMG (B. Daya Sagar, Q. Cheng, F. Agterberg, eds.), 113. Medina-Cetina, Z., N. Yousefpour, and J.-L. Briaud.
Springer, Cham, 2018, pp. 613–643. Probabilistic Evaluation of Unknown Foundations for
100. Ahmad, H., C. Ningsheng, M. Rahman, M. M. Islam, H. Scour Susceptible Bridges. Journal of Bridge Engineering,
R. Pourghasemi, S. F. Hussain, J. M. Habumugisha, et Vol. 25, No. 10, 2020, p. 04020074.
al. Geohazards Susceptibility Assessment along the Upper 114. Yousefpour, N., Z. Medina-Cetina, and J. L. Briaud.
Indus Basin Using Four Machine Learning and Statistical Evaluation of Unknown Foundations of Bridges Sub-
Models. ISPRS International Journal of Geo-Information, jected to Scour Physically Driven Artificial Neural Net-
Vol. 10, No. 5, 2021, p. 315. work Approach. Transportation Research Record:
101. Cao, J., Z. Zhang, J. Du, L. Zhang, Y. Song, and G. Sun. Journal of the Transportation Research Board, 2014.
Multi-Geohazards Susceptibility Mapping Based on 2433: 27–38.
Machine Learning—A Case Study in Jiuzhaigou, China. 115. Pan, S. J., and Q. Yang. A Survey on Transfer Learning.
Natural Hazards, Vol. 102, No. 3, 2020, pp. 851–871. IEEE Transactions on Knowledge and Data Engineering,
102. Chou, J. S., and A. D. Pham. Hybrid Computational Vol. 22, No. 10, 2010, p. 1345.
Model for Predicting Bridge Scour Depth Near Piers and 116. Sagi, O., and L. Rokach. Ensemble Learning: A Survey.
Abutments. Automation in Construction, Vol. 48, 2014, Wiley Interdisciplinary Reviews: Data Mining and Knowl-
pp. 88–96. edge Discovery, Vol. 8, No. 4, 2018, p. e1249.
103. Najafzadeh, M., F. Saberi-Movahed, and S. Sarkamar- 117. Picard, R. R., and R. D. Cook. Cross-Validation of
yan. NF-GMDH-Based Self-Organized Systems to Pre- Regression Models. Journal of the American Statistical
dict Bridge Pier Scour Depth Under Debris Flow Effects. Association, Vol. 79, No. 387, 1984, pp. 575–583.
Marine Georesources and Geotechnology, Vol. 36, No. 5, 118. Hashem, T., and N. Yousefpour. Application of Long-
2018, pp. 589–602. Short Term Memory and Convolutional Neural Net-
104. Bateni, S. M., S. M. Borghei, and D. S. Jeng. Neural Net- works for Real-Time Bridge Scour Forecast. arXiv Pre-
work and Neuro-Fuzzy Assessments for Scour Depth print arXiv:2404.16549, 2024. https://doi.org/10.48550/
around Bridge Piers. Engineering Applications of Artificial arXiv.2404.16549.
Intelligence, Vol. 20, No. 3, 2007, pp. 401–414. 119. Yang, L., and A. Shami. On Hyperparameter Optimiza-
105. Kaya, A. Artificial Neural Network Study of Observed tion of Machine Learning Algorithms: Theory and Prac-
Pattern of Scour Depth around Bridge Piers. Computers tice. Neurocomputing, Vol. 415, 2020, pp. 295–316.
and Geotechnics, Vol. 37, No. 3, 2010, pp. 413–418. 120. Wardhana, K., and F. C. Hadipriono. Analysis of Recent
106. Lee, T. L., D. S. Jeng, G. H. Zhang, and J. H. Hong. Bridge Failures in the United States. Journal of Perfor-
Neural Network Modeling for Estimation of Scour Depth mance of Constructed Facilities, Vol. 17, No. 3, 2003,
around Bridge Piers. Journal of Hydrodynamics, Vol. 19, pp. 144–150.
No. 3, 2007, pp. 378–386. 121. Briaud, J. L., Z. Medina-Cetina, S. Hurlebaus, M. Ever-
107. Yousefpour, N., S. Downie, S. Walker, N. Perkins, and ett, S. Tucker, N. Yousefpour, and R. Arjwech. Unknown
H. Dikanski. Machine Learning Solutions for Bridge Foundation Determination for Scour. Texas Transporta-
Scour Forecast based on Monitoring Data. Transportation tion Institute, College Station, TX, 2012.
Research Record: Journal of the Transportation Research 122. Colorado Water Science Center. Real-time Monitoring
Board, 2021. 2675: 745–763. and Evaluation of Bridge Scour. https://www.usgs.gov/
108. Yousefpour, N., and O. Correa. Towards an AI-Based centers/colorado-water-science-center/science/real-time-
Early Warning System for Bridge Scour. Georisk, Vol. 17, monitoring-and-evaluation-bridge-scour#publications
No. 4, 2023, pp. 713–739. 123. Hunt, B. E. Monitoring Scour Critical Bridges - A Synth-
109. Sharafati, A., M. Haghbin, D. Motta, and Z. M. Yaseen. esis of Highway Practice. NCHRP Synthesis 396. Trans-
The Application of Soft Computing Models and Empirical portation Research Board, Washington, D.C., 2009.
Formulations for Hydraulic Structure Scouring Depth 124. Henneberg, M. F. Real-time Streambed Scour Monitoring
Simulation: A Comprehensive Review, Assessment and Pos- at Two Bridges over the Gunnison River in Western Col-
sible Future Research Direction. Archives of Computational orado, 2016–17: U.S. Geological Survey Scientific Investi-
Methods in Engineering, Vol. 28, No. 2, 2021, pp. 423–447. gation Report. 2018. https://doi.org/10.3133/sir20185123
110. Hornik, K. Approximation Capabilities of Multilayer 125. Sauer, V., and D. Turnipseed. Stage Measurement at
Feedforward Networks. Neural Networks, Vol. 4, No. 2, Gaging Stations. U.S. Geological Survey (USGS), Virgi-
1991, pp. 251–257. nia, 2010.
24 Transportation Research Record 00(0)

126. Sauer, V., and D. Turnipseed. Discharge Measurement at 131. Yousefpour, N., and F. Fazel Mojtahedi. Early Detection
Gaging Stations. U.S. Geological Survey 536 (USGS), of Internal Erosion in Earth Dams: Combining Seismic
Virginia, 2010. Monitoring and Convolutional AutoEncoders. Georisk:
127. Karimi, D., H. Dou, S. K. Warfield, and A. Gholipour. Assessment and Management of Risk for Engineered Sys-
Deep Learning with Noisy Labels: Exploring Techniques tems and Geohazards, 2023, pp. 1–21.
and Remedies in Medical Image Analysis. Medical Image 132. Hochreiter, S., and J. Schmidhuber. Long Short-Term
Analysis, Vol. 65, 2020, p. 101759. Memory. Neural Computation, Vol. 9, No. 8, 1997,
128. Moritz, S., and T. Bartz-Beielstein. imputeTS: Timeseries pp. 37–45.
Missing Value Imputation in R. R Journal, Vol. 9, No. 1, 133. Bengio, Y., P. Simard, and P. Frasconi. Learning Long-
2017, p. 207. Term Dependencies with Gradient Descent is Difficult.
129. Rodrigues, F., K. Henrickson, and F. C. Pereira. Multi- IEEE Transactions on Neural Networks, Vol. 5, No. 2,
Output Gaussian Processes for Crowdsourced Traffic 1994, pp. 157–166.
Data Imputation. IEEE Transactions on Intelligent Trans- 134. Yousefpour, N., and M. Pouragha. Prediction of the
portation Systems, Vol. 20, No. 2, 2019, pp. 594–603. Post-Failure Behavior of Rocks: Combining Artificial
130. van Dijk, D., T. Teräsvirta, and P. H. Franses. Smooth Tran- Intelligence and Acoustic Emission Sensing. International
sition Autoregressive Models — A Survey of Recent Develop- Journal for Numerical and Analytical Methods in Geome-
ments. Econometric Reviews, Vol. 21, No. 1, 2002, pp. 1–47. chanics, Vol. 46, No. 10, 2022, pp. 1874–1894.