Computers & Geosciences 139 (2020) 104475

Contents lists available at ScienceDirect

Computers and Geosciences

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

Evaluation of machine learning methods for lithology classification using

geophysical data
Thiago Santi Bressan 1, *, Marcelo Kehl de Souza, Tiago J. Girelli, Farid Chemale Junior
Universidade do Vale do Rio do Sinos – Unisinos, S~
ao Leopoldo, RS, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: Specific computational tools assist geologists in identifying and sorting lithologies in well surveys and reducing
Lithological group operational costs and practical working time. This allows for the management of professional output, the efficient
Pattern recognition interpretation of data, and completion of scientific research on data collected in geologically distinct regions.
Multivariate data
Machine learning methods and applications integrate large sets of information with the goal of efficient pattern
Sedimentary rocks
recognition and the capability of leveraging accurate decision making. The objective of this study is to apply
machine learning methods to the supervised classification of lithologies using multivariate log parameter data
from offshore wells from the International Ocean Discovery Program (IODP). According to the analysis of the
lithologies proposed in the IODP Expeditions and for the application of our methods, the lithologies were divided
into four groups. The IODP Expeditions were organized into four templates for better results in analyzing the set
of expeditions and practical application of the methods. The templates were submitted to training, validation,
and testing by multilayer perceptron (MLP), decision tree, random forest, and support vector machine (SVM)
methods. The evaluation was randomly divided into training (70%), validation (10%), and testing (20%) using
the classification methods as an evaluation of the results. In the results, it was observed that Template1 (IODP
Expedition 362) obtained better results with the MLP method, Template2 (IODP Expeditions 354, 355, and 359)
and Template3 (IODP Expeditions 354, 355, 359, and 362) obtained better results with the random forest
method with greater than 80.00% accuracy. For cross-validation, the random forest method performed well in all
scenarios. In the practical template, the G2 group obtained a better result with the MLP method with an average
accuracy above 85.00%. It is expected that machine learning methods can help improve the study of geology
with accurate and rapid answers related to interpreting collected data in different study regions.

1. Introduction the identification and recognition of patterns in sedimentary rocks. The

The accurate classification of information intensifies work, improves
scientific results, and automates the execution of activities with reduced
time and work hours. Appropriate methods and computational tools
assist geologists in identifying lithologies in onshore or offshore well
surveys by allowing for efficient data interpretation. The efficiency of
lithological classification can be improved using applications capable of
producing more objective decisions rather than more conventional
methods of interpretation (Hsieh et al., 2005; Jahdhami and Anboori,
2017).
Integration between machine learning and geology with numerous
scientific and technological studies, search forms and processes aids in
collection of physical samples of sedimentary rocks is essential, as the
professional needs material available for visual interpretations and
laboratory analysis to generate information with a high degree of val­
idity and accuracy. This process is slow and costly, requiring highly
trained and dedicated professionals. In addition, the geophysical prop­
erties of sedimentary rocks acquired directly in the field with specific
equipment or in the laboratory can be utilized, reducing time and costs
in the identification of lithologies. These methods include applications
in the estimation of permeability and porosity using well profiles (Rafik
and Kamel, 2017), determination of lithofacies with seismic sections
(Brckovic et al., 2017), and geological mapping of surfaces and minerals
(Latifovic, Pouliot and Campbell, 2018; Bruno et al., 2018).

* Corresponding author.
E-mail addresses: thiago.bressan@iffarroupilha.edu.br (T.S. Bressan), marcelo.k.souza@gmail.com (M. Kehl de Souza), tjgirelli@gmail.com (T.J. Girelli),
faridchemale@gmail.com (F.C. Junior).
1
Instituto Federal de Educaç~
ao Ci^encia e Tecnologia Farroupilha - IFFar

https://doi.org/10.1016/j.cageo.2020.104475
Received 2 October 2019; Received in revised form 6 March 2020; Accepted 18 March 2020
Available online 23 March 2020
0098-3004/© 2020 Elsevier Ltd. All rights reserved.
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

This study differs by its integration of geology and neural networks in Table 1
intelligent learning to assist professional geologists with practical work Division of lithology into groups. Each group contains its composition models,
in the laboratory or the field. This paper applies the methods of machine lithological composition, and lithological code.
learning in the supervised classification of lithologies using multivariate Groups Models Lithological Composition Lithology
data of log parameters of offshore wells from the International Ocean Code
Discovery Program (IODP). GP Litho1 Very fine sand/sandstone, Fine sand/ 10
sandstone, Medium sand/sandstone, Coarse
2. Materials sand/sandstone, Sand, Sand/Sandstone

Litho2 Alternating sand/sandstone and mud/ 11

2.1. Machine learning mudstone layers, Sandy clay/claystone, Sandy
silt/siltstone
Litho3 Alternating silt/siltstone and clay/claystone 12
Machine learning can be defined as the most current method of using layers, Clayey silt/siltstone
a computer to discover new knowledge. In the last decade, computers Litho4 Sand/sandstone-silt/siltstone-clay/claystone, 13
have been used for simple learning or information storage only. With the Clayey sand/sandstone, Silty sand/sandstone
advancement of computer science and the creation of new, ultrafast Litho5 Silt/Siltstone, Silty clay/claystone, Clay/ 14
claystone, Clay, Silt, Mudstone
hardware, one can enhance algorithms and exponentially increase pro­
Litho6 Calcareous silty clay/claystone, Calcareous 15
cessing by considering computational thinking from the analysis of large silt/siltstone, Calcareous ooze, Chalk,
quantities of data in a short time. Machine learning is used in areas such Marlstone, Rudstone, Floatstone, Grainstone,
as natural language processing, speech recognition, stock prediction, Packstone, Wackestone, Boundstone,
identification of diseases in healthcare, calculation of porosity in sedi­ Limestone, Calcareous claystone

mentary rocks, and identification of specific minerals in sedimentary G1 Litho1 Very fine sand/sandstone, Fine sand/ 10
rocks. Machine learning creates responses from the intensive processing sandstone, Medium sand/sandstone, Coarse
sand/sandstone, Sand, Sand/Sandstone
of information, predicting highly reliable outputs for decision making
Litho2 Silt/siltstone, Silty clay/claystone, Clay/ 14
(Raschka, 2015). The supervised learning method processes reference claystone, Clay, Silt, Alternating silt/siltstone
data input to create a model for the prediction of new data. For this and clay/claystone layers, Clayey silt/siltstone
training, the algorithm requires data in a standard format and type, as Litho3 Calcareous silty clay/claystone, Calcareous 15
well as data with reliable and accurate values, extracted from relevant silt/siltstone, Calcareous ooze, Chalk,
Marlstone, Rudstone, Floatstone, Grainstone,
sources with the ability to improve feedback. In this study, we used a Packstone, Wackestone, Boundstone,
dataset organized from the division of lithologies into GP (group GP), G1 Limestone, Calcareous claystone
(group 1), G2 (group 2), and G3 (group 3) according to Table 1. These
G2 Litho1 Very fine sand/sandstone, Fine sand/ 10
lithologies are linked to IODP Expeditions 349, 354, 355, 356, 359, 361, sandstone, Medium sand/sandstone, Coarse
and 362 and organized into four templates according to Table 4. They sand/sandstone, Sand, Sand/sandstone
were processed by the application of machine learning methods for su­ Litho2 Sand/sandstone-silt/siltstone-clay/Claystone, 13
pervised classification, including MLP, decision tree, random forest, and Clayey sand/sandstone, Silty sand/sandstone,
Alternating sand/sandstone and mud/
SVM. mudstone layers, Sandy clay/claystone, Sandy
silt/siltstone
2.1.1. Machine learning methods for the classification lithology Litho3 Silt/siltstone, Silty clay/claystone, Clay/ 14
In addition to the methods used in this work, MLP, decision tree, claystone, Clay, Silt, Alternating silt/siltstone
and clay/claystone layers, Clayey silt/siltstone
random forest, and SVM, for supervised data, lithological classification
Litho4 Calcareous silty clay/claystone, Calcareous 15
covers the use of several other important methods, such as naïve Bayes silt/siltstone, Calcareous ooze, Chalk,
(Rosid et al., 2019; Kong et al., 2014), probabilistic neural networks Marlstone, Rudstone, Floatstone, Grainstone,
(Al-Mudhafar, 2017a), logistic boosted regression (Al-Mudhafar, Packstone, Wackstone, Boundstone,
2017a), kernel support vector machine (Al-Mudhafar, 2015, 2017b), Limestone, Calcareous claystone

support vector regression (Awad and Khanna, 2015), methods for un­ G3 Litho1 Very fine sand/sandstone, Fine sand/ 10
supervised data, such as cluster analysis (Lee and Datta-Gupta, 1999; sandstone, Medium sand/sandstone, Coarse
sand/sandstone, Sand, Sand/sandstone
Pirrone et al., 2014; McCreery and Al-Mudhafar, 2017) and gaussian
Litho2 Alternating sand/sandstone and mud/ 11
mixture models (Wallet and Hardisty, 2019), and methods that integrate mudstone layers, Sandy clay/claystone, Sandy
functions of math and multivariate statistics with supervised and un­ silt/siltstone, Sand/sandstone-silt/siltstone-
supervised data such as principal components analysis (Lee and clay/claystone, Clayey sand/sandstone, Silty
Datta-Gupta, 1999), linear discriminant analysis (Lee and Datta-Gupta, sand/sandstone
Litho3 Alternating silt/siltstone and clay/claystone 12
1999; Al-Mudhafar, 2015c; Hong et al., 2004), multinomial logistic layers, Clayey silt/siltstone, Silty clay/
regression (Hong et al., 2004), singular value decomposition (Romp­ claystone
panem et al., 2017) and fuzzy logic (Devanand and AuthorAnonymous, Litho4 Silt/Siltstone, Silt 13
2015; Orozco-del-Castillo et al., 2011), with application-specific char­ Litho5 Clay/claystone, Clay 14
Litho6 Calcareous silty clay/claystone, Calcareous 15
acteristics and support for certain data types.
silt/siltstone, Calcareous ooze, Chalk,
Marlstone, Rudstone, Floatstone, Grainstone,
2.1.1.1. Decision tree. The decision tree is a practical, fast, and robust Packstone, Wackestone, Boundstone,
learning method for supervised inductive learning (Maimon and Limestone, Calcareous claystone
Rokach, 2010). It is a useful method in the process of previously un­
known information extraction from the analysis of a large volume of
through a structure of nodes and sheets. When applied to database re­
data. Examples of applications that use a decision tree as a learning al­
cords, it results in the classification of records and is a robust method for
gorithm include landslides (Alkhasawneh et al., 2014), classification
data with considerable noise, as well as nonstandard data (S� aez et al.,
and identification of natural minerals (Akkas et al., 2015) and image
2013). Configurations such as maximum tree depth, number of features
classification (Loussaief and Abdelkrim, 2018).
for the best split, maximum number of nodes, maximum number of
A decision tree is essentially a series of if-else statements aligned
sheets, and the function for division and choice of nodes can be defined

2
T.S. Bressan et al.
and improved with training.
In relation to the criteria for division and choice of nodes, two
methods are available: Gini (1) and entropy (2).
X
c
GiniðEÞ ​ ¼ 1 p2j
j¼1
X
c
HðEÞ ​ ¼ ​ pj log pj
j¼1
The use of each criterion is defined according to the quantity and
format of data available for training. For large quantities of data, the
entropy criterion may be altered using a logarithm, requiring more
computational processing and making it relatively slower in relation to
the Gini criterion (Jing et al., 2017). In relation to the data format type,
the Gini criterion works well with binary values.
2.1.1.2. Random forest. Random forest is an ensemble method for
pattern classification developed based on decision trees. The random
forest creates a set of multiple decision trees and calculates the average
of all processed trees. In this context, the random forest includes groups
of calculations including averaging calculations that aggregate several
estimators and return the mean of their predictions and reduction in
their variances, and boosting calculations that integrate several small,
specific estimators, which seek an appropriately large return with the
sum of the results (James et al., 2013; Raschka, 2015).
Random forest is integral to the configuration criteria for division
and choice of nodes (Gini and entropy), depth of decision trees, and
quantity of decision trees.
The averaging method includes a bootstrap aggregation procedure as
a statistical approach used to quantify the degree of uncertainty asso­
ciated with a certain learning estimator.
The random forest method is applied in several areas, such as genetic
association studies in ecology (Brieuc et al., 2018), specifically in ge­
ology and geoscience in group mapping of vegetation (Dubeau et al.,
2017), predicting the spatial distribution of seabed hardness (Li et al.,
2016), fault detection (Puggini et al., 2015), facies and fracture pre­
diction (Bhattacharya and Mishra, 2018) and data analytics to model
daily hydrocarbon production (Bhattacharya et al., 2019).
2.1.1.3. Support vector machine (SVM). The SVM is a supervised
learning method that constructs a border (hyperplane) on a represen­
tation of the data, thus improving its presentation, grouping, and sep­
aration between different instances of data classes (Vapnik, 1998).
Its application depends on the format and the relationship between
the data, with a possible application on linearly separable and linearly
nonseparable data. The method calculates the vector that best represents
the hyperplane. Support vectors (Fig. 1) are points closest to the hy­
perplane line, and the distance between these vectors represent the
margins. The method chooses the hyperplane that has the greatest
margin, being considered the more robust model and less tolerant to
errors.
For a model with linearly separable data (as seen in Fig. 1), the
representation performed by SVM is fast and accurate, resulting in an
excellent practical return for classification output. A model with linearly
nonseparable data requires a method for interpreting the data group,
which applies a kernel method to find patterns and relationships after
applying the representation of the hyperplane to the group data.
SVM is applied in several areas, such as geology and geoscience for
3D geological modeling (De Boissieu et al., 2017; Bhattacharya and Carr,
2019), regolith-geology mapping (De Boissieu et al., 2017), hyper­
spectral classification of images (Peng et al., 2015), automated litho­
logical classification using ASTER remote sensing data (YU et al., 2012)
and identification of organic-rich lithofacies (Wang et al., 2014).
2.1.1.4. Multilayer perceptron (MLP). An artificial neural network
Computers and Geosciences 139 (2020) 104475
1
2
Fig. 1. Representation of SVM in a linearly separable model. The groups are
converted into data vectors and separated by a hyperplane calculated by the
method that best represents the separation of the data. Margins represent the
distance between the data vectors. Figure modified from Vapnik (1998).
(ANN) can be defined as a linear model based on brain architecture,
developed in an attempt to transfer learning capacity to a computerized
system (Castro et al., 2017; Souza et al., 2019). The best-known archi­
tecture of an ANN is the MLP, which is made up of a series of layers,
neurons, and connections.
It is important to note that due to its definition, the MLP architecture
depends on the adjustment of the main points: architecture configura­
tion and training. The configuration of the architecture includes the
definition of the connections and the number of neurons necessary for
good learning and adequate resolution of the proposed problem.
Consequently, training MLP networks can be very time consuming for
large datasets. The determination of the ideal architecture configuration
is a constant research objective, in which the quadratic error, prediction,
precision, and its metrics are adjusted for the expected results. In these
points and using these metrics, the results are related to the type and
quality of the data used in the process of forming the network.
Because it is a traditional network in the context of the generation
and application of knowledge, areas such as mudstone classification
(Bhattacharya et al., 2016), reservoir characterization (Fajana et al.,
2019) and pattern recognition in lithology classification of complex
aquifer systems (Sahoo and Jha, 2017) utilize the practical organization
of this important method.
2.1.1.5. Calibration of methods used. Further description of the method
hyperparameter configurations, as well as the result of calibration and
choice of values, can be found in the accompanying material.
2.1.1.6. Validation and evaluation. A confusion matrix is the represen­
tation of real values and predicted values that allows for visualization of
the performance of a machine learning classifier method on the pro­
posed templates (Navin and Pankaja, 2016). The confusion matrix
configures itself as a table generated for the classification of a binary
dataset and is used to describe the performance of a classification or
method. The main diagonal indicates the accuracy of the evaluated re­
cords, combining the true results in an organized structure in the matrix.
Its organization is presented in Fig. 2.
Accuracy (Eq. (3)) is formed by the division between true positive
and true negative values and the total positive and negative values,
according to Eq. (3):
TP þ TN
Accuracy ​ ¼ ​ 3
PþN
It is important to highlight that this classification metric can result in
3
T.S. Bressan et al.
Fig. 2. Definition of the confusion matrix. The matrix is divided into real data
and predicted data (rows and columns), combining true positive (TP) data, false
positive (FP) data, false negative (FN) data and true negative (TN) data.
Figure modified from Navin and Pankaja (2016).
false method performance because this metric calculates the average
between the return of the classes of a dataset (Hossin and Sulaiman,
2015).
Precision (Eq. (4)) and recall (Eq. (5)) belong to the F1-score metric.
Precision is calculated by the division of true positive values by the sum
between true positive values and false positive values. The recall is
calculated by the division of the true positive values by the sum between
true positive and false negative values. Their equations are presented
below:
TP
Precision ​ ¼ ​
TP þ FP
TP
Recall ​ ¼ ​
TP þ FN
F1-score or F-measure (Eq. (6)) is the harmonic mean between pre­
cision and recall. Its use is beneficial in dataset processing with diver­
sified classes that are highly disproportionate. This equation is given
below:
2*Precision*recall
F1 ​ ¼ ​
Precision þ recall
The cross-validation method is used to evaluate the performance of
the data. This method randomly partitions the total untrained dataset
into k smaller groups of equal size (Haykin, 2009). Processing of the data
is repeated k times until all groups are trained and tested. Processing
returns are described through rating metrics such as accuracy, precision,
recall, and F1-score according to Fig. 3. In this way, the entire set of
available data is evaluated, returning precise classification of the data
and integrating the various characteristics of data formation and
grouping.
Fig. 3. Graphical representation of cross-validation with the selected data in
the datasets of the templates. In each round, the dataset is divided into n groups
of equal size, with n groups for training and n groups for validation. The cross-
validation result produces the measurements of accuracy, precision, recall, and
F1-score. Figure modified from Haykin (2009).
Computers and Geosciences 139 (2020) 104475
The receiver operating characteristic (ROC) is a graphical method for
the evaluation, presentation, and selection of prediction systems
(Tharwat, 2018). It uses the confusion matrix to construct the results and
represents two parameters of the probability of accuracy (true positive
rate (TPR) and false positive rate (FPR)). TPR (Eq. (7)) and FPR (Eq. (8))
are defined as follows:
TP
TPR ​ ¼ ​ 7
TP þ FN
FP
FPR ​ ¼ 8
FP þ TN
The ROC curve plots TPR vs. FPR at different classification thresh­
olds. Its multidimensional capability allows for better visualization of
the result variables throughout the spectrum of the graph. The
descending diagonal (0,1) represents the classification model that plays
equally in both classes. Points belonging to the upper left triangle of this
diagonal represent the best results, and points belonging to the lower
right triangle represent the worst results. Its origin is related to the
detection of signals and the evaluation of the transmission quality of a
noise signal (Egan, 1975). ROC graphics are used in medicine (Tilaki-­
Hajian, 2013), economics (Gajowniczek et al., 2014), weather fore­
casting (Zhao et al., 2011), and geology (Vakhshoori and Zare, 2018;
Chen and Wu, 2016; Airola et al., 2018).
2.2. Geological setting
4
This study is based on IODP Expeditions 349, 354, 355, 356, 359,
361, and 362. The holes were drilled in different regions of the Indian
5
Ocean. As a result, there is a large amount of information resulting in a
good data group for this study in machine learning. Further description
of the study areas can be found in the supplementary material.
All of the expeditions described in the present study contain similar
successions of sediments/sedimentary rocks from the Bengal-Nicobar
Fan. Therefore, it is possible to perform a grouping of lithologies and
to map a pattern between depth and lithology, as well as the distribution
6
of sedimentary rocks in all of the sites and holes surveyed for the seven
expeditions.
The grouping of lithologies seeks to organize the sets of lithologies
present in the IODP Expeditions by separating the records, creating a
wide combination of data and directions for the heterogeneous sedi­
mentary rocks identified and described in the visual core description
(VCD). Fan et al. (2019), Korolev et al. (2018), and Rahim et al. (2009)
described the division into groups as integral to the heterogeneous re­
ality of lithological analysis in the field as it is related to the multivariate
physical characteristics of the site or core sampled.
For each lithological group, models and lithological code were
assigned for the classification of the lithology by machine learning
methods. Table 1 presents the model of the division of lithologies into
four groups, denominated GP (group GP), G1 (group 1), G2 (group 2),
and G3 (group 3). The core images of different lithologic groups can be
found in the supplementary material.
During IODP Expeditions, the onboard data collected onboard pro­
vided the first steps for working with the sampled rocks. The collection
of these data occurred in core or collected samples. Among the mea­
surements of the log parameters of rocks, we can highlight the gamma-
ray attenuation bulk density (GRA), P-wave velocity logger system (PWL
or P-wave), magnetic susceptibility (MS), reflectance spectrophotom­
etry and colorimetry (RSC), and shock remanent magnetization (SRM).
GRA is a measure of the density expressed in g/cm3, with a high
degree of penetration, emitted spontaneously from an atomic nucleus
(137 Ce) during radioactive decay. It has a gamma-ray peak of 662 KeV
and is attenuated as it passes through the core. This attenuation is
related to Compton spreading, where a known sample thickness is
proportional to the bulk density. Bulk density can also be affected by
vertical compaction during the collection of cores. This measure is used
4
T.S. Bressan et al.
in the identification or classification of rocks, mineral composition,
grain size, and porosity calculation (ODP, 2007).
PWL values are measurements of the sound wave velocity through a
sample. The PWL velocity varies according to the physical composition,
porosity, density, and degree of fracture. In marine environments, PWL
values are influenced by the degree of consolidation and lithification,
fractures, and hydrocarbon occurrence (Brckovic et al., 2017; Doveton,
1994). Together with the GRA, the PWL measurements are used to
calculate the acoustic impedance and reflection coefficients to construct
synthetic seismic profiles and to estimate the depths of seismic horizons.
MS is the intensity with which the material can be magnetized in an
external magnetic field (Blum, 1997; Brckovic et al., 2017; Mcneill et al.,
2017). The ratio of magnetization is expressed in units of volume,
defined as:
k ¼ M=H
where M is the volume of magnetization applied to a magnetic suscep­
tibility k by an applied external field (H). Susceptibility is measured with
the main recording devices, for which calibration factors must be met for
geometry and effects of transport and core coatings. They can be clas­
sified by the magnetization volume value (M) into three groups:
diamagnetic materials ( 1 < M < 0), paramagnetic materials (0 < M �
1), and ferromagnetic materials (M � 1). MS varies according to the type
and concentration of magnetic grains and corresponds to variation in
sediment composition, mainly the granulometry and mineralogical
composition. Sediments with the presence of clay have relatively lower
magnetic susceptibility, and materials with the presence of water tend to
have slightly negative values.
RSC is a unit of measurement related to two widely used techniques
in the visual identification of rock characteristics: colorimetry and
reflectance spectrophotometry. Colorimetry is used to measure the color
value of a surface. Many numerical systems have been developed to
express the visual values of colors. The International Commission on
Illumination (CIE) proposed a standard method for the numerical
measurement of colors, the L*, a*, b* system, considering the nonlinear
perception of the human eye and the combination of illumination and
basic colors (Hughes and Langlois, 2010; CIE, 2004). L* defines the
brightness value of the respective sample, with values between L* ¼
0 (totally dark) and L* ¼ 100 (totally bright). The a* defines the value of
the red-green coordinate (a* positive defines values with more red and
a* negative defines values with more green) and b* defines the value of
the blue-yellow coordinate (b* positive defines values with more yellow
and b* negative defines values with more blue) (Blum, 1997; CIE, 2004).
Colorimetry is widely used to calculate relative brightness and colors on
a surface or sample and is combined with other properties to create a
more accurate evaluation of the analyzed object (Hughes and Langlois,
2010).
Magnetization is correlated with the proportional response of mag­
netic susceptibility to a magnetic field. In some cases, for pure magne­
tization, this relationship may undergo changes, where a medium
exhibits a magnetic field, even with the absence of a magnetic field
applied to it. This process is called shock remanent magnetization (SRM)
(Buschow and Boer, 2004; Goodrich, 2007; Jovane et al., 2013). The
vector of magnetization of an object is the sum of the values of the
induced magnetic field and the magnetic field remanence. Magnetic
remanence can be classified into five types, of which the main type,
natural remanent magnetization (NRM), was used in the IODP. NRM is a
more reliable method since remanence is transmitted from an object
naturally by its favorable chemical compositions and without the
interference of equipment or sensors.
The definition of use for these seven log parameters is based on the
fact that the data acquisition resolution is closest to the depth, and the
joint capability of these logs facilitates the recognition and classification
of lithologies. Through analysis of the GRA, PWL, MS, L*, a*, b*, and
SRM in joint processing using machine learning, it is possible to identify
5
Computers and Geosciences 139 (2020) 104475
which member of the lithological group belongs to the value sought.
Fig. 4 shows a model of the log parameters for IODP Expedition 362,
sites U1480, and holes E, F, and G, which form Template1, lithology
group GP, with a plot of the classification of the lithologies in relation to
the depth of the wells. The other expeditions and sites follow the model
in Fig. 4, and adjust depth with the lithological classification.
2.3. Methods and data configuration
2.3.1. Data preparation
The total number of records for the IODP Expeditions, divided into
groups, is shown in Table 2. Each record includes values for the seven
log parameters GRA, PWL, MS, RSC (L*, a*, b*), and SRM. The dataset
formed by the log parameters is transformed into the matrix. A table of
the complete dataset with IODP Expeditions, groups, and lithologies is
found in Table S1 of the supplementary material.
2.3.2. Programming language and library for machine learning
Python is a high-level programming language that is interpretable,
easy to learn and use, and supports numerous add-ons, which makes it a
powerful language for calculating and analyzing large quantities of in­
formation (Spronck, 2017). In this machine learning study, the
scikit-learn library was used, which integrates all methods for learning
processing, with the support of supervised and unsupervised training on
data, creating extremely elaborate and understandable outputs. The
modules used in scikit-learn are the MLPClassifier module for the MLP
algorithm, the DecisionTreeClassifier module for the decision tree al­
gorithm, the RandomForestClassifier module for the random forest al­
gorithm, and the SVC module for the SVM algorithm.
2.3.3. Training
To carry out the training, the data were divided according to Table 4.
Three templates were created with data combinations between IODP
Expeditions 354, 355, 359, and 362, using 70.00% for training data,
10.00% for validation data and 20.00% for test data, randomly sepa­
rated. A practical template was created to simulate a real exercise that
applies the methods and predicts the lithologies of a hole through the
training of the entire dataset. The composition of the expeditions in the
templates was given by the proximity of the expeditions in the same
geological region to direct the context as practical analysis in the field or
directly on the ship.
2.3.4. Prediction
For each of the templates, the training described in the Train column
in Table 4 was performed using the MLP, decision tree, random forest,
and SVM methods. To carry out the prediction, data were used according
to the test column in Table 4.
The configurations of the MLP, decision tree, random forest, and
SVM methods are shown in Table 3. The configurations use features of
the methods to create algorithms in the Python language, and the main
points were adjusted for use with possible characteristics such as the
number of nodes, trees, depth, and activation function. The SVM method
was configured with the default configuration provided by the pro­
gramming language. The default setting maintenance is because it pre­
sents the best result compared to adjustments in the method parameters.
The distribution between training, validation, and testing in tem­
plates 1, 2, and 3 was related to the performance of practical tests with
the methods and in accordance with Storkey (2013) and Korjus et al.
(2016) with a percentage of 10% for validation and 10–20% for testing.
For the practical template, training, validation, and testing were divided
according to the segmentation of the expeditions.
After training and predictions, a table was compiled comparing the
output result for each template and each method of the analyzed dataset.
For each processing, classification metrics and confusion matrix values
were tabulated for better organization and presentation. A detailed table
with the results is found in Table S1 of the supplementary material.
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

Fig. 4. Log parameters for the IODP Expedition 362, Site U1480, Holes E, F, and G, which form Template1, lithology group GP. Lithology group GP is divided into
Litho1, Litho2, Litho3, Litho4, Litho5, and Litho6. Litho1 consists of sand with variable granulation; Litho2 is composed of interlayered sand and mud with layers of
clay and silt; Litho3 is composed of layers of clay and silt; Litho4 consists of mixed layers of sand, silt, and clay; Litho5 is composed of lithologies divided between silt,
clay, and mudstone; and finally, Litho6 is composed of lithologies based on carbonate composition and its derivatives. The color legend in the figure refers to the
lithologies present in the GP group. The unit of measurement for GRA is g/cm3, P-wave is m/s, MS is m3/kg, L* is %, a* is %, and b* is %. The SRM log parameter is
not included.

6
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

Table 2 G3 group, Template1 obtained 79.00% accuracy in the SVM method,

Total number of records by lithological groups. Template2 obtained 77.89% accuracy in the random forest method, and
Number of Records – IODP-Expeditions Group Template3 obtained 79.42% accuracy in the random forest method. A
detailed table with these results is found in Table S1 of the supple­
24,504 GP
23,967 G1 mentary material.
24,504 G2 The results of the application of cross-validation are shown in Fig. 5.
24,504 G3 They were divided into five folds, performing training and testing for all
Total number: 97,479 GP, G1, G2 and G3 templates and datasets. Accuracy, precision, recall, and F1-score were
used to measure the performance of templates for five-fold cross-vali­
dation. The decision tree method in the entropy criterion obtained a
Table 3 better result in relation to the Gini criterion, with an increase in accuracy
Configuration of the methods for training, validation and of 5.00%, as well as a proportional increase in cross-validation with 5 k-
testing. folds.
Methods Configuration In Fig. 6, log parameters are presented for the practical template in
MLP solver ¼ lbfgs the defined groups and methods (data of G3 group, IODP Expedition
activation ¼ relu 362, Site U1481). For the practical template, the execution of the
random_state ¼ 8 methods obtained excellent results, which can be observed in Fig. 6 and
DecisionTree random_state ¼ 8 Table S1 of the supplementary material.
max_depth ¼ 20
criterion ¼ entropy
It should be noted that the G2 group obtained better mean results in
RandomForest max_depth ¼ 20 accuracy for the four methods. The random forest method obtained an
n_estimators ¼ 1000 excellent result for the practical template, with more than 79.00% ac­
random_state ¼ 8 curacy for the classification of lithologies in the GP and G1 groups.
n_jobs ¼ 1
Emphasis is placed on the MLP method with 85.08% accuracy in the G2
SVM Default setting
group.
IODP Expedition 362, Site U1481 obtained the best result for the
3. Results and discussions lithological classification in the G3 group, MLP method. IODP Expedi­
tion 361, Site U1478 obtained the best result in the GP group, SVM
In this section, the results of the application of the four methods are method and site U1477 obtained the best result in the G2 group, MLP
discussed, according to the templates and dataset presented. method. IODP Expedition 356, Site U1462 obtained the best result in the
The sum of the records of the dataset used by the templates in this G2 group, MLP method. IODP Expedition 349, Site U1431 obtained the
study was 97,474 records. According to Strauβ (2018) and Venkatara­ best result in the G2 group, MLP method, Site U1432 obtained the best
man (2017), to obtain results above 70.00% accuracy using machine results in the G2 group, MLP and decision tree methods, Site U1433
learning, it is necessary to have thousands or millions of records for obtained excellent results in the G2 group, MLP method. Therefore, it
training. Such results depend on the organization of a precise and reli­ was concluded that the best method for the practical template is the
able historical database that adjusts itself with training and testing. In MLP, with the best results in 85.70% of the total practical template.
this case, it was determined that for the GP group, Template1 obtained The confusion matrix highlighted in Fig. 7 (related to Fig. 6) and the
an accuracy of 76.00% in the SVM method, Template2 obtained an ac­ supplementary material: boxplot represents the method result of the
curacy of 82.61% in the random forest method, and Template3 obtained respective lithological classification for IODP Expedition 362, Site
an accuracy greater than 82.74% in the random forest method. For the U1481 (G3 group). The MLP method (A) has the best average result for
G1 group, Template1 obtained an accuracy of 89.51% in the random the hole, with an accuracy of 76.30%. For Litho1 (lithology code 10), the
forest method, Template2 obtained an accuracy of over 75.00% in the best accuracy was obtained by the random forest method (B). For Litho3
four methods analyzed, and Template3 obtained an accuracy of over (lithology code 12), the best accuracy was obtained by the MLP method
73.00% in the four methods analyzed. For the G2 group, Template1 (A). For Litho4 (lithology code 13), the best accuracy was obtained by
obtained 82.84% accuracy in the random forest method, Template2 the MLP (A) and SVM (D) methods. Finally, for Litho5 (lithology code
obtained 82.61% accuracy in the random forest method, and Template3 14), the best accuracy was obtained by the MLP (A) and decision tree (C)
obtained 84.20% accuracy in the random forest method. Finally, for the methods. The complete table with classification metrics is found in

Table 4
Division for training according to IODP Expeditions.
Groups Template Composition IODP Expeditions – Sites Train Validation Test Methods

GP, G1, G2 Template1 362 (U1480) 70.00% 10.00% 20.00% MLP

and G3 Template2 354 (U1449, U1450, U1451) 70.00% 10.00% 20.00% DecisionTree
355 (U1456) RandomForest
359 (U1465, U1466, U1467, U1468, SVM
U1470, U1471, U1472)
Template3 354 (U1449, U1450, U1451) 70.00% 10.00% 20.00%
355 (U1456)
359 (U1465, U1466, U1467, U1468,
U1470, U1471, U1472)
362 (U1480)
Practical 349 (U1431, U1432, U1433) Expeditions 354,355,359 and Expedition Expeditions
Template 354 (U1449, U1450, U1451) 362 (U1480) 349 349, 356, 361 and 362
355 (U1456) (U1433) (U1481)
356 (U1462)
359 (U1465, U1466, U1467, U1468,
U1470, U1471, U1472)
361 (U1477, U1478)
362 (U1480, U1481)

7
T.S. Bressan et al.
Fig. 5. Results of the application of cross-validation in the four defined groups:
GP, G1, G2, and G3, divided into Template1, Template2, and Template3.
Template1 covers IODP Expedition 362, Site U1480; Template2 covers IODP
Expeditions 354, 355, and 359; and Template3 covers IODP Expeditions 354,
355, 359, and 362. The methods applied to the data were MLP, decision tree,
random forest, and SVM.
Table S1 of the supplementary material.
Fig. 8 shows the ROC curve representing four methods for the G3
group, practical template in IODP Expedition 362, Site U1481, divided
by models (Litho1, Litho3, Litho4, Litho5) and lithological code
8
Computers and Geosciences 139 (2020) 104475
(10,12,13, and 14). The results are good in all plots, with a positive ROC
curve above the main diagonal (0,1) and corroborate with precise
identification (from 70.00% to 86.00%) of the lithologies, as seen in
Fig. 7. The MLP method has a better and growing stability of the plotted
results relative to the other methods.
The best result for the lithological classification linked to the G2
group is because the grouping was organized into four lithologies:
Litho1, Litho2, Litho3, and Litho4, which formed a larger organization,
agglomerating more lithologies and facilitating recognition by the
methods. The GP group did not obtain good results in the general
context due to greater refinement of the division and grouping of the
lithologies, making it difficult to recognize the analyzed methods.
Table 5 summarizes the results of machine learning processing
methods on the groups and data templates proposed in this article. The
best method was random forest. The best data group was the G2 group.
The best template was Template2. Referring to group vs. method, in the
GP group, the best method was random forest; in the G1 group, the best
method was random forest; in the G2 group, the best method was MLP;
and in the G3 group, the best method was random forest. In the com­
bination of template vs. method, in Template1, Template2, and Tem­
plate3, the best method was random forest, and in the practical
template, the best method was MLP. The best result in applying cross-
validation was the random forest method.
In all processes, there are limitations that create restrictions on
achieving results with more than 80.00% accuracy (Peng and Bai, 2017;
Papernot et al., 2016), especially due to the characteristics of the data­
sets, the quantity of data in the datasets, the number of features (GRA,
PWL, MS, L*, a*, b*), and the inability to store training and testing for
future use, which strengthens datasets and improves results.
Regarding the performance and complexity of the execution of the
algorithms, the algorithms that use decision trees, random trees, nodes
and edges are executed very quickly and accurately than algorithms that
use mathematical calculations or statistical functions (Maniriho and
Ahmad, 2018). The MLP algorithm has in its organizational structure a
sequential node-to-node and layer-by-layer execution runs quickly with
optimal performance in relation to the hardware involved. Finally, the
SVM algorithm, acting directly on variables, vectors, performs calcula­
tions in a two-dimensional or three-dimensional plane using an appro­
priate kernel (Gaussian, polynomials or sigmoid) requires compatible
hardware and high processing time compared to other algorithms. It is
observed that in the results proposed by this article, the SVM algorithm
always obtained inferior results to the algorithms that use decision trees
or nodes and layers.
The ash layer was not part of the division of the lithological groups
since its presence in the lithological profile of the wells is restricted to
mm- and cm-thick layers, with few log records and inadequate resolu­
tion, making it impossible for machine learning applications in all of the
configurations used at this stage. Further details referring to the ash
layer records are found in Table S2 of the supplementary material with
ash layer records for the IODP Expedition 362, Site U1480.
4. Conclusions
In this study, four different machine learning methods were applied
to three standard data templates and a practical data template in a
lithological classification problem for wells from IODP Expeditions.
The results indicate that the MLP method had better results in the
lithological classification for the practical template, considering the li­
thology of the G2 group proposed in this research and the characteristics
of the datasets used.
In datasets Template1, Template2, and Template3, the best results
were in the G1 group with the random forest method, with an accuracy
of more than 85.00%. The G1 group had better results for lithological
classification for the organized templates due to the advantageous
grouping of the three lithologies (Litho1, Litho2, and Litho3). The Litho2
grouping composite for joining lithologies between silt and clay
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

Fig. 6. Resulting lithological classification – practical template, G3 group, IODP Expedition 362, Site U1481. The log parameters of the well GRA, P-wave, MS, L*, a*,
and b* are displayed with respect to the real lithology related to the depth presented. Graph A displays the total number of records per method vs. lithology. Graph B
displays the resulting lithological classification by method relative to the real lithology. The methods are MLP, random forest, decision tree, and SVM, presented in
the sequence of identification of the lithologies. (Aa) to (Dd) correspond to the method distribution by percentage of hits by lithology. Litho1 refers to lithology code
10, Litho3 refers to lithology code 12, Litho4 refers to lithology code 13 and Litho5 refers to lithology code 14. Litho2 and Litho6 are not recorded in this hole range.
The SRM log parameter is not included.

increased the accuracy according to the quantity of the dataset, as it
facilitated the execution of the classifications by the methods proposed.
The SVM method obtained poor results in cross-validation analysis
and accuracy. The geological data analyzed do not have a highly defined
pattern, and this method depends on many external and natural factors
to obtain excellent results in its application. To successfully use machine
learning with this method, it is necessary to have a dataset with a large
quantity of data, and the data must be balanced regarding each training
for a specific group of data. Indeed, we must perform an adjustment of
the configuration parameters, making it difficult to use this for the
classification of lithologies as proposed in this study.
The best result for cross-validation was obtained by the random
forest method in all groups and templates analyzed. The characteristics
of the random forest method allowed for better results of lithological
classification, as it has a simple configuration and its design includes a
set of several decision trees, leading to accurate results with smaller

9
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

Fig. 7. Confusion matrix for the G3 group, practical template (IODP Expedition 362, Site U1481). (A) MLP method, (B) random forest method, (C) decision tree
method, (D) SVM method. A to D represent the diagonal matrix for the result of each method, where the MLP method had the best result. (A*) represents the number
of true and predicted G3 group (lithology codes 10, 12, 13, 14) lithology records without normalization. (B*) highlights the percentage of normalized data between
true and predicted. Colors represent the frequency of normalized and nonnormalized records. Lithology code 11 and lithology code 15 are not recorded in this
hole range.

10
T.S. Bressan et al. Computers and Geosciences 139 (2020) 104475

Table 5
Summary table with the best accuracy per method, group, and
template.
Table summary best accuracy

Method:
RandomForest
Group:
G2
Template:
Template2
Group vs. Method:
GP RandomForest
G1 RandomForest
G2 MLP
G3 RandomForest
Template vs. Method:
Template1 RandomForest
Template2 RandomForest
Template3 RandomForest
Practical Template MLP
Cross-Validation:
RandomForest

statistical variance.
The main contributions of this work are a rapid approach to litho­
logical classification in offshore wells, the proposal of a lithological
classification using supervised training methods, lithological classifica­
tion using multivariate data and the support of a large number of vari­
ables, and strengthened and improved learning using neural network
methods and machine learning.

Computer code availability

Name of code: LithoPy. Developer: Thiago Santi Bressan. Contact

address: Programa de Po �s-Graduaça~o em Geologia, Universidade do
Vale do Rio dos Sinos (Unisinos), Av. Unisinos, 950, Cristo Rei, S~ ao
Leopoldo, RS, Brasil. Telephone number: þ55-55-99601-6723. e-mail.
tsbressan@gmail.com. Year of first release: 2019. Hardware required: I3
CPU or better with 16 GB memory RAM. Software required: Anaconda
(>¼5.3), Jupyter Notebook (>¼5.2), Sklearn (>¼0.20), NumPy
(>¼1.8.2), SciPy (>¼ 0.13.3), Matplotlib (>¼3.0.0) and Windows (10)
or Linux (Ubuntu or such as system). Program language: Python. Pro­
gram size: 10 MB. How to access the source code: Available at https://
github.com/tsbressan/LithoPy.

Authorship statement

Thiago Santi Bressan designed and developed the algorithms,

worked on the main writing of the manuscript. Marcelo Kehl de Souza
contributed to the statistical analysis, algorithm tests, review and
writing of the manuscript. Tiago J Girelli contributed to the develop­
ment of figures and tables and writing of the manuscript. Farid Chemale
Junior contributed to the writing and revision of the manuscript.

Declaration of competing interest

The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to influence
the work reported in this paper.

Fig. 8. ROC curve of the MLP, random forest, decision tree, and SVM methods
applied to the G3 group, practical template (IODP Expedition 362, Site U1481).
Lithology code 10 refers to the G3 group, models Litho1. Lithology code 12
refers to the G3 group, models Litho3. Lithology code 13 refers to the G3 group,
models Litho4. Lithology code 14 refers to the G3 group, models Litho5. Li­
thology code 11 and lithology code 15 are not recorded in this hole range.
Acknowledgement
This work was financed through a grant from the Coordenaça
~o de
Aperfeiçoamento de Pessoal de Nível Superior – Brasil -grant #
88887.091717/2014–01). This research used data provided by the In­
ternational Ocean Discovery Program (IODP) (www.iodp.org/access-d
ata-and-samples).

11
T.S. Bressan et al.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.cageo.2020.104475.
Abbreviations used in this manuscript
ANN Artificial Neural Network
CIE International Commission on Illumination
GRA Gamma Ray Attenuation
IODP International Ocean Discovery Program
MS Magnetic Susceptibility
MLP Multi-Layer Perceptron
NRM Natural Remanence Magnetization
PWL P-Wave Velocity Logger System
RSC Reflectance Spectrophotometry
ROC Receiver Operating Characteristic
SRM Magnetic Remanence
SVM Support Vector Machine
VCD Visual Core Description
References
Airola, A., Pohjankukka, J., Torppa, J., Middleton, M., Nyk€ anen, V., Heikkonen, J.,
Pahikkala, T., 2018. The spatial leave-pair-out cross-validation method for reliable
AUC estimation of spatial classifiers. Data Min. Knowl. Discov. 33, 730–747. https://
doi.org/10.1007/s10618-018-00607-x.
Akkas, E., Akin, L., Çubukçu, E., Artuner, H., 2015. Application of Decision Tree
Algorithm for classification and identifi-cation of natural minerals using SEM–EDS.
Comput. Geosci. 80, 38–48. https://doi.org/10.1016/j.cageo.2015.03.015, 2015.
Al-Mudhafar, W.J., 2015. Integrating component analysis & classification techniques for
comparative prediction of continuous & discrete lithofacies distributions. In: OTC-
25806-MS, the Offshore Technology Conference. https://doi.org/10.4043/25806-
MS (4-7 May), Houston, TX, USA.
Al-Mudhafar, W.J., 2017a. Integrating well log interpretations for lithofacies
classification and permeability modeling through advanced machine learning
algorithms. J. Petrol. Explor. Prod. Technol. (2017) 7, 1023–1033. https://doi.org/
10.1007/s13202-017-0360-0.
Al-Mudhafar, W.J., 2017b. Integrating kernel support vector machines for efficient rock
facies classification in the main pay of Zubair formation in South Rumaila oil field,
Iraq. Model. Earth Syst. Environ. (2017) 3, 12. https://doi.org/10.1007/s40808-
017-0277-0.
Alkhasawneh, M.S., Ngah, U.K., Tay, L.T., Isa, N.A.M., Al-Batah, M.S., 2014. Modeling
and testing landslide hazard using decision tree. J. Appl. Math. 2014 https://doi.
org/10.1155/2014/929768, 929768.
Awad, M., Khanna, R., 2015. Support Vector Regression. Efficient Learning Machines.
Apress, Berkeley, CA. https://doi.org/10.1007/978-1-4302-5990-9_4.
Bhattacharya, S., Carr, T.R., 2019. Integrated data-driven 3D shale lithofacies modeling
of the Bakken Formation in the Williston basin, North Dakota, United States.
J. Petrol. Sci. Eng. 177, 1072–1086. https://doi.org/10.1016/j.petrol.2019.02.036,
2019.
Bhattacharya, S., Mishra, S., 2018. Applications of machine learning for facies and
fracture prediction usingBayesian Network Theory and Random Forest: case studies
from theAppalachian basin, USA. J. Petrol. Sci. Eng. 170, 1005–1017. https://doi.
org/10.1016/j.petrol.2018.06.075, 2018.
Bhattacharya, S., Carr, T.R., Pal, M., 2016. Comparison of supervised and unsupervised
approaches for mudstonelithofacies classification: case studies from the Bakken
andMahantango-Marcellus Shale, USA. J. Nat. Gas Sci. Eng. 33, 1119–1133. https://
doi.org/10.1016/j.jngse.2016.04.055, 2016.
Bhattacharya, S., Ghahfarokhi, P.K., Carr, T.R., Pantaleone, S., 2019. Application of
predictive data analytics to model daily hydrocarbon production using
petrophysical, geomechanical, fiber-optic, completions, and surface data: a case
study from the Marcellus Shale, North America. J. Petrol. Sci. Eng. 176, 702–715.
https://doi.org/10.1016/j.petrol.2019.01.013, 2017.
Blum, P., 1997. Physical properties handbook: a guide to the shipboard measurement of
physical properties of deep-sea cores. Coll. Stat. Texas, USA http://www-odp.tamu.
edu.
Brckovic, A., Kovacevic, M., Cvetkovic, M., Mocilac, I.K., Rukavina, D., Saftic, B., 2017.
Application of artificial neural networks for lithofacies determination based on
limited well data. Central Eur. Geol. 3 (60), 299–315. https://doi.org/10.1556/
24.60.2017.012.
Brieuc, M.S.O., Waters, C.D., Drinan, D.P., Naish, K.A., 2018. A practical introduction to
Random Forest for genetic association studies in ecology and evolution. Mol. Ecol.
Resour. 18 (4), 755–766. https://doi.org/10.1111/1755-0998.12773.
Bruno, A.E., Charbonneau, P., Newman, J., Snell, E.H., So, D.R., Vanhoucke, V.,
Watkins, C.J., Williams, S., Wilson, J., 2018. Classification of crystallization
outcomes using deep convolutional neural networks. PloS One 13, 6. https://doi.
org/10.1371/journal.pone.0198883.
12
Computers and Geosciences 139 (2020) 104475
Buschow, K.H.J., Boer, F. R. de, 2004. Physics of Magnetism and Magnetic Materials.
Kluwer Academic Publishers, ISBN 0-306-47421-2, pp. 105–115. https://doi.org/
10.1007/b100503.
Castro, W., Oblitas, J., Santa-Cruz, R., Avila-George, H., 2017. Multilayer perceptron
architecture optimization using parallel computing techniques. PloS One 12, 12.
https://doi.org/10.1371/journal.pone.0189369.
Chen, Y., Wu, W., 2016. A prospecting cost-benefit strategy for mineral potential
mapping based on ROC curve analysis. Ore Geol. Rev. 74, 26–38. https://doi.org/
10.1016/j.oregeorev.2015.11.011, 2016.
CIE, 2004. Technical Report, third ed. COLORIMETRY, ISBN 3901906339, p. 82pp
https://archive.org/details/gov.law.cie.15.2004.
De Boissieu, F., Sevin, B., Cudahy, T., Mangeas, M., Chevrel, S., Ong, C., Rodger, A.,
Maurizot, P., Laukamp, C., Lau, I., Touraivane, T., Cluzel, D., Despinoy, M., 2017.
Regolith-geology mapping with support vector machine: a case study over
weathered Ni-bearing peridotites, New Caledonia. Int. J. Appl. Earth Obs. Geoinf.
(64), 377–385. https://doi.org/10.1016/j.jag.2017.05.012.
Devanand, Kumar, N., 2015. Prediction of CMRS rock mass rating using fuzzy logic.
International Conference on Advances in Computer Engineering and Applications.
https://doi.org/10.1109/ICACEA.2015.7164685. ICACEA - 2015.
Doveton, J.H., 1994. Geological log interpretation. SEPM Society for Sedimentary
Geology (29). https://doi.org/10.2110/scn.94.29.
Dubeau, P., King, D.J., Unbushe, D.G., Rebelo, L., 2017. Mapping the dabus wetlands,
Ethiopia, using random forest classification of landsat, PALSAR and topographic
data. Rem. Sens. (9), 1056. https://doi.org/10.3390/rs9101056.
Egan, J.P., 1975. Signal Detection Theory and ROC Analysis. Academic Press, New York,
2017.
Fajana, A.O., Ayuk, M.A., Enikanselu, P.A., 2019. Application of multilayer perceptron
neural network and seismic multiattribute transforms in reservoir characterization of
Pennay field, Niger Delta. J. Petrol. Explor. Prod. Technol. (9), 31–49. https://doi.
org/10.1007/s13202-018-0485-9, 2019.
Fan, B., Shi, L., Li, Y., Zhang, T., Lv, L., Shikai, T., 2019. Lithologic heterogeneity of
lacustrine shale and its geological significance for shale hydrocarbon-a case study of
Zhangjiatan Shale. Open Geosci. 2019 (11), 101–112. https://doi.org/10.1515/geo-
2019-0009.
Gajowniczek, K., Zabkowski, T., Szupiluk, R., 2014. Estimating the roc curve and its
significance for classification models’ assessment. Quantit. Methods Econ. 15 (2),
382–391, 2014.
Goodrich, W.E., 2007. Characterization and Quantification of Magnetic Remanence in
Unexploded Ordnance. Master of Science (Geophysics). Faculty and Board of
Trustees of the Colorado School. https://mountainscholar.org/bitstream/handle
/11124/79225/T06342.pdf?sequence¼1.
Haykin, S., 2009. Neural Networks and Learning Machines, third ed. Pearson Prentice
Hall, New York City, U. S., p. 938pp
Hong, T., White, C.D., Gani, M.R., Bhattacharya, J., 2004. Comparison of multivariate
statistical algorithms for wireline log facies classification. AAPG Ann. Meet Abst.
(88), 13.
Hossin, M., Sulaiman, M.N., 2015. A review on evaluation metrics for data classification
evaluations. Int. J. Data Min. Knowl. Manag. Process 2 (5), 1–11. https://doi.org/
10.5121/ijdkp.2015.5201.
Hsieh, B., Lewis, C., Lin, Z., 2005. Lithology identification of aquifers from geophysical
well logs and fuzzy logic analysis: shui-Lin Area, Taiwan. Comput. Geosci. 31 (3),
263–275. https://doi.org/10.1016/j.cageo.2004.07.004.
Hughes, V.K., Langlois, N.E.I., 2010. Use of reflectance spectrophotometry and
colorimetry in a general linear model for the determination of the age of bruises.
Forensic Sci. Med. Pathol. 6 https://doi.org/10.1007/s12024-010-9171-z, 275–28.
Jahdhami, N.A., Anboori, A., 2017. The Application of Specific Drilling Energy to
Identify Overburden Lithological Boundaries and Aid Well Operations - Oman
Khazzan Field. Abu Dhabi International Petroleum Exhibition & Conference, Abu
Dhabi, UAE. https://doi.org/10.2118/188413-MS.
James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An Introduction to Statistical
Learning. Springer, New York City, U.S, p. 187. https://doi.org/10.1007/978-1-
4614-7138-7.
Jing, S., Liu, C., Li, G., Yan, G., Zhang, Y., 2017. An efficient algorithm for parallel
computation of rough entropy using CUDA. Int. Conf. Comput. Intell. Sec. 13
https://doi.org/10.1109/CIS.2017.00009.
Jovane, L., Hinnov, L., Housen, B.A., Herrero-Barvera, E., 2013. Magnetic Methods and
the Timing of Geological Processes, vol. 373. Geological Society Special Publication
Nº https://doi.org/10.1144/SP373.17.
Kong, G., Xia, Y., Qiu, C., 2014. Cost-sensitive bayesian network classifiers and their
applications in rock burst prediction. ICIC 2014, LNCS 8588, 101–112. https://doi.
org/10.1007/978-3-319-09333-8_12.
Korjus, K., Hebart, M.N., Vicente, R., 2016. An efficient data partitioning to improve
classification performance while keeping parameters interpretable. PloS One 11 (8).
https://doi.org/10.1371/journal.pone.0161788.
Korolev, E.A., Usmanov, S.A., Nikolaev, D.S., Gabdelvaliyeva, R.R., 2018. Effect of
lithological heterogeneity of bitumen sandstones on SAGD reservoir development.
IOP Conf. Ser. Earth Environ. Sci. 155 https://doi.org/10.1088/1755-1315/155/1/
012019, 2018.
Latifovic, R., Pouliot, D., Campbell, J., 2018. Assessment of convolution neural networks
for surficial geology mapping in the south rae geological region, northwest
territories, Canada. Rem. Sens. 10 (2), 307. https://doi.org/10.3390/rs10020307.
Lee, S.H., Datta-Gupta, A., 1999. Electrofacies characterization and permeability
predictions in carbonate reservoirs: role of multivariate analysis and nonparametric
regression. In: SPE Annual Technical Conference and Exhibition (3–6 October).
https://doi.org/10.2118/56658-MS. Houston, Texas.
T.S. Bressan et al.
Li, J., Tran, M., Siwabessy, J., 2016. Selecting optimal random forest predictive models: a
case study on predicting the spatial distribution of seabed hardness. Geosci. Australia
GPO Box 378. https://doi.org/10.1371/journal.pone.0149089.
Loussaief, S., Abdelkrim, A., 2018. Machine Learning framework for image classification.
ASTESJ 3, 1. https://doi.org/10.1109/SETIT.2016.7939841.
Maimon, O., Rokach, L., 2010. Data Mining and Knowledge Discovery Handbook, 2
editions. Springer, New York City, U.S. https://doi.org/10.1007/978-0-387-09823-4
(Chapter 9).
Maniriho, P., Ahmad, T., 2018. Analyzing the Performance of Machine Learning
Algorithms in Anomaly Network Intrusion Detection Systems. In: 4th International
Conference on Science and Technology. https://doi.org/10.1109/
ICSTC.2018.8528645. ICST 2018.
McCreery, E., Al-Mudhafar, W., 2017. Geostatistical classification of lithology using
partitioning algorithms on well log data - a case study in forest hill oil field, east
Texas basin. 79th eage conference and exhibition. At Paris, France. https://doi.org/
10.3997/2214-4609.201700905.
Mcneill, L.C., Dugan, B., Petronotis, K.E., Backman, J., Bourlange, S., Chemale, F.,
Chen, W., Colson, T.A., Frederik, M.C.G., Guerin, G., Hamahashi, M., Henstick, T.,
House, B.M., Hupers, A., Jeppson, T.N., Kachovich, S., Kenigsberg, A.R.,
Kuranaga, M., Kutterolf, S., Milliken, K.L., Mitchison, F.L., Mukoyoshi, H., Nair, N.,
Owari, S., Pickering, K.T., Pouderoux, H.F.A., Yehua, S., Song, I., Torres, M.E.,
Vannucchi, P., Vrolijk, P.J., Yang, T., Zhao, X., 2017. Expedition 362 summary. Proc.
Int. Ocean Discov. Program 362. https://doi.org/10.14379/iodp.
proc.362.101.2017.
Navin, M., Pankaja, R., 2016. Performance analysis of text classification algorithms using
confusion matrix. Int. J. Eng. Tech. Res. 4 (6), 75–78.
ODP, 2007. ODP Prime Scientific Data: Collection, Archive, and Quality ODP. Technical
Note 37. URL: http://www-odp.tamu.edu/publications/tnotes/tn37/TNOTE_37.
PDF.
Orozco-del-Castillo, M.G., Ortiz-Aleman, C., Urrutia-Fucugauchi, J., Rodr0 ıguez-
Castellanos, A., 2011. Fuzzy logic and image processing techniques for the
interpretation of seismic data. J. Geophys. Eng. 8, 185–194. https://doi.org/
10.1088/1742-2132/8/2/006, 2017.
Papernot, N., McDaniel, P., Jha, S., Fredrikson, M., Celik, Z.B., Swami, A., 2016. The
limitations of deep learning in adversarial settings. In: European Symposium on
Security and Privacy. https://doi.org/10.1109/EuroSP.2016.36, 2017, Germany.
Peng, H., Bai, X., 2017. Limits of Machine Learning Approach on Improving Orbit
Prediction Accuracy Using Support Vector Machine. Advanced Maui Optical and
Space Surveillance (AMOS) Technologies Conference, Hawaii, 2017.
Peng, J., Zhou, Y., Chen, C.L.P., 2015. Region-kernel-based support vector machines for
hyperspectral image classification. IEEE Trans. Geosci. Rem. Sens. 9 (53),
4810–4824. https://doi.org/10.1109/TGRS.2015.2410991.
Pirrone, M., Battigelli, A., Ruvo, L., 2014. Lithofacies Classification of Thin Layered
Turbidite Reservoirs through the Integration of Core Data and Dielectric Dispersion
Log Measurements. https://doi.org/10.2118/170748-MS.
Puggini, L., Doyle, J., Mcloone, S., 2015. fault detection using random forest similarity
distance. IFAC-PapersOnLine 48 (21), 583–588. https://doi.org/10.1016/j.
ifacol.2015.09.589.
Rafik, B., Kamel, B., 2017. Prediction of permeability and porosity from well log data
using the nonparametric regression with multivariate analysis and neural network,
Hassi R’Mel Field, Algeria. Egyptian J. Petrol. 26 (3), 763–778. https://doi.org/
10.1016/j.ejpe.2016.10.013.
13
Computers and Geosciences 139 (2020) 104475
Rahim, I.A., Tahir, S. Hj, Musta, B., Omang, S.A.K., 2009. Lithological unit thickness
approach for determining intact rock strength (IRS) of slope forming rock material of
crocker formation. Borneo Sci. 25, 23–32.
Raschka, S., 2015. Python Machine Learning. Packt Publishing Ltd, England, p. 454pp.
Romppanem, S., H€ akk€anem, H., Kaski, S., 2017. Singular value decomposition approach
to the yttrium occurrence in mineral maps of rare earth element ores using laser-
induced breakdown spectroscopy. Spectrochim. Acta, Part B 134, 69–74. https://doi.
org/10.1016/j.sab.2017.06.002, 2017.
Rosid, M.S., Haikel, S., Haidar, M.W., 2019. Carbonate reservoir rock type classification
using comparison of Naïve Bayes and Random Forest method in field “S” East Java.
AIP Conf. Proc. 2168, 020019 https://doi.org/10.1063/1.5132446, 2017.
S�
aez, J.A., Galar, M., Luengo, J., Herrera, F., 2013. Tackling the problem of classification
with noisy data using Multiple Classifier Systems: analysis of the performance and
robustness. Inf. Sci. 247 https://doi.org/10.1016/j.ins.2013.06.002.
Sahoo, S., Jha, M.K., 2017. Pattern recognition in lithology classification: modeling using
neural networks, self-organizing maps and genetic algorithms. Hydrogeol. J. 25,
311–330. https://doi.org/10.1007/s10040-016-1478-8, 2017.
Souza, J.F.L., Santos, M.D., Magalh~ aes, R.M., Neto, E.M., Oliveira, G.P., Roque, W.L.,
2019. Automatic classification of hydrocarbon ‘‘leads’’ in seismic images through
artificial and convolutional neural networks. Comput. Geosci. 132, 23–32. https://
doi.org/10.1016/j.cageo.2019.07.002.
Spronck, P., 2017. The Coder’s Apprentice - Learning Programming with Python 3.
Spronck Create Commons Licences, p. 398. Version 1.0.16. http://www.spronck.
net/pythonbook/pythonbook.pdf.
Storkey, A.J., 2013. When Training and Test Sets Are Different: Characterising Learning
Transfer. https://doi.org/10.7551/mitpress/9780262170055.003.0001.
Strauß, S., 2018. From big data to deep learning: a leap towards strong AI or ‘intelligentia
obscura’? Big Data Cognitive Computing 2 (3), 16. https://doi.org/10.3390/
bdcc2030016.
Tharwat, A., 2018. Classification assessment methods. Appl. Comput. Inform. https://
doi.org/10.1016/j.aci.2018.08.003.
Tilaki-Hajian, K., 2013. Receiver operating characteristic (ROC) curve analysis for
medical diagnostic test evaluation. Caspian J Intern Med 4 (2), 627–635pp.
Vakhshoori, V., Zare, M., 2018. Is the ROC curve a reliable tool to compare the validity of
landslide susceptibility maps? Geomatics, Nat. Hazards Risk 1 (9). https://doi.org/
10.1080/19475705.2018.1424043, 2018.
Vapnik, V.N., 1998. Statistical Learning Theory. A Wiley-Interscience Publication, New
York City, U. S., p. 740pp
Venkataraman, S., 2017. System Design for Large Scale Machine Learning. Electrical.
University of California at Berkeley, EECS Department. https://www2.eecs.berkeley.
edu/Pubs/TechRpts/2017/EECS-2017-219.html.
Wallet, B., Hardisty, R., 2019. Unsupervised seismic facies using Gaussian mixture
models. SEG Library 3 (7), 1A–T725. https://doi.org/10.1190/INT-2018-0119.1.
Wang, G., Carr, T.R., Ju, Y., Li, C., 2014. Identifying organic-rich Marcellus Shale
lithofacies by support vectormachine classifier in the Appalachian basin. Comput.
Geosci. 64, 52–60. https://doi.org/10.1016/j.cageo.2013.12.002, 2014.
YU, L., Porwal, A., Holden, E., Dentith, M.C., 2012. Towards automatic lithological
classification from remote sensing data using support vector machines. Comput.
Geosci. 45 https://doi.org/10.1016/j.cageo.2011.11.019.
Zhao, L.N., Tian, F.Y., Wu, H., Qi, D., Wang, Z., 2011. Verification and comparison of
probabilistic precipitation forecasts using the TIGGE data in the upriver of Huaihe
Basin. Adv. Geosci. (29), 95–102pp. https://doi.org/10.5194/adgeo-29-95-2011,
2011.