Automatic Lithology Well Logging

Automatic Lithology Prediction from Well
Logging Using Kernel Density Estimation
Anisa Noor Corina
Petroleum Engineering
Submission date: June 2016
Supervisor: Sigve Hovda, IPT
Norwegian University of Science and Technology

Department of Petroleum Engineering and Applied Geophysics
Preface
This thesis is a one of my fruitful result of a keen work for the past one semester. This
thesis is also part of the requirement for a master degree in Petroleum Engineering,
Department of Petroleum Technology and Applied Geophysics, Norwegian University
of Science and Technology (NTNU). The study described herein began in spring 2016
to the extent of 30 educational points. Apart from the efforts of myself, the success of
this study depends on the guideline of many others. I take this opportunity to thank
these people.
First, I express my deep gratitude to my supervisor, Sigve Hovda, for the time and
constructive response to the work. His enthusiasm working in this project has pre-
served the optimistic attitude and made this project possible. I would also like to thank
Erik Skogen for the supportive discussion of well logging which I found very valuable
for this work. I would also like to thank AGR company for providing access to iQx soft-
ware, including the access to the data. And, I would also like to thank MATLAB © 2015
The MathWorks, Inc.,Natick, Massachusetts, United States, for supporting the compu-
tation within the study. Among all of the benefits provided by Institute Petroleum of
Technology (IPT) NTNU, my curiosity of interdisciplinary working on this project to-
gether with these people have benefited myself more than most.
Despite the academic support, I would thank my all of my friends from Petroleum
Engineering/Petroleum Geosciences, especially cohort 2014, who became my loyal
partners during the long journey of 2 years of the master degree. I would also thank
my friends from Indonesia who bring the joy during the rough hour working on my
thesis.
Most importantly, none of this could have happened without my family. To my
mom, the person that I adore the most, I sent my deepest love and thank for being
there all the times for me. I would also to thank my dad, sister, and brother who never
stop giving me comfort even though we are miles away and I am forever grateful.
Trondheim, 2016-06-09
Anisa Noor Corina
i
Abstract
This thesis presents an automatic real-time analysis of lithology interpretation

through a method of statistical analysis: kernel probability density method. The goal
of this thesis is to develop a method for interpreting and predicting lithology from the
borehole geophysical data in real time. Prior to the development, the data is explored
to check the data quality and the requirement of data correction. In addition, from
exploratory data analysis, the data characteristics can be observed thus the best-fit
classification method can be selected. The study focuses on the univariate analysis
of gamma-ray data in classifying shale and non-shale lithology. In addition to univari-
ate analysis, a preliminary study of bivariate analysis is also provided in this thesis. The
bivariate analysis combines the gamma-ray and the neutron data.
Within the study, the models of probability density are constructed by using ker-
nel estimator. The data source for the models are extracted from 3 wells in the North
Sea, they are Well 15/5-7 A, Well 15/6-11 S, and Well 15/6-9 S. The application of ker-
nel method on gamma-ray data returns a good estimation and appropriates the non-
parametric distribution of the data. There are two different types of model constructed
based on the type of classification rule. The first model is constructed solely using
gamma-ray data. While, the second model is constructed by combining gamma-ray
data and geological description which is represented as prior probability value in the
classification rule. Once the models are ready, the models are validated and tested with
a set of testing data in order to assess the misclassification rate.
There are three different experiments performed based on the source of the testing
data set. These experiments are executed in order to assess how precise is the model
classifying lithology by using testing data from different well locations. One of the ex-
periment tests the models with a dataset taken from the same source of the model.
Meanwhile, the other two experiments test the models with a dataset taken from the
different source of the models. The validation shows promising result and proves that
gamma-ray is a representative variable in classifying lithology.
The evaluation techniques within this study can be applied in the practice to inter-
pret and predict the lithology. By applying the technique, it is expected that the reading
from the logging tools can be processed and the result of lithology type with the pre-
diction can be automatically returned in the surface. The application is also beneficial
to reduce the time required for lithology interpretation during drilling operation.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
1 Introduction 1
2 Background Theory 4
2.1 Well Logging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Gamma Ray Logging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Neutron Log . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Boxplot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Hypothesis testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Motivation of hypothesis testing . . . . . . . . . . . . . . . . . . . . . 12
2.4 Kernel density estimation (KDE) . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Motivation of univariate density estimation . . . . . . . . . . . . . . 15
2.4.2 Properties of kernel density estimator . . . . . . . . . . . . . . . . . . 16
2.5 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Methodology and data 21

3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Data management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.1 Data source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.2 Filtering and collecting data . . . . . . . . . . . . . . . . . . . . . . . 21
4 Data exploration on GR data 25

4.1 Data exploration results and discussions . . . . . . . . . . . . . . . . . . . . 25
4.1.1 GR data exploration under lithology grouping . . . . . . . . . . . . . 25
4.1.2 GR data exploration under lithology grouping and hole size sub-
group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2 Hypothesis testing results and discussions . . . . . . . . . . . . . . . . . . . 38
4.2.1 Hypothesis testing #1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.2 Hypothesis testing #2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5 Univariate KDE analysis on GR data 43

5.1 KDE analysis on GR data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.1.2 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Validation of probability density on GR data by using KDE . . . . . . . . . . 51
ii
CONTENTS
5.2.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2.2 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3 Preliminary study of bivariate analysis . . . . . . . . . . . . . . . . . . . . . 70
5.4 Concluding remark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6 Conclusions 73
A Acronyms 75
B Geological Description 76
C ANOVA Table Results of Hypotheses Test 80
D Results of GR probability density of two categories: shale and non-shale 83

D.1 Well 15/5-7 A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
D.2 Well 15/6-11 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
D.3 Well 15/6-9 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
E Results of bivariate analysis of GR and neutron 87

E.1 Well 16/1-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
E.2 Well 16/2-7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
E.3 Well 16/2-13 A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Bibliography 92
iii
List of Figures
2.1 Example of time-based surface logging (Bourgoyne et al., 1985) . . . . . . . 5

2.2 Gamma ray reading for various lithology (Glover, 2001) . . . . . . . . . . . 6
2.3 Comparison of shale volume by using different methods (Glover, 2001) . . 7
2.4 The graph of neutron life after neutron is emitted by the neutron tool . . . 8
2.5 Typical apparent porosity from neutron log for varies lithologies . . . . . . 9
2.6 Histograms of a data set with 15, 35, and 100 bins (Scott, 2004) . . . . . . . 10
2.7 An example of boxplot with whiskers of 1.5 IQR of upper and lower quartiles 12
2.8 (a) Simple density estimation, and (b) smooth density estimation (with
underlying Gaussian kernel function) for certificated of deposit (CR) rates 16
2.9 Distribution of K (u) in various kernel functions . . . . . . . . . . . . . . . . 18
2.10 Description of misclassification regions P (1|2) and P (2|1) . . . . . . . . . . 20
3.1 Flowchart of methodology in this study . . . . . . . . . . . . . . . . . . . . . 22

3.2 The spreadsheet for recording the availability of logging data in some of
the wells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 The interface for data visualization of Well 15/5-7 A, Well 15/6-11 S, and
Well 15/6-9 S. In each well, the geophysical data were plotted in log
traces, alongside casing data and geological description. . . . . . . . . . . . 24
4.1 Graphical description of GR data in Well 15/5-7 A which grouped based

on the lithology type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Graphical description of GR data in Well 15/6-11 S which grouped based
4.3 Graphical description of GR data in Well 15/6-9 S which grouped based
4.4 Shifted GR value from logging visualization: (a) 26" (blue area) and 17 12 "
(red area) in Well 15/5-7 A and b) 17 21 " (blue area) and 12 14 " (red area)
in Well 15/6-11 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.5 Graphical description of GR data in Well 15/5-7 A which grouped accord-
ing to the lithology type and hole size . . . . . . . . . . . . . . . . . . . . . . 35
4.6 Graphical description of GR data in Well 15/6-11 S which grouped ac-
cording to the lithology type and hole size . . . . . . . . . . . . . . . . . . . 36
4.7 Graphical description of GR data in Well 15/6-9 S which grouped accord-
ing to the lithology type and hole size . . . . . . . . . . . . . . . . . . . . . . 37
5.1 Density probability plots with KDE and kernel bandwidth results of Well
15/5-7 A for present lithology in borehole size: (a)36", (b)26", (c)17 12 ",
and (d)8 12 " . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
iv
LIST OF FIGURES
15/6-11 S for present lithology in borehole size: (a)36", (b)26", (c)17 12 ",
(d)12 41 ", and (e)8 21 " . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
15/6-9 S for present lithology in borehole size: (a)24", (b)17 21 ", and (c)8 12 " 47
5.4 Comparison of histogram and KDE in estimating probability density in
(a)Well 15/5-7 A , section 26", (b)Well 15/6-9 S , section 17 12 ", and (c) sec-
tion 24" . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.5 Probability density of Well 15/5-7 A for hole size 17 12 ". The gray area is
the overlapping distribution between shale and sandstone lithology. . . . 49
5.6 Comparison of probability density from group of 8 12 " in Well 15/6-9 S
which are (a) not-merged and (b) merged . . . . . . . . . . . . . . . . . . . . 50
5.7 Preview of the interface for validation process built by using MATLAB . . . 54
5.8 Training and testing data distributions from different experiments for
group of 17 12 ". The testing data was from Well 15/6-9 S . . . . . . . . . . . . 64
5.9 Plot of probability density for case in experiment 1: 17 12 " in Well 15/6-9 S . 66
5.10 Distribution of model 1, model 2, and testing data of experiment 1 group
26" in Well 15/5-7 A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.11 Distribution of model 1, model 2, and testing data of experiment 2 for
group 8 12 ". The model was from Well 15/6-11 S and the testing data was
from Well 15/6-9 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.12 Bivariate analysis plots of group 8 12 " from Well 16/1-14 . . . . . . . . . . . . 71
5.13 Bivariate analysis plots of group 17 21 " from Well 16/2-7 . . . . . . . . . . . . 71
D.1 Probability density results of each hole section in Well 15/5-7 A using
kernel estimator grouped into shale and non-shale lithology . . . . . . . . 84
D.2 Probability density results of each hole section in Well 15/6-11 S using
kernel estimator grouped into shale and non-shale lithology . . . . . . . . 85
D.3 Probability density results of each hole section in Well 15/6-9 S using ker-
nel estimator grouped into shale and non-shale lithology . . . . . . . . . . 86
E.1 Scatter plot and probability density plots of GR and neutron for Well
16/1-14 in hole section: (a) 12 14 " and (b) 8 21 " . . . . . . . . . . . . . . . . . 88
16/2-7 in hole section: (a) 17 12 ", (b) 12 41 ", and (c) 8 21 " . . . . . . . . . . . . 90
16/2-13 A in hole section: (a) 12 14 " and (b) 8 21 " . . . . . . . . . . . . . . . . 91
v
List of Tables
2.1 Methods of calculating shale volume . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Various kernel functions and forms . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Classification probabilities table . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1 Statistic description of GR data in Well 15/5-7 A grouped according to the

present lithology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Statistic description of GR data in Well 15/6-11 S grouped according to
4.3 Statistic description of GR data in Well 15/6-9 S grouped according to
4.4 Statistic description of Well 15/5-7 A grouped according to the present
lithology and subgroup of hole size . . . . . . . . . . . . . . . . . . . . . . . 32
4.5 Statistic description of Well 15/6-11 S grouped according to the present
4.6 Statistic description of Well 15/6-9 S grouped according to the present
4.7 P-value of the first hypothesis test for indicating any lithology group in
GR data. The detailed lithology group investigated in each well is pro-
vided in Table 4.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.8 Mean ranks of each lithology group in Well 15/5-7 A, Well 15/6-11 S, and
Well 15/6-9 S from the first hypothesis test . . . . . . . . . . . . . . . . . . . 39
4.9 P-value result of the second hypothesis test. Empty (-) p-value results in-
dicated that there are no hole size group found in the particular lithology
group. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.10 Mean rank for each category. Empty (-) mean rank results indicated that
there are no mean rank for selected hole size and lithology group. . . . . . 41
5.1 Confusion matrix table of 2 sub-population, π1 and π2 . . . . . . . . . . . . 52

5.2 Summary of data and results for validation within each hole section in
Well 15/5-7 A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Well 15/6-11 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Well 15/6-9 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.5 Summary of data and results for validation of Well 15/6-11 S and Well
15/6-9 S , by using Well 15/5-7 A as training data . . . . . . . . . . . . . . . 56
5.6 Summary of data and results for validation of Well 15/5-7 A and Well
15/6-9 S , by using Well 15/6-11 S as training data . . . . . . . . . . . . . . . 57
vi
LIST OF TABLES
5.7 Summary of data and results for validation of Well 15/5-7 A and Well
15/6-11 S , by using Well 15/6-9 S as training data . . . . . . . . . . . . . . . 58
5.8 Summary of data and results for validation of wells in Block 16 by using
Well 15/5-7 A as training data . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Well 15/6-11 S as training data . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Well 15/6-9 S as training data . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.11 Table of minimum, maximum, and averaged value of misclassification
error in each model and experiment . . . . . . . . . . . . . . . . . . . . . . . 62
5.12 The values of the intersection points of the models and the testing data
from experiment 1, group of 17 12 " in Well 15/6-9 S . . . . . . . . . . . . . . . 67
C.1 ANOVA table of hypothesis testing #1 in Well 15/5-7 A . . . . . . . . . . . . 81

C.2 ANOVA table of hypothesis testing #1 in Well 15/6-11 S . . . . . . . . . . . . 81
C.3 ANOVA table of hypothesis testing #1 in Well 15/6-9 S . . . . . . . . . . . . 81
C.4 ANOVA table of hypothesis testing #2 for shale lithology in Well 15/5-7 A . 81
C.5 ANOVA table of hypothesis testing #2 for sandstone lithology in Well
15/5-7 A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
C.6 ANOVA table of hypothesis testing #2 for shale lithology in Well 15/6-11 S 81
15/6-11 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
C.8 ANOVA table of hypothesis testing #2 for shale lithology in Well 15/6-9 S . 82
15/6-9 S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
vii
Chapter 1
Introduction
Subsurface lithology within drilling operation is interpreted with the usage of dif-
ferent techniques, involving seismic records, mud logging, and well logging. In the
petroleum industry, well logs are the main source of information concerning the sub-
surface formations. And, the variation in borehole geophysical data usually is used
to relate any change in lithology and geological properties. In addition to mud log-
ging, real-time drilling measurement nowadays is improving through the development
of measurement-while-drilling. Moreover, logging within drilling phase is even possi-
ble through logging-while-drilling (LWD). Continuous transmission of the information
from MWD and LWD has shown the benefits in helping the drilling decisions and real-
time formation evaluation (Bonner et al., 1992).
Despite of well logging, mud logging also provides information for lithology inter-
pretation. Within mud logging, measurement of the progress of the drilling operation
and the contents of the formation are recorded. The drilling parameters which are
recorded include weight on bit (WOB), hook load, and mud properties. Meanwhile,
the cuttings from the drilled formation inside borehole is circulated to the surface. By
visualizing the cutting sample, the lithology can be approximated. Generally, cuttings
samples require an amount of time to be circulated to the surface which lead to a de-
layed interpretation. Restoring the information from cuttings also can be difficult due
to the requirement of correlating the cutting origin, fluid loss in the borehole, flushed
rock fragment. These factors cause cutting visualization alone does not provide an ac-
curate lithology interpretation and requires a combination with information from well
logs and mud logs.
A wide span of methods in lithology interpretation has been proposed through
combinations of various measurements both in the qualitative and quantitative eval-
uation. Some of the qualitative evaluations from mud logging include ROP interpreta-
tion (such as identification of drilling break and drill-off trend), drilling force, and bit
evaluation, and specific energy deflection (Provost (1987), Ziaja and Roegiers (1998),
Laosripaiboon et al. (2015) ). Meanwhile, qualitative evaluations from logging mea-
surement include visualizations of multiple logs , photoelectric (P e ) factor interpre-
tation, and gamma-ray evaluation for shale identification (Gardner and Dumanoir
(1980), Serra et al. (1985), Dewan (1986) ). Qualitative methods itself are inadequate
for lithology interpretation, especially in formation with complex lithologies which re-
quire a large set of logging information.
Within time, lithology interpretation has expanded and starts to consider the usage
of quantitative methods, such as crossplot, statistical analysis, and neural network.
1
Crossplot is one of the basic method of quantitative evaluation which is performed by
plotting data points of two or more than two different log data. The geophysical data
which are widely used for crossplotting are density, neutron, sonic, and P e . The types
and the applications of crossplot method has been studied by Burke et al. (1969) and
Clavier and Rust (1976). However, these methods still require manual analyses and can
not be applied for automatic interpretation.
There are a plethora of statistical classification methods, such as discrimination
analysis, linear regression, kernel estimation, etc. The selection between these meth-
ods is greatly dependent on the character of the data, so that the method can pro-
vide an interpretation that is important for model building. Probabilistic classification
is a common classification method which can predict the belonging of a member by
returning the probability. An early research by Delfiner et al. (1987) has shown how
statistical analysis can be applied for lithology determination and prediction. This re-
search proposed a procedure which combines modern wireline measurement in order
to produce automatic lithologic description. Within the procedure, lithology classifi-
cation by discriminant analysis and probability calculation by Bayesian rule was in-
troduced. The application of this research was shown in a case study by Busch et al.
(1987). This research showed that the statistical discriminant analysis is possible to
predict lithology formation. However, the proposed method in these researches is lim-
ited for geophysical data with normal (Gaussian) distribution thus, this method is not
flexible to be applied in non-parametric distribution.
A statistical method called kernel density estimator provides an estimation of the
probability density function for a non-parametric distribution and examines the mul-
timodality of data. A research by Silverman (1986) showed that kernel density estima-
tion is the excellent tool for estimating the univariate, bivariate, or trivariate data. The
application of kernel density estimation in borehole geophysical data was performed
by Mwenifumbo (1993). Within this research, the kernel density method was applied
for analysis of univariate and bivariate data to identify lithology and sulfide mineral-
ization. However, the assessment of the statistical significance was not performed.
Until now, there has been no study of lithology prediction based on borehole geo-
physical data which applied statistical analysis. Within this study, a method of lithology
classification and prediction from borehole geophysical data is developed by applying
kernel density method as the statistical analysis. The objective of this study is that the
developed method can be applied to give interpretation and prediction in a real-time
operation. Specifically, the main focus of this study is the application of univariate
kernel density which is not extensively used. The models from gamma-ray data con-
structed by kernel estimator are assessed by the used of the confusion matrix to un-
derstand the model accuracy in classifying shale and non-shale lithology. The results
are presented in term of misclassification rate. In addition to that, a brief insight of
bivariate analysis to improve lithology classification is also provided.
The models are tested with two different classification rules to assess the effect from
adding prior probability value which value is taken from the geological description.
The models are also tested in different experiments with testing datasets which are
taken from different well locations. This study shows that gamma-ray is a good variable
for lithology classification. But, the accuracy is dependent on the source of data which
is tested towards the models. Moreover, the method proposed in this study can be
applicable for lithology prediction, even though the application is still limited because
it processes the data from the current depth of logging tool, not beyond the logging
2
CHAPTER 1. INTRODUCTION
tool. However, this application can be beneficial to improve lithology interpretation

during the drilling operation and provides an automatic lithology prediction.
The rest of report is structured as follows. Chapter 2 introduces the theory of the
well logging and several methods of statistical analysis, they are exploratory data anal-
ysis, kernel density estimation, and classification. Chapter 3 introduces the source
of the data which is used within this study and the data management. Chapter 4 in-
troduces the data exploration of gamma-ray data and hypothesis testing on the data.
Chapter 5 introduces the application of kernel density method into gamma-ray data
and the validation of models for classification purpose.
3
Chapter 2
Background Theory
This purpose of this chapter is to discuss the theories which relevant to this study. The
first theory discusses the well logging in petroleum industry, including the descrip-
tion of gamma ray and neutron logging which are relevant to the geophysical data
which is used in this study. The next theories discussed are relevant to the method-
ology adapted in this study, they are exploratory data analysis, hypothesis testing, and
kernel density estimation.
2.1 Well Logging

In the petroleum industry, well logging plays a crucial role as a tool to interpret down-
hole conditions. Well logging is divided into two types, surface and downhole logging.
Surface logging records all information during drilling operation through sensors lo-
cated at the surface. Meanwhile, downhole logging records information from sensors
located at the downhole tools. The information recorded in surface logging are (i)
drilling parameters, such as hookload, torque, and Rate of Penetration (ROP), (ii) mud
returns, and (iii) cuttings from downhole. By closely monitoring the surface measure-
ments, events occurring in the borehole can be identified by looking at values outside
of the normal ranges. In addition, cuttings from downhole can be used to describe the
geological properties and detect hydrocarbon traces (Wilson, 1955).
Downhole logging, also called as wireline logging, measures the properties of rocks
surrounds the borehole, such as rock radioactivity, resistivity, etc. The tool is sus-
pended on a cable or wire and can be run between drilling operations and at the end
of drilling. The recent development allows wireline logging to be run during a drilling
operation, which is called as Logging while Drilling (LWD). By combining LWD with
the Measurement while Drilling (MWD) system, information can be transmitted from
downhole to surface almost continuously during the drilling operation.
Some of the wireline tools measure properties that give a direct result and do not
require to be interpreted, while some of the others require interpretation. Most of the
times, the interpretation requires a collaboration of results from different wireline tools
because each tool has a limited measurement and the results can be masked by the
rock or fluid properties. As an example, resistivity measurement is affected by forma-
tion temperature because the resistivity tools are not sensitive to temperature. This
event can lead to misinterpretation of reservoir fluids in the formation. Therefore, a
combination of resistivity and temperature data will give a better interpretation.
4
CHAPTER 2. BACKGROUND THEORY
Figure 2.1: Example of time-based surface logging (Bourgoyne et al., 1985)
2.1.1 Gamma Ray Logging

Gamma Ray (GR) tool measures the natural radioactivity of minerals contained in the
rocks. Most of the rocks contain natural occurring radioactive elements, such as potas-
sium, uranium, and thorium in different amounts, and all of these emit gamma rays
(Schlumberger Educational Services, 1989).
GR log is useful for correlating zones from one well to others and indicating shale in
the formation, due to high content of radioactive minerals in shale. A rough estimation
of clay volumes (Vcl ) can be calculated using GR reading. By setting sand point, min-
imum GR reading (γmi n ) which indicates 100 % sand content, and shale point, maxi-
mum GR reading (γmax ) which indicates 100 % shale content, GR index (IGR ) can be
calculated by linear scaling (Ellis and Singer, 2010) :
γl og − γmi n
IGR = (2.1)
γmax − γmi n
Poupon and Gaymard (1970) proposed that shale volume is equal with IGR . Beside
linear scaling, there are several different approaches that consider the effect of clay
distribution in the reservoir rock, clay mineral, and clay bound. These methods are
summarized in Table 2.1 and visualized in Figure 2.3.
5
2.1. WELL LOGGING
Figure 2.2: Gamma ray reading for various lithology (Glover, 2001)
6
Method Equation
Clavier et al. (1971) Vsh = 1.7 − [3.38 − (IGR + 0.7)2 ]0.5
Larionov (1969), for tertiary rock Vsh = 0.083 × [2(3.7058×IGR ) − 1]

Larionov (1969), for older rock Vsh = 0.33 × (2IGR − 1)
3IGR
Stieber (1970) Vsh =
(1 + 2IGR )
Table 2.1: Methods of calculating shale volume
Figure 2.3: Comparison of shale volume by using different methods (Glover, 2001)
However, determining lithology shaliness only by using GR index can cause misin-
terpretations, such as in cases of uranium-rich formations, sandstone containing mica,
and nonradioactive clays. These misinterpretations can be prevented by the use of
spectral gamma ray which measures not only the total radioactivity, but also the con-
centration of potassium (K), thorium (Th), and uranium (U).
GR tools are sensitive to a number of factors (Bateman, 2012):
1. Eccentricity of gamma ray tool
2. Hole size
3. Mud weight
4. Casing weight and size
5. Cement thickness
Modern logs usually have automatic corrections applied to GR readings. However,
7
2.1. WELL LOGGING
a set of correction charts is available to correct GR manually to environmental condi-

tions such as hole size and mud weight (Schlumberger Wireline & Testing, 1998). The
corrections are considered to be crucial so that the logging data can be representative.
2.1.2 Neutron Log

In general, neutron log measures the amount of hydrogen in formations. The neutron
tool releases high energy neutron into the formation which will scatter elastically with
nuclei. The energy will be reduced to the thermal energy level (≈ 0.025eV) and then
the neutron will be absorbed by the nucleus while emitting γ-rays.
Figure 2.4: The graph of neutron life after neutron is emitted by the neutron tool
The energy loss due to elastic scattering is maximum when a neutron collides nu-
cleus with the same mass (i.e. hydrogen). Therefore, the count rate to slow down the
neutron and the distance traveled by neutron depend on the amount of hydrogen. Be-
cause hydrogen is mostly found in pores (composed in water or hydrocarbon), the neu-
tron log is related to porosity function. The count rate in high porosity rocks is slower
than in low porosity rocks.
Mainly, there are three different types of neutron tools available, they are:
1. The gamma ray/neutron tool (GNT)
2. The sidewall neutron porosity tool (SNP)
3. The compensated neutron log (CNL)
The detailed explanation of each tool can be found in a textbook written by Bateman
(2012). The data measured by neutron tools are shown in the unit of porosity or hy-
drogen index. The neutron tools are calibrated in limestone filled with fresh water.
Therefore, the results often presented in equivalent limestone porosity units.
The neutron logs are used for porosity calculation which assumes that the contri-
bution of elements other than hydrogen is negligible. The second use of neutron logs is
lithology determination. The source of a hydrogen atom is not only from fluids occupy-
8
ing the pore space but it can be originated from bound water molecules in shales, crys-
tallized water in evaporites, or hydrated minerals in igneous and metamorphic rocks
(Glover, 2001).
The apparent porosity of shale is varying, but it is usually higher than the apparent
porosity identified in carbonate and sandstone rocks. This high porosity reading by
neutron tool is caused by the effect of hydrogen contained in the bound water in shale.
However, shale identification by using neutron log requires extra concern due to the
effect of hydrocarbon gas which may present and disturbs the log.
Figure 2.5: Typical apparent porosity from neutron log for varies lithologies
2.2 Exploratory data analysis

The motivation of data exploration is to extract any important information of the data
and understand the data behavior. This step is found to be crucial because we would
set assumptions and test hypotheses within this process (Tukey, 1977). The method
adapted in data exploration is called as Exploratory Data Analysis (EDA). This method
9
2.2. EXPLORATORY DATA ANALYSIS
provides the most appropriate way to explore, summarize data, and also create a visu-
alization of the data. By implementing EDA, it is expected to gain some confidences of
the data. Most of the EDA use graphical techniques to get data visualization, such as
histograms, box plots, and steam-leaf plots.
2.2.1 Histogram
Histogram is one of graphical techniques which serves the data by using bars to show
data distribution. During the construction, the data values are divided into series of
intervals which illustrated in bars. The number of intervals are often referred as bin,
which each interval contains a certain range of values. Each bar usually has a consis-
tent and equal ranges with others, and its interval is not overlapping with others and
adjacent.
Histograms are often constructed as frequency or density histogram. The descrip-
tion of each type of histogram is explained below:
Class frequency histogram measures the number of occurrences which values are
falling within a given class interval.
Relative frequency histogram calculates the frequency of each interval divided by the
total number of measurements. This histogram has a total height of bars equal
to 1 or 100%.
Density histogram calculates the relative frequency of each interval divided by bin
width. In other words, density histogram shows relative frequency as area of each
bar. In density histogram, the total area of bars is normalized to 1.
Histogram is very well suited to illustrate a large set of data and continuous data.
With the advantage of its simplicity of construction, histogram is a graphical technique
which is commonly used. However, describing data using histogram may encounter
some difficulties and requires some understandings. A histogram is very sensitive to
bin width because of its effect to the graph smoothness. It is evidenced that larger bin
width (fewer bins) reduces noises and makes the graph oversmoothed, while smaller
bin width (more bins) makes the graph undersmoothed.
Figure 2.6: Histograms of a data set with 15, 35, and 100 bins (Scott, 2004)
An example is taken from a paper written by Scott (2004) to give an illustration of

the effect of the bin width. A set of data with 21,640 data points are constructed using
histograms with 3 different bins Figure 2.6). Undersmoothed histogram contains high
variability in value even though it has a smaller bias, while oversmoothed histogram
has the opposite effect. Thus, choosing the most optimal bin width is crucial to avoid
misinterpretation (Simonoff, 1996).
10
There are many theories have been developed in determining the optimum bin
width. However, the application of these methods depends on the data distribution
and the goal of analysis. Some of the methods proposed calculation of bin numbers, k,
and some proposed calculation of bin width, h. The relationship of bin numbers and
bin width is:
x max − x mi n
k= (2.2)
h
where x is random variable. The most common methods on calculating optimum bin
width are explained below.
Sturges method. This method suggests to calculate bin width with formula (Scott,
1992),
k = 1 + l og 2 n (2.3)
where n is the number of data points. Sturges derived the formula above based
on binomial distribution with normal distribution data. This method is popu-
lar due to its simplicity in the calculation. However, this method may give poor
result if the data is not normal.
Scott method. This method is derived by minimizing the integrated means squared
error of the density estimate for normally distributed data (Scott, 1992). The cal-
culation requires standard deviation 1 , σ.
3.491σ
h= (2.4)
n 1/3
Freedman-Diaconis. This method is suitable for data containing a large number of

outliers or heavy-tailed distribution. The method is a modification of Scott
method, replacing σ with parameter interquartile range (IQR), the distance be-
tween the lower and upper quartiles (Scott, 1992). Description of quartile can be
found in Chapter 2.2.2.
2(IQR)
h= (2.5)
n 1/3
Histogram is also referred as simple univariate density estimator to approach prob-

ability density function. To improve the smoothness in estimate density function, the
application of histogram often combined with kernel density estimator. The detailed
explanation of kernel estimator will be covered under 2.4.
2.2.2 Boxplot
Boxplot is a graphical technique to examine the shape of data distribution by using
parameters called quartiles. In addition, boxplot is also used to study the variability of
values in a set of data (Ott and Longnecker, 2010).
Quartiles are three points dividing a dataset which is arranged from the lowest to
the highest value equally into 4 groups. The first quartile (Q 1 ), which is called lower
quartile, has a value between the smallest value and median of a dataset. The second
quartile (Q 2 ) is called as the median of a dataset. The median value is taken from a
qP
2
1 i (x−x̄)
Standard deviation measures the variation of data set ,σ = n−1 (Ott and Longnecker, 2010)
11
2.3. HYPOTHESIS TESTING
data point located in the middle of a dataset which is arranged from the lowest to the
highest value. In other words, a median is the center of data distribution. Last, third
quartile (Q 3 ), which is called upper quartile, has a value between the median and the
highest value of a dataset. The difference between upper and lower quartiles value is
defined as an interquartile range.
Figure 2.7: An example of boxplot with whiskers of 1.5 IQR of upper and lower
quartiles
The boxplot is constructed by creating a box with two sides which values are equal
to the upper and lower quartiles. Within the box, a line is drawn to indicate median
(see figure 2.7). The boxplot is often constructed with whiskers to indicate data vari-
ability. There are several ways to plot the whiskers depend on the information to be
represented. In most cases, the whiskers represent the minimum and maximum val-
ues of the data. The whisker can also set at a value equal to 1.5 IQR of the upper and
lower quartiles, and this boxplot is often referred as Tukey boxplot (Frigge et al., 1989).
2.3 Hypothesis testing

2.3.1 Motivation of hypothesis testing
Hypothesis testing is a method for testing hypothesis about a group within a popula-
tion (Privitera, 2015). The hypothesis testing is started by defining the null hypothesis
(H0 ), a statement of a population parameter that is assumed to be true. Hypothesis
testing tests the null hypothesis in order to check whether the statement is likely to
be true or not. The statement which opposes the null hypothesis is called alternative
hypothesis (H1 ).
The methods to compute the test statistic are varied depends on the data character-
istic. The hypothesis test applied in this study was Kruskal-Wallis test. Kruskal-Wallis
test is a nonparametric test (distribution free) for assessing differences in a continuous
dependent variable which is presumed containing independent variables (3 or more
groups) (Kruskal and Wallis, 1952). Kruskal-Wallis test is a rank-based test which is an
extension of Mann-Whitney test. However, the test does not reveal which group of in-
dependent variables that is significantly different from each other. In other words, the
test is only limited to inform that at least two groups are different. The assumptions
required for Kruskal-Wallis test, are:
1. The dependent variable is a continuous variable.
2. The independent variables should consist three or more independent groups.
Mann-Whitney U Test is commonly used to test two groups within the popula-
tion.
12
3. The observations in each group are independent and there is no relationship be-
tween the groups.
The stated H0 of the test is that the data set comes from same distribution. Be-
fore getting into Kruskal-Wallis test, firstly we would explain the underlying theory of
rank-sum test for two independent variables from Mann-Whitney U test (Mann and
Whitney, 1947).
Mann-Whitney U test: rank sum test
Rank-sum test is a method ranking the raw data from the lowest (rank #1) to the highest
value (rank #N), with tied ranks included. Tied ranks are assigned if there are two or
more than two tied values in the raw data, thus the ranks are adjusted and equalized.
Define a population with 2 group of samples, group 1 and group 2, where n 1 is the
size of observations in group 1 and n 2 is the size of observations in group 2. The sum
rank of each group is defined with R 1 and R 2 respectively for group 1 and group 2. For
any combination of n 1 and n 2 , the maximum possible value of sum rank, R max , in each
group can be calculated as follow
n 1 (n 1 + 1)
R max1 = n 1 n 2 + (2.6a)
2
n 2 (n 2 + 1)
R max2 = n 1 n 2 + (2.6b)
2
Mann-Whitney proposed a parameter U which is equal to the difference between
maximum possible value of rank sum, R max , and the actual rank sum observed, R. The
equation U for both groups follows
U1 = R max1 − R 1 (2.7a)
U2 = R max2 − R 2 (2.7b)
For any samples sizes, n a and n b , U parameter has identities:
U1 +U2 = n 1 n 2 (2.8a)
U1 = n 1 n 2 −U2 (2.8b)
U2 = n 1 n 2 −U1 (2.8c)
U = min(U1 ,U2 ) (2.8d)
Depending on n 1 , n 2 , and level of significance, critical value of U , Ucr i t , can be calcu-

lated. If U < Ucr i t , then the test is significant and the null hypothesis is rejected.
In case where the number of samples, n 1 and n 2 , are large (both equal to or greater
than 5), U is calculated by using different approach. In this case, U approximates the
normal distribution N (µ, σ), where
n1 n2
µ= (2.9a)
ss
n 1 n 2 (n 1 + n 2 + 1)
σ= (2.9b)
12
From these identities, standardized variable, z-values, can be calculated following
|U − µ| − 0.5
z= (2.10)
σ
13
2.3. HYPOTHESIS TESTING
The value of - 0.5 is a correction for continuity to accommodate the sampling distri-
butions of U which are discrete. Then, the probability value (p-value) of z is generated
from normal distribution. If p-value < level of significance, then the null hypothesis is
rejected.
Kruskal-Wallis test
The concept of Kruskal-Wallis test is quite similar with Mann-Whitney which also
adapts the rank-sum test. Defined a population with k independent group of sam-
ples, where n = (n 1 , n 2 , . . . , n k ) represents the number of observation in each group,
R = (R 1 , R 2 , . . . , R k ) represents the sum rank of the kth group, and M = (M 1 , M 2 , . . . , M k )
represents the mean of rank of the kth group. In addition, R T is the sum of R of all k
group and M T is the sum of M of all k group, following:
k
X
RT = Rj (2.11a)
j =1
k
X
MT = Mj (2.11b)
j =1
Kruskal-Wallis measures a parameter SS bg (R) which is defined as the between-

groups sum of squared deviates based on the rank value. The conceptual formula of
SS bg (RR) is shown as
k £
n j (M j − M T )2
X ¤
SS bg (R) = (2.12)
j =1
and the computational formula is shown as
(R j )2 (R T2 )
" #
X k
SS bg (R) = − (2.13)
j =1 nj N
where N is the size of data population.

Kruskal-Wallis hypothesis test is concluded by defining a test statistic H, which
value is equal to the ratio between SS bg (R) and the mean of sampling distribution of
SS bg (R) ,
SS bg (R)
H= (2.14)
N (N + 1)/12
An alternative way to write the formula H above is
Ã !
12 k (R )2
X j
H= − 3(N + 1) (2.15)
N (N + 1) j =1 n j
The p-value can be approximated by inputting the calculated H into chi-square distri-
bution because H value has a close approximation to the chi-square distribution for
d f = k − 1, or following
H ∼ χ2(k−1) (2.16)
where χ2 is chi-squared distribution. If the returned p-value is less than level of signifi-
cance (typically set at 5%), then we rejected the statement of the null hypothesis. Small
values of p-value remove the doubt of the validity of H0 .
14
2.4 Kernel density estimation (KDE)

2.4.1 Motivation of univariate density estimation
The fundamental concept underlining the analysis of univariate data is the probability
density function for non-parametric distribution (Simonoff, 1996). The density func-
tion of a random variable X which has probability density function f (x) is shown as
Z b
P (a < X < b) = f (u)d u (2.17)
a
By using the definition of density function, an estimation of density function can
be constructed. There are two types of probability density estimator: simple density
estimator, which is often referred as histogram, and smooth density estimator.
Simple density estimator
Recall forward approximation of density function,
d F (x + h) − F (x)
f (x) ≡ F (x) ≡ lim , (2.18)
dx h→0 h
where F (x) represents the cumulative distribution function of X .
Assume that density f consists random samples with size n which samples are in-
dependent and identically distributed, represented as {x 1 , . . . , x n }. By dividing Equa-
tion 2.18 into a set of K bin numbers with width h and replace F (x) with the empirical
cumulative distribution function,
#{x i ≤ x}
F̂ (x) = , (2.19)
n
This equation leads histogram to be a density function estimator with each bin
value equal to
(#{x i ≤ b j +1 } − #{x i ≤ b j })/n

fˆ(x) = , x ∈ (b j , b j +1 ], (2.20)
h
where x ∈ (b j , b j +1 ] is the boundaries of j th bin. In simpler way, density estimator of
histogram can also be defined as
nj
fˆ(x) =
, x ∈ (b j , b j +1 ], (2.21)
nh
where n j represents the number of observations in j th bin and bin width h = b j +1 −b j .
Histogram is considered as the simplest method to estimate the distribution of uni-
variate data. However, the shortcomings of using histogram are that the histogram is
not giving a smooth estimation and not sensitive to f . Histogram may also distort de-
pending on the bar width.
Smooth univariate estimator
Recall central approximation of density function,
d F (x + h) − F (x − h)
f (x) ≡ F (x) ≡ lim , (2.22)
dx h→0 2h
15
2.4. KERNEL DENSITY ESTIMATION (KDE)
Different from histogram, the smooth estimator approaches the density function
by estimating the derivative at each point x separately. By replacing F (x) with empiri-
cal cumulative distribution,
{#x i ∈ (x − h, x + h)}
fˆ(x) = (2.23)
2nh
The equation above also can be written as
1 X n ³x −x ´
i
fˆ(x) = K , (2.24)
nh i =1 h
where
(
1
2
, if − 1 < u ≤ 1,
K (u) =
0, otherwise.
The Equation 2.24 is the form of kernel density estimator, with uniform kernel func-
tion K . This estimator fˆ(x) counts the percentage of the observations in each data
point over the local neighbourhood which is close to the examined data point x. By
merging all of the smooth kernel function at each data point, we will have a smooth
density estimation for one population.
The comparison of simple and smooth density estimator in estimating density
function can be seen in Fig. 2.8. The example given by Simonoff (1996) shows the com-
parison of probability density function from kernel estimator and histogram. Kernel
estimator gives a estimation which is smoother compared to discreteness of the his-
togram.
(a) (b)
Figure 2.8: (a) Simple density estimation, and (b) smooth density estimation (with
underlying Gaussian kernel function) for certificated of deposit (CR) rates
2.4.2 Properties of kernel density estimator

Kernel density estimator is dependent on the kernel function, K , and the bandwidth, h.
The bandwidth h is often referred as smoothing parameter which control the smooth-
ness of the data function. A very small bandwidth will give undersmoothed estima-
16
tion with more peaks and bumps, meanwhile a very large bandwidth will give over-
smoothed graph. Bandwidth also have strong relation to bias and variance which cre-
ates a dilemma in selecting optimal bandwidth. A small bandwidth will reduce the bias
of fˆ(x), but it will trigger larger variance of fˆ(x), and vice versa. The criterion on choos-
ing optimal bandwidth mostly quantified through the measurement of mean squared
error (MSE).
¤2
MSE fˆ(x) = E f fˆ(x) − f (x)
£ ¤ £
= Bias2 fˆ(x) + Var fˆ(x)

£ ¤ £ ¤
£ 00 ¤2 (2.25)
f (x)R(K ) h σK f (x)
4 4
− + + O(n −1 ) + O(h 6 )
nh 4
By integrating MSE over the entire line, we will get MISE (integrated mean squared
error). And the asymptotic MISE (AMISE) will follow
R(k) h σK R( f )
4 4 00
AMISE(h) = + (2.26)
nh 4
where R(K ) = K (u)2 du, K satisfies condition u 2 K (u)du = σ2K > 0, and f 00 is the
R R
second derivative of density function f . The optimal bandwidth, h 0 , is selected by

minimizing AMISE through differential equation and resulting
" #1/5
R(K )
h0 = n −1/5 (2.27)
σ4K R( f 00 )
and the minimum AMISE follow
5
AMISE0 = [σK R(K )]4/5 R( f 00 )1/5 n −4/3 (2.28)
4
The formula of h 0 consists an unknown density function f which value is out of
the control of data analyst, thus this formula can not be applied directly. However,
the term [σK R(K )[4/5 can be minimized by choosing the optimal kernel function K ,
and this kernel function is often called as Epanechnikov kernel. Several other kernel
functions which commonly user are shown in Table 2.2 and Fig. 2.9.
By comparing the inefficiency of each kernel function to Epanechnikov kernel, it is
obvious that the selection of kernel functions is insensitive with MSE (Simonoff, 1996).
Therefore, kernel function should be selected based on other consideration, such as
properties of fˆ.
17
2.4. KERNEL DENSITY ESTIMATION (KDE)
Table 2.2: Various kernel functions and forms
Kernel Form
1
Uniform 2
Triangular k(u) = (1 − |u|)

2 2
k(u) = 15
¡ ¢
Biweight 16 1 − u
35
¢3
k(u) = 32 1 − u 2
¡
Triweight
2
/2
Gaussian k(u) = (2π)−1/2 e −u
k(u) = 34 1 − u 2
¡ ¢
Epanechnikov
(b) Triangular kernel

(a) Uniform kernel function (c) Biweight kernel function
function
(d) Triweight kernel (f) Epanechnikov kernel

(e) Gaussian kernel function
function function
Figure 2.9: Distribution of K (u) in various kernel functions
Choosing the optimal bandwidth
If the reference of density function f is based on the Gaussian function, then the Gaus-
sian density can be substituted into equation 2.27, and resulting
h 0 = 1.059σn −1/5 (2.29)
This Gaussian reference density can also be used and converted into other types
of kernel function. The optimal bandwidth of other kernel function, h 0,K ∗ satisfies the
condition,
h 0,K ∗ = c K ∗ h 0,G , (2.30)
where
18
" p #1/5
2 πR(K ∗)
cK ∗ = (2.31)
σ4K ∗
and h 0,G is the optimal bandwidth of Gaussian kernel. This method is often referred
as the Silverman Rule-of-Thumb of selecting optimal bandwidth. Depending on the
true density, this method can give the optimal bandwidth if the true density is normal.
However, if the true density is close to normal, the bandwidth will be close to optimal
(Hansen, 2009).
2.5 Classification
In a population which consists several independent groups, very often we wish to look
for the characteristics or features to separate the multivariate samples into the known
groups. Based on the features, classification rule could be developed in order to iden-
tify and allocate an object from new observations into one of the groups.
Consider a population consists two sub-populations, denoted as π1 and π2 . The
probability density of each population is denoted as f 1 (x) and f 2 (x), with random vari-
able of X = X 1 , . . . , X p . Denote that Ω is the collection of all possible outcomes x, R 1
¡ ¢
is the possible outcomes x which are classified as population π1 , and R 2 = Ω−R 1 is the
possible outcomes of x which are classified as population π2 .
The classification probabilities can be presented in the following table:
Table 2.3: Classification probabilities table
Classified as:
π1 π2
π1 P (1|1) P (2|1)
True population:
π2 P (1|2) P (2|2)
The probability misclassifying an object as a belonging to population π2 when the ob-

ject actually belong to population π1 calculated as
Z
P (2|1) = P (X ∈ R 2 |X ∈ π1 ) = f 1 (x)d x (2.32)
R2
and the probability misclassifying an object as a belonging to population π1 when the

actual belonging is population π2 equals
Z
P (1|2) = P (X ∈ R 1 |X ∈ π2 ) = f 2 (x)d x (2.33)
R1
In some cases, prior probability and costs of misclassification are taken account
into classification rules. Prior probability is the probability of one population from
prior observation and denoted as p 1 for prior probability of population π1 and p 2 for
prior probability of population π2 . The total of prior probability is equal to 1, p 1 + p 2 =
1. Costs of misclassification are defined as the prices to pay if an object is misclassified
19
2.5. CLASSIFICATION
Figure 2.10: Description of misclassification regions P (1|2) and P (2|1). The purple
shaded area indicates region of P (1|2) and the light green shaded area indicates region
of P (2|1)
and denoted as c 1 for the cost of classifying an object of π1 as π2 and c 2 for the cost of
classifying an object of π2 as π1 .
The classification rules are evaluated in terms of the expected cost of misclassifica-
tion (ECM)
EC M = c(2|1)P (2|1)p 1 + c(1|2)P (1|2)p 2 (2.34)
The optimal classification rule is calculated by minimizing the ECM, resulting
f 1 (x) c(1|2) p 2 o
n µ ¶µ ¶
R 1 = x ∈ Ω; ≥
f 2 (x) c(2|1) p 1
(2.35)
f 1 (x) c(1|2) p 2 o
n µ ¶µ ¶
R 2 = x ∈ Ω; <
f 2 (x) c(2|1) p 1
Special classification rules prevail for conditions such as:
1. Equal (or unknown) prior probabilities: p 1 = p 2 . The classification rule now de-
pends on probability density ratio and cost ratio.
f 1 (x) c(1|2) f 1 (x) c(1|2)
R1 : ≥ , R2 : < (2.36)
f 2 (x) c(2|1) f 2 (x) c(2|1)
2. Equal (or undefined) misclassification cost: c(1|2) = c(2|1). The classification

rule now depends on prior probability and density ratio.
f 1 (x) p(2) f 1 (x) p(2)
R1 : ≥ , R2 : < (2.37)
f 2 (x) p(1) f 2 (x) p(1)
3. Equal prior probabilities and equal misclassification cost: p 1 = p 2 , c(1|2) = c(2|1).

The classification rule now only depends on probability density ratio.
f 1 (x) f 1 (x)
R1 : ≥ 1, R2 : <1 (2.38)
f 2 (x) f 2 (x)
20
Chapter 3
Methodology and data
This chapter contains the overview of the methodology applied and the management
of data which would be used in this study. Within data management section, the data
source, the process of collecting data, and the data limitation are also discussed.
3.1 Methodology
This study was performed based on a set of systematic methods, which is referred as
methodology. The methodology applied was summarized in a flowchart, described in
Figure 3.1. The results from each process are presented and discussed further in this
report.
3.2 Data management

3.2.1 Data source
The data used in this study was obtained from iQx software built by AGR company. The
software provides well data management from approximately 6,000 wells in Norwegian
Continental Shelf (NCS) which are grouped into ± 1693 geological blocks. The software
records various types of well data, such as geological description, well schematic, sur-
face logging, and also well logging (or geophysical) data. However, not all of the wells
listed in this software have a complete set of data, which later became obstacles in our
study. The detailed explanation of the obstacles is discussed in the next section after
the data were extracted.
3.2.2 Filtering and collecting data

As explained in the previous section, there was a large number of wells recorded in iQx
software. In order to keep the simplicity of this study, we focused working on a small
number of wells which have similarity in the geological description. From the aspect
of data quality, we selected wells which had a complete set geophysical data to aid any
ambiguity in the result.
The filtering process to select the appropriate wells for the study was executed
manually due to limitations in software functionality. Therefore, the existence and the
21
3.2. DATA MANAGEMENT
Start
Collecting data
Data exploration
Hypothesis testing
Analyzing data using univari-

ate kernel density estimation
Validation for lithology clas-

sification and prediction
Stop
Figure 3.1: Flowchart of methodology in this study
quality of geophysical data from various wells were checked and recorded in a spread-
sheet. Considering the large number of wells available and time limitation, the evalu-
ation within this process was limited to 16 geological blocks.
From the evaluation, we chose 3 neighboring wells from Block 15, Well 15/5-7
A,Well 15/6-11 S, and Well 15/6-9 S. Apart from the availability and the good quality
of logging data, these wells also had similar geological features. The evaluation also
discovered that Block 15 has a large resource of wells which would be beneficial for
validation process later in this study.
Once the filtering process was completed and the wells were chosen, we started ex-
tracting the data which related to the study, they were geophysical data, geological de-
scription and well schematic. The well schematic data provided information regarding
casing size, hole size, and shoe depth. While the geological description provided infor-
mation regarding formation distribution, characteristic, basal stereotype, depositional
environment, and dominant lithology.
According to the geological description, Block 15 consisted ± 35 formations in total
which were divided into 6 big formation groups. There were 4 major lithologies found
in these formations, they were sandstone, shale, chalk, and carbonate lithology. How-
ever, the available geological description only provided one generalized description
of formation for all of the selected wells and also lacked of detailed information (i.e.
lithology mixture, minerals). The geological description of the formations is shown in
22
CHAPTER 3. METHODOLOGY AND DATA
Figure 3.2: The spreadsheet for recording the availability of logging data in some of
the wells
Appendix B.
After investigated the data attributes, we discovered that the geophysical data were
only recorded based on the measured depth (MD), without the availability of true ver-
tical depth (TVD)- and time-based geophysical data. The other data which were not
available from the software were the information of the run wireline tools and well tra-
jectory data. Such these data constraints would lead to limitations in the study analyses
and results.
To begin with, we visualized all the extracted data into one figure of interface which
constructed by using MATLAB, shown in Fig. 3.3. Eventually, we set some assumptions
in this study:
1. The lithology information from the geological description represented the real
lithology condition.
2. All the geophysical data were recorded from well-calibrated logging tools, thus
the geophysical data were valid and represented real borehole condition.
23
3.2. DATA MANAGEMENT
Figure 3.3: The interface for data visualization of Well 15/5-7 A, Well 15/6-11 S, and
Well 15/6-9 S. In each well, the geophysical data were plotted in log traces, alongside
casing data and geological description.
24
Chapter 4
Data exploration on GR data
This chapter contains the process of data exploration of GR data of Well 15/5-7 A, Well
15/6-11 S, and Well 15/6-9 Sand the results. This process was performed to check the
quality of GR data and discover any variables grouping the GR data. From this process,
we expected that the GR data would be valid to be used for further analysis. After the
data exploration was completed, hypothesis testings were carried out in order to test
and support the discovery from data exploration.
4.1 Data exploration results and discussions

Within the process of data exploration, we plotted the graphical description and calcu-
lated the numerical description of GR data which later would be used for observations.
The graphical description included the data visualization by using histogram and box-
plot while the numerical description included the measurement of data variability by
calculating the mean, median, standard deviation, and IQR of the data. The bin width
of the histogram was chosen from Scott rule (Chapter 2.2.1) while the boxplot visual-
ization was adapted from Tukey method (Chapter 2.2.2).
4.1.1 GR data exploration under lithology grouping

In the theory of GR explained in Chapter 2.1.1, GR values are dependent on the types of
lithology that present in the borehole and tends to have a similar value for one lithology
group. According to this, the GR data description would be presented and observed
based on lithology type. The graphical descriptions of GR data in Well 15/5-7 A,Well
15/6-11 S, and Well 15/6-9 S are presented in Fig. 4.1 - 4.3 with statistic description
summarized in Table 4.1 - 4.3.
In Well 15/5-7 A, there were three different lithology types indicated, they were
shale, sandstone, and chalk, while Well 15/6-11 S and Well 15/6-9 S had additional
carbonate lithology. From the observation of histogram plots, GR of shale and sand-
stone from all of these wells had bimodal distribution which indicated by two major
peaks presented in the histogram, while chalk and carbonate had uniform distribu-
tion. The observed bimodal distribution might indicate that the GR data of shale and
sandstone lithology contained sub-groups.
25
4.1. DATA EXPLORATION RESULTS AND DISCUSSIONS
Table 4.1: Statistic description of GR data in Well 15/5-7 A grouped according to the
present lithology
Lithology Mean Median Standard IQR

Deviation
Chalk 57.39 55.18 22.67 21.20
Sandstone 94.71 115.66 41.63 73.45
Shale 118.89 132.88 45.78 68.69
Table 4.2: Statistic description of GR data in Well 15/6-11 S grouped according to

present lithology

Deviation
Carbonate 50.39 50.21 5.67 9.82
Chalk 35.66 35.62 10.42 14.01
Sandstone 91.66 100.05 31.33 55.28
Shale 97.47 92.31 33.85 52.89
Table 4.3: Statistic description of GR data in Well 15/6-9 S grouped according to

present lithology

Deviation
Carbonate 43.68 40.48 14.98 21.14
Chalk 45.06 45.66 10.05 13.25
Sandstone 102.58 104.32 18.39 19.12
Shale 116.99 122.42 21.58 22.88
26
CHAPTER 4. DATA EXPLORATION ON GR DATA
(a) Histograms
(b) Boxplots
Figure 4.1: Graphical description of GR data in Well 15/5-7 A which grouped based on
the lithology type
27
(a) Histograms
(b) Boxplots
Figure 4.2: Graphical description of GR data in Well 15/6-11 S which grouped based
on the lithology type
28
(a) Histograms
(b) Boxplots
Figure 4.3: Graphical description of GR data in Well 15/6-9 S which grouped based on
the lithology type
29
By looking into the GR distribution from histograms and boxplots, low GR values
were observed in chalk and carbonate lithology while the GR values of shale and sand-
stone were varying. It was also discovered that lithology is a variable that divide the GR
data into groups and each group appeared to be independent with each other. This
presumption tested and proved later in the hypothesis test Chapter 4.2.1.
High variance of GR value in sandstone and shale lithology was detected from the
standard deviation (σ) value, while carbonate and chalk had less variance. The widest
span of GR values was detected in shale lithology with many outliers and long whiskers
indicated from the boxplots. According to the discoveries, the GR data of shale and
sandstone were not convincing due to indication of bimodal distribution and high
variance of the data. We sensed that there were sub-groups might exist within each
lithology group. Hence, another investigation was performed and the detailed expla-
nation is provided in the next section.
4.1.2 GR data exploration under lithology grouping and hole size

subgroup
GR reading is affected by factors from borehole environment. Therefore, the quality
of GR data depends on the correction was made or not. The error factors are: (i) tool
eccentricity, (ii) hole size, (iii) mud weight, (iv) casing size, and (v) cement thickness.
Referring to this theory, we presumed that the sub-groups within lithology group were
emerged due to uncorrected GR data.
This presumption followed by an investigation to discover the variable of the sub-
group which was performed by visualizing the GR data in log traces. By observing the
log, we discovered that the values of GR in one hole size appeared to be shifted from
GR values in the other hole size (see Fig. 4.4). The investigation was improved by con-
structing statistical graphs and calculating the statistic descriptions of GR data which
grouped based on the lithology type and the hole size. The results are shown in Fig. 4.5
- 4.7 and Table 4.4 - 4.6.
By observing the minimum and maximum GR value of groups within shale and
sandstone lithology, it was clearly seen that the span of GR value in each group was
different from others and the variance of GR value was reduced. As an example, the GR
value of sandstone lithology in Well 15/5-7 A without hole size grouping was ranged
12.47-155.91 API. But, by hole size grouping, now we discovered that the GR value of
sandstone between 26" and 17 12 " had a huge gap. Group 26" had fairly small GR value
(13.64-50.57 API) compared to group 17 12 " (93.94-155.91 API)
Another approach proving that GR data distribution depends on the hole size was
by observing the median of groups within one lithology type. Such an example, the
median value of shale lithology in Well 15/6-11 S|: from group of 26" and 12 41 " (70.89
API and 89.27 API, respectively) were rather small compared to median observed from
group of 17 12 " and 8 21 " (125.39 API and 101.08 API, respectively).
Based on the histogram plots, shale and sandstone lithology showed better GR dis-
tribution with indication of unimodal distribution. In some cases, the hole size group-
ing resulted in symmetric distribution, such as groups of shale lithology with hole
size 17 12 " from Well 15/5-7 A and from Well 15/6-9 S. Such this symmetric distribu-
tion had mean (x̄) and median (M d ) values which were relatively closed (x̄ = 142.69,
M d = 143.58 for shale 17 21 " Well 15/5-7 A, and x̄ = 126.69, M d = 126.68 for shale 17 21 "
Well 15/6-9 S).
30
(a) (b)
Figure 4.4: Shifted GR value from logging visualization: (a) 26" (blue area) and 17 12 "
(red area) in Well 15/5-7 A and b) 17 21 " (blue area) and 12 14 " (red area) in Well
15/6-11 S
31
Table 4.4: Statistic description of Well 15/5-7 A grouped according to the present
lithology and subgroup of hole size
Well 15/5-7 A
Lithology Hole Mean Median St. Dev IQR Min Max

Size
26" 69.02 74.42 17.48 23.13 16.28 105.38

Shale 17 21 " 142.69 143.58 12.74 12.95 50.53 175.64
8 12 " 169.02 162.51 48.54 75.60 75.47 278.76
26" 24.97 24.20 5.76 6.39 13.64 50.57

Sandstone 17 21 " 122.47 122.39 10.72 14.32 93.94 155.91
8 12 " 56.31 53.10 20.48 32.29 12.47 104.66
26" - - - - - -
Chalk 17 21 " - - - - - -
8 12 " 57.39 55.18 22.67 21.20 12.16 184.99
32
Table 4.5: Statistic description of Well 15/6-11 S grouped according to the present
Well 15/6-11 S

Size
26" 69.13 70.89 13.76 14.74 11.94 97.09

17 12 " 125.90 125.39 15.51 18.70 76.04 175.73
Shale
12 14 " 97.83 89.27 39.83 54.68 40.75 330.87
8 12 " 98.17 101.18 29.33 33.69 36.04 269.62
26" - - - - - -
17 12 " 110.16 110.82 16.02 22.78 77.37 162.63
Sandstone
12 14 " 50.50 46.54 9.67 13.79 35.75 84.67
8 12 " 50.52 42.33 21.77 39.23 14.63 109.02
26" - - - - - -
17 1/2" - - - - - -
Chalk
12 14 " 35.66 35.62 10.42 14.01 12.99 72.67
8 12 " - - - - - -
26" - - - - - -
17 12 " - - - - - -
Carbonate
12 14 " 50.39 50.21 5.67 9.82 39.96 60.69
8 12 " - - - - - -
33
Table 4.6: Statistic description of Well 15/6-9 S grouped according to the present
Well 15/6-9 S

Size
24" 94.55 98.57 21.39 14.66 20.84 133.09

Shale 17 21 " 126.69 126.68 8.78 10.79 85.16 156.34
8 12 " 113.29 116.03 29.06 44.30 40.05 170.61
24" - - - - - -
Sandstone 17 21 " 107.84 108.49 11.73 16.54 85.31 149.65
8 12 " 68.71 70.03 17.47 32.09 36.41 109.81
24" - - - - - -
Chalk 17 21 " - - - - - -
8 12 " 45.06 45.66 10.05 13.25 23.09 82.65
24" - - - - - -
Carbonate 17 21 " - - - - - -
8 12 " 43.68 40.48 14.98 21.14 22.16 124.79
34
(a) Histograms
(b) Boxplots
Figure 4.5: Graphical description of GR data in Well 15/5-7 A which grouped

according to the lithology type and hole size
35
(a) Histograms
(b) Boxplots
Figure 4.6: Graphical description of GR data in Well 15/6-11 S which grouped

36
(a) Histograms
(b) Boxplots
Figure 4.7: Graphical description of GR data in Well 15/6-9 S which grouped

37
4.2. HYPOTHESIS TESTING RESULTS AND DISCUSSIONS
Summing up the investigation, hole size grouping in each lithology group had im-
proved the GR data distribution significantly and it was proved that our presumption
was correct. The results also indicated that each groups are independent with others.
To prove these discoveries and the presumptions, we performed another hypothesis
testing in Chapter 4.2.2.
Grouping the GR data based on the hole size can also remove other error factors,
such as mud weight and casing size, because in the application usually the value of
mud weight and casing size are constant during drilling one hole section. We contem-
plated on correcting the data based on the borehole environment, but the correction
was not possible because we were constrained with the availability of GR tool descrip-
tion from the data source.
4.2 Hypothesis testing results and discussions

4.2.1 Hypothesis testing #1
The first hypothesis testing was directed to check and test our presumption that GR
data contained an independent variable, which was lithology, from data exploration in
Chapter 4.1.1. The stated hypothesis are following:
H0 : the GR data had similar distribution and identical.
H1 : lithology type affected GR data value and GR data behavior in each lithology was
independent with the others.
The Kruskal-Wallis test was executed by using sample data from GR data and
grouping variable of the lithology type. The level of significance for the test was set to
5%. The summary of p-value and mean rank is presented in Table 4.7 and the detailed
result of ANOVA table is provided in Appendix C.
Table 4.7: P-value of the first hypothesis test for indicating any lithology group in GR
data. The detailed lithology group investigated in each well is provided in Table 4.8
Well P-Value
Well 15/5-7 A < 1.00 x 10-323

Well 15/6-11 S < 1.00 x 10-323
Well 15/6-9 S < 1.00 x 10-323
The lithology effect on GR data has been evaluated by using Kruskal-Wallis H test.
According to p-value given in each well, the null hypothesis was rejected due to p-
value<0.05. As stated in the theory, the Kruskal-Wallis test is an omnibus test, thus
we could not indicate which specific groups are statistically and significantly different
with the others. However, observation of the mean rank value in each group can give a
picture how each group differs from the others.
If a group has a mean rank value which relatively closed with the value from other
groups, then these groups are considered identical. In Well 15/5-7 A, the mean rank
38
Table 4.8: Mean ranks of each lithology group in Well 15/5-7 A, Well 15/6-11 S, and
Well 15/6-9 S from the first hypothesis test
Mean Group Rank

Well
Shale Sandstone Chalk Carbonate
Well 15/5-7 A 5.129e+03 3.8037e+03 2.0172e+03 -

Well 15/6-11 S 4.8042e+03 4.4590e+03 1.0064e+03 2.1327e+03
Well 15/6-9 S 4.8096e+03 3.3632e+03 0.8340e+03 0.7671e+03
values of shale, sandstone, and chalk were significantly different. Meanwhile, sand-
stone and shale lithology in Well 15/6-11 S had quite similar mean rank (4804 and
4459, respectively). However at this state, we could not conclude that shale and sand-
stone lithology were identical because it was proved that there were subgroups of hole
size within shale and sandstone lithology (see Chapter 4.1.2).
4.2.2 Hypothesis testing #2

The second hypothesis testing was conducted in order to test our presumption from
Chapter 4.1.2 that the GR data not only contained independent variable of lithology
but also hole size. The stated hypotheses are following:
H0 : the GR data in lithology groups had similar and identical distribution.
H1 : GR data in lithology groups were divided into another independent variable

which was borehole size.
The level of significance was set to 5%. The results of p-value are summarized in Ta-
ble 4.9 and the ANOVA table is provided in Appendix C.
Referring to the results of p-value in Table 4.9, all of the groups had p-value<0.05,
thus the null hypothesis was rejected. this also proved that there were at least 2 sub-
groups existed in each lithology group. In order to acknowledge any similarity between
the groups, we observed the mean rank value.
Observing groups in Well 15/5-7 A, we discovered that the shale lithology in 17 12 "
and 8 12 " group had similar mean rank values, respectively 2313 and 2556. However,
a contrast of the mean rank values between these groups was identified in sandstone
lithology. This observation showed that shale from group 17 12 " and 8 21 " were identical
but not for the sandstone lithology.
Another case from Well 15/6-11 S, both sandstone and shale had similar mean rank
values which identified in hole 12 14 " and 8 12 " group. Thus, the GR data from these
groups were identical for both sandstone and shale lithology and might be originated
from the same population. However, we took a careful approach and did not merge the
GR data of these groups in order to avoid errors.
In previous hypothesis testing, the observation of mean rank values of lithology
groups were limited because indication of hole size sub-group. Now, after grouping GR
data according to the hole size, we observed that lithologies in a group of hole size had
a diverse mean rank value. A significant variation of the mean rank of each lithology
39
4.2. HYPOTHESIS TESTING RESULTS AND DISCUSSIONS
was observed in the group of hole size 8 12 " in Well 15/6-11 S with mean rank of shale =
1545, sandstone = 455, chalk = 795, carbonate = 75. This observation proved that each
lithology group was independent and not identical with others.
Based on the observations of the results of hypotheses testing, the subsequent anal-
ysis on GR data would follow the discoveries from hypotheses test. Henceforth, the
analysis would be performed and presented separately according to the group of lithol-
ogy type and hole size.
Table 4.9: P-value result of the second hypothesis test. Empty (-) p-value results
indicated that there are no hole size group found in the particular lithology group.
Well Lithology P-Value
Shale < 1.00 x 10-323

Well 15/5-7 A Sandstone < 1.00 x 10-323
Chalk -
Shale < 1.00 x 10-323

Sandstone < 1.00 x 10-323
Well 15/6-11 S
Chalk -
Carbonate -
Shale < 1.00 x 10-323

Sandstone 1.061 x 10-132
Well 15/6-11 S
Chalk -
Carbonate -
40
Table 4.10: Mean rank for each category. Empty (-) mean rank results indicated that
there are no mean rank for selected hole size and lithology group.
Lithology
Well Hole Size
Shale Sandstone Chalk Carbonate
26" 652 224 - -

Well 15/5-7 A 17 12 " 2313 1598 - -
8 12 " 2556 584 1010 -
26" 784 - - -
17 12 " 2435 1886 - -
Well 15/6-11 S
12 14 " 1545 455 795 47
8 12 " 1570 417 - -
26" 757 - - -
Well 15/6-11 S 17 12 " 2572 1043 - -
8 12 " 1883 152 570 145
41
4.3. CONCLUDING REMARKS
4.3 Concluding remarks

• Data exploration on GR data has been performed. The observation of the results
showed that there were at least two group types within the GR data distribution,
they were lithology type and borehole size.
• Hypothesis tests were conducted by using Kruskal-Wallis H test to support the

discoveries from data exploration.
• Two hypothesis tests were run and the p-value from the results had shown that
GR data was dependent to the lithology type and the borehole size.
42
Chapter 5
Univariate KDE analysis on GR data
This chapter introduces the application of KDE on GR data to get the estimation of
probability density. Later on, a validation process was performed to assess the effec-
tiveness of GR variable on classifying lithology by using the results acquired from KDE
application. The results and discussions from validation process are provided in a sep-
arated section in this chapter.
5.1 KDE analysis on GR data

The KDE application was performed in order to get a smooth probability density esti-
mation of non-parametric distribution of GR data. The kernel estimation was applied
for wells: Well 15/5-7 A, Well 15/6-11 S, and Well 15/6-9 S to get the probability density
based on Equation 2.24.
1 X n ³x −x ´
i
fˆ(x) = K , (2.24 revisited)
nh i =1 h
The kernel function applied for this analysis was Epanechnikov function by con-
sidering that this function minimizes the AMISE value (see Chapter 2.4.2). However,
it was stated that the choice of kernel function does not affect the result significantly
(Silverman, 1981).
The estimation was carried out by using MATLAB R2015A built-in function,
ksdensity, which returns the estimation of the probability density and the sample
vector (Deveaux, 1999). The optimal bandwidth of kernel smoothing, h, was also cal-
culated and selected by using this function automatically. The default of the optimal
bandwidth from this function is based on a distribution of normal densities. The val-
ues of the optimal bandwidth for each data group are provided in the result section
together with the results of probability density plots.
43
5.1. KDE ANALYSIS ON GR DATA
5.1.1 Results
The results of probability density plot for each well are provided based on hole size.
The lithology groups which presented in each hole size group were plotted together in
one axis with different line colors (indicated in the legend).
Well 15/5-7 A
(a)
(b)
(c)
44
CHAPTER 5. UNIVARIATE KDE ANALYSIS ON GR DATA
(d)
Figure 5.1: Density probability plots with KDE and kernel bandwidth results of Well
15/5-7 A for present lithology in borehole size: (a)36", (b)26", (c)17 21 ", and (d)8 12 "
Well 15/6-11 S
(a)
(b)
45
(c)
(d)
(e)
15/6-11 S for present lithology in borehole size: (a)36", (b)26", (c)17 12 ", (d)12 41 ", and
(e)8 21 "
46
Well 15/6-9 S
(a)
(b)
(c)
15/6-9 S for present lithology in borehole size: (a)24", (b)17 12 ", and (c)8 21 "
5.1.2 Discussions
There are two discussion points presented within this discussion section. The first dis-
cussion point discusses the results from histogram and kernel estimator. While, the
second discussion point discussed about the GR data distribution for each lithology.
Simple and smooth probability density estimator
The plots of probability density estimated by kernel estimator in the result section
showed smoothed probability density lines along GR values. In addition, we also dis-
covered that the kernel estimator was able to match the non-parametric distribution
47
of GR data. To assess the effectiveness of kernel estimator over histogram in estimating

probability density, we plotted data of several groups using kernel estimator and his-
togram in one figure, shown in Fig. 5.4. In this comparison, the histogram was plotted
by using the same bin width method as in Chapter 4.1, which was Sturges method.
In figure 5.4a, both histogram and kernel estimator returned similar estimation of
the probability density and were able to approach the non-parametric distribution of
GR data. In another investigation of figure 5.4b, we observed a bimodal distribution of
GR in sandstone lithology with two adjacent modes. The histogram returned poor es-
timation of probability density around these modes which was indicated by the abrupt
change of probability density values on GR values between these modes. Meanwhile,
kernel estimator returned a better estimation which indicated by the gradual changes
between these modes. The better estimation from kernel estimator was due to the
ability of this estimator to calculate the superimpose effect of GR values in the local
neighborhood around these modes. Fig. 5.4c showed that the kernel estimator was
also able to approach skewed distribution.
(a)
(b)
(c)
Figure 5.4: Comparison of histogram and KDE in estimating probability density in

(a)Well 15/5-7 A , section 26", (b)Well 15/6-9 S , section 17 12 ", and (c) section 24"
48
As stated in theory, kernel estimator depends on the smoothing parameter, h, just

as histogram depends on the bin width, thus, the probability density distribution from
histogram and kernel depends on how these parameters are chosen. However, by ob-
serving the results in general, it was clearly seen that the visualization of probability
density from kernel estimator was much better compared to the discreteness of the
histogram.
Despite of that, histogram has another weakness comparing distribution of several
group samples. The histogram of each sample could not be plotted and merged to-
gether in one figure unless the samples have a uniform bandwidth. Because histogram
plot contains discrete bins, the varying bandwidths of samples in one figure can cause
a subjective comparison between one sample to others. This weakness of histogram
can be avoided by using the kernel estimator because the probability density returned
by kernel estimator is continuous.
Data distribution of GR from probability distribution
Observing the probability density of groups which contain two or more than two
lithologies, there were overlapped distributions of lithologies which forming region(s).
This region indicated that there were multiple lithologies exist for any GR value located
within it. Moreover, the size of the overlapped regions affects the misclassification rate.
The bigger the size of this region, more misclassification will likely occur.
Figure 5.5: Probability density of Well 15/5-7 A for hole size 17 21 ". The gray area is the
overlapping distribution between shale and sandstone lithology.
Group of 17 12 " from Well 15/5-7 A was dominated with high GR value of shale
and low GR value of sandstone. The separation between these lithologies was clear
if we observed the peak of probability density of each lithology. However, there was an
overlapped region formed because of the left-skewed distribution from shale lithology.
Similar cases also found in other groups which contain shale and sandstone lithology.
In another case, several overlapped regions were formed within groups which con-
tain additional chalk and/or carbonate lithology. This typical case was identified in
49
group of 8 12 " from Well 15/5-7 A and Well 15/6-9 S and 12 14 " from Well 15/6-11 S.
These multiple overlapped regions were formed due to low GR value of chalk, carbon-
ate, and sandstone lithology and the separation of these lithologies was difficult to be
identified. This discovery was consistent with the discussion from GR data exploration
in Chapter 4.
Based on this, chalk, carbonate, and sandstone lithology were grouped into one
group, which was non-shale lithology group, because these lithologies had identical
value of GR. This decision was also made to reduce the complexity of classification
of these lithologies. Henceforth, the lithology classification was simplified into shale
and non-shale lithology group. This type of classification was in accordance with the
usage of GR data which is shale identification. The results of probability density plots
categorized into shale and non-shale lithology are shown in Appendix D.
(a) Probability density without merging sandstone, chalk, and carbonate
(b) Probability density with merging sandstone, chalk, and carbonate into non-shale lithology
group
Figure 5.6: Comparison of probability density from group of 8 21 " in Well 15/6-9 S
which are (a) not-merged and (b) merged
A significant change of probability density distribution was observed in groups

which contain more than two lithologies (group of 8 12 " from Well 15/5-7 A and Well
15/6-9 Sand 12 14 "). Meanwhile, there was no significant change indicated in the groups
which only contain shale and sandstone lithologies. Fig. 5.6 showed the alteration by
merging chalk, carbonate, and sandstone lithologies into a non-shale group. The over-
lapped regions in Fig. 5.6a were reduced and simplified into one overlapped region in
figure Fig. 5.6b due to the skewed distribution of non-shale lithology group.
So, what is the explanation of the overlapped lithology distribution in term of geo-
logical aspect? Based on the theory of GR, the GR tool works by measuring the emitted
radioactivity of rock in the borehole. Even though shale appertains as a lithology with
50
high radioactivity, there are some cases when high radioactivity is failed to be indi-
cated as shale, such as in the shaly sandstone formation, uranium bearing formation,
high radioactivity sandstone formation. These failed cases often cause misinterpreta-
tion which leads to lithology misclassification. This condition explained why the over-
lapped regions formed between shale and non-shale lithologies. There are methods
that have been proposed to improve the lithology classification in cases mentioned
above, such as the usage of P e measurement or quantitative analysis to indicate min-
erals in shaly sandstone (Poupon and Gaymard, 1970), uranium measurement from
spectral GR logging to indicate uranium-rich formation, and potassium reading to in-
dicate sandstone with mica formation (Ellis and Singer, 2010). However, the investiga-
tion on these cases was beyond the scope of this study.
5.2 Validation of probability density on GR data by using

KDE
From the results and discussions in the previous section, kernel estimator was proved
to have the capability to estimate probability density of GR data which has non-
parametric distribution. Based on this discovery, we performed a validation into the
models which were constructed from GR data by using kernel estimator. The motiva-
tion of validation was to evaluate the capability of GR variable to discriminate litholo-
gies by testing unknown dataset into the model and investigating how the model will
generalize to the testing dataset. The goal from this process was to estimate the accu-
racy of the GR model in practice (lithology classification and prediction for this case).
Prior to the validation process, the model must be constructed from a known
dataset (or referred as training dataset). In addition to the model, the dataset to be
tested into the model (or referred as testing dataset) which was taken from unknown
data also should be prepared. Once the models and testing data were prepared, the val-
idation was started by classifying elements in the testing data toward the model based
on the classification rule.
Within the classifying process, the number of observations of elements in testing
data which correctly classified and misclassified was noted in the confusion matrix.
The confusion matrix is a table that reports the number of false positives, false nega-
tives, true positives, and true positives (see Table 5.1). From the final result of confu-
sion matrix, the misclassification rate can be calculated following
FP + FN
Misclassification rate = , (5.1)
Total number of observation
where total number of observation = TN + FP + FN + TP.
From a brief explanation above, there are three main properties to be set within the
validation process: the models (testing dataset), the testing dataset, and the classifica-
tion rule.
The classification rule
Referring the theory of classification in Chapter 2.5, prior probability and cost of mis-
classification can be taken account in classification rule. In this study, we neglected
the misclassification cost, thus c(1|2) = c(2|1). However, we would investigate the
51
5.2. VALIDATION OF PROBABILITY DENSITY ON GR DATA BY USING KDE
Table 5.1: Confusion matrix table of 2 sub-population, π1 and π2
Predicted
π1 π2
π1 True Negative (TN) : False Positive (FP):

Number of observations correctly Number of observations
classified as π1 that belong to π1 incorrectly classified as π2 that
Actual
belong to π1
π2 False Negative (FN): True Positive (TP):
Number of observations Number of observations correctly
incorrectly classified as π1 that classified as π2 that belong to π2
belong to π2
effect of prior probability and compared the result with another result which neglected
the effect of prior probability. Therefore, there were two classification rules applied in
this study, they were rules from Eq. 2.38 and Eq. 2.37.
The models (training dataset)

The models selected in this study were originated from GR data of Well 15/5-7 A, Well
15/6-11 S, and Well 15/6-9 S. The models in each well were grouped based on the
hole size to get in accordance to results from data exploration. Moreover, we removed
groups which only contain one type of lithology from the models to avoid any error
since the purpose of this process was lithology classification.
According to the selected classification rules from the explanation above, the mod-
els were constructed into two different types:
1. Model 1, model without prior probability. This model was constructed by as-
suming that prior probabilities between two lithology groups were equal (p(1) =
p(2) = 0.5). This model would agree with the classification rule from Eq. 2.38.
2. Model 2, model with prior probability. This model was constructed by taking
account the effect of prior probabilities. The prior probability values were deter-
mined by calculating the number of observations of shale and non-shale lithol-
ogy from the geological description of testing dataset. The number of observa-
tions then standardized into 1 in order to fulfill the condition p(1)+p(2) = 1. This
model complied the classification rule from Eq. 2.37.
The testing dataset

In this study, the testing data were not randomly selected, instead the data were se-
lected on basis of depth span. This adjustment was based on the fact that rock in for-
mation is formed as a bed or layer, thus term of rock lithology related to group of rock
in particular depth range.
Based on the source of testing data, there were 3 different types of experiment
which performed in this study:
1. Experiment 1, a validation within groups in the same well. The source of testing
data for this experiment was generated from the same source of training data.
52
The ratio of training and testing dataset was set to ± 70 % /30 %, respectively. The
section which selected for training and testing data was not randomly selected,
but the selection was based on the consideration if we apply this experiment
in the field. In the practice, we would have a training dataset from formations
that have been drilled and with known lithology. Meanwhile, the testing dataset
was taken from the formations that recently been drilled with unknown lithology.
According to this, the testing dataset would always be taken from a section in a
greater depth than the depth of the section for testing data. Hence, the GR data
from the upper depth region was set as training data and the lower depth region
was set as testing data.
2. Experiment 2, a validation with neighboring wells within Block 15. The testing
data would be picked from neighboring wells in Block 15 and tested according to
the hole size. One additional well from Block 15, Well 15/6-12, was used in the
validation process for group of 12 14 " in Well 15/6-11 S.
3. Experiment 3, a validation with wells from different block, which is Block 16.
The wells used in this experiment were Well 16/1-14, Well 16/2-7, and Well 16/2-
13 A. Similar to experiment 2, the testing data would be tested according to the
hole size.
The validation process was performed by using a program which constructed by

using MATLAB. In addition to the program, we also built an interface using MATLAB
GUI toolbox to improve the speed of analysis and provide visualizations of the results.
This interface provides several functionalities, such as preview of probability plots for
training and testing data, misclassification error calculation, lithology prediction plot,
intersection points table, and confusion table. Moreover, the interface allows the user
to select preferred depth of training and testing data, and prior probability data source.
53
54
Figure 5.7: Preview of the interface for validation process built by using MATLAB
5.2.1 Results
Experiment 1
1. Well 15/5-7 A
Table 5.2: Summary of data and results for validation within each hole section in Well 15/5-7 A
Training data Testing data Prior probability Misclassification Rate (%)

Hole size
Depth (m) N Depth (m) N Shale Not-shale Model 1 Model 2

26" 194 - 850 1313 850 - 1038.5 378 0.254 0.746 2.12 11.38
17 12 " 1039 - 2180 2283 2180 - 2656.5 954 0.144 0.856 13 12.58
8 12 " 2657 - 3800 2287 3800 - 4119 639 0.541 0.459 11.27 9.86
55
2. Well 15/6-11 S
Table 5.3: Summary of data and results for validation within each hole section in Well 15/6-11 S

Hole size
17 12 " 690 - 1730 2081 1730 - 2181 903 0.158 0.842 12.62 15.84
12 14 " 2181.5 - 3320 2278 3320 - 3816.5 994 0.447 0.553 13.88 12.07
3. Well 15/6-9 S
Table 5.4: Summary of data and results for validation within each hole section in Well 15/6-9 S

Hole size
17 12 " 753 - 2180 2855 2180 - 2785.5 1212 0.368 0.632 43.48 38.28
8 12 " 2786 - 3590 1609 3590 - 3942 705 0.684 0.352 23.69 25.39
Experiment 2
1. Well 15/5-7 A as training data

56
Table 5.5: Summary of data and results for validation of Well 15/6-11 S and Well 15/6-9 S , by using Well 15/5-7 A as training data
Training data: Misclassification

Testing Data Prior probability
Hole Well 15/5-7 A Rate (%)
size Depth (m) N Well Depth (m) N Shale Not-shale Model 1 Model 2
26" 194 - 1038.5 1690 Well 15/6-11 S 187 - 689.5 1006 1 0 2.19 0
Well 15/6-11 S 690 - 2181 2983 0.334 0.666 34.43 34.56

17 12 " 1039 - 2656.5 3236
Well 15/6-9 S 753 - 2785.5 4066 0.607 0.393 52.26 48.55
Well 15/6-11 S 3817 - 4043 453 0.768 0.232 33.77 22.3

8 12 " 2657 - 4119 2925
Well 15/6-9 S 2786 - 3942 2313 0.276 0.724 8.13 11.59
2. Well 15/6-11 S as training data
Table 5.6: Summary of data and results for validation of Well 15/5-7 A and Well 15/6-9 S , by using Well 15/6-11 S as training data

Hole Well 15/6-11 S Rate (%)
26" 187 - 689.5 1006 Well 15/5-7 A 194 - 1038.5 1690 0.763 0.237 23.67 23.67

Well 15/5-7 A 1039 - 2656.5 3236 0.514 0.486 35.97 36.53
17 12 " 690 - 2181 2983
Well 15/6-9 S 753 - 2785.5 4066 0.607 0.393 15.1 15.15
12 14 " 2181.5 - 3816.5 3271 Well 15/6-12 2754 - 3628.5 1750 0.01 0.99 2.17 1.14
57
Well 15/5-7 A 2657 - 4119 2925 0.17 0.83 12.27 6.39

8 12 " 3817 - 4043 453
Well 15/6-9 S 2786 - 3942 2313 0.276 0.724 7.57 8
Table 5.7: Summary of data and results for validation of Well 15/5-7 A and Well 15/6-11 S , by using Well 15/6-9 S as training data

Well 15/5-7 A 1039 - 2656.5 3236 0.514 0.486 35.51 35.97

17 12 " 753 - 2785.5 4066
Well 15/6-11 S 690 - 2181 2983 0.334 0.666 29 28.03
Well 15/5-7 A 2657 - 4119 2925 0.17 0.83 26.39 13.64

8 12 " 2786 - 3942 2313
Well 15/6-11 S 3817 - 4043 453 0.768 0.232 18.98 14.79
58
Experiment 3
1. Well 15/5-7 A as training data
Table 5.8: Summary of data and results for validation of wells in Block 16 by using Well 15/5-7 A as training data

Well 16/1-14 371 - 1477.5 2214 0.642 0.358 41.00 38.44

17 12 "

1039 - 2656.5 3236
Well 16/2-7 700 - 1771.5 2144 0.628 0.372 62.78 62.64
Well 16/1-14 2228 - 2548 641 0.716 0.284 36.97 26.21

8 21 " 2657 - 4119 2925 Well 16/2-7 2098 - 2498 801 0.648 0.352 11.74 12.36
59
Well 16/2-13 A 2487 - 2772 571 0.595 0.405 30.65 30.30

Table 5.9: Summary of data and results for validation of wells in Block 16 by using Well 15/6-11 S as training data

Well 16/1-14 371 - 1477.5 2214 0.642 0.358 29.49 28.27

17 21 " 690 - 2181 2983
Well 16/2-7 700 - 1771.5 2144 0.628 0.372 50.98 45.24
Well 16/1-14 1478 - 2227.5 1500 0.703 0.297 12.87 14.00

12 41 " 2181.5 - 3816.5 3271 Well 16/2-7 1772 - 2097.5 652 0.189 0.811 55.83 47.55
Well 16/2-13 A 717 - 2486.5 3540 0.674 0.326 25.82 25.51
60
Well 16/1-14 2228 - 2548 641 0.716 0.284 19.34 22.93

8 12 " 3817 - 4043 453 Well 16/2-7 2098 - 2498 801 0.648 0.352 19.35 24.84
Well 16/2-13 A 2487 - 2772 571 0.595 0.405 22.42 29.95
Table 5.10: Summary of data and results for validation of wells in Block 16 by using Well 15/6-9 S as training data

Well 16/1-14 371 - 1477.5 2214 0.642 0.358 49.10 48.24

17 12 " 753 - 2785.5 4066
Well 16/2-7 700 - 1771.5 2144 0.628 0.372 80.88 77.66

Well 16/1-14 2228 - 2548 641 0.716 0.284 20.12 21.53
8 21 " 2786 - 3942 2313 Well 16/2-7 2098 - 2498 801 0.648 0.352 23.22 23.10
Well 16/2-13 A 2487 - 2772 571 0.595 0.405 26.80 28.72
61
5.2.2 Discussions
There are 3 main points discussed in this section, they are the results of misclassifi-
cation rate, the effect of prior probability, and the explanation of how the validation
method for lithology prediction can be applied in the field. Each point is discussed in
the separated section.
Misclassification error from the experiments
The misclassification rate in each experiment is summarized in Table 5.11. In gen-

eral, the table showed that the misclassification rate increased significantly if the well
sources between the models and the testing data were different. Testing the models
by using datasets from the neighboring wells in experiment 2 increased the averaged
misclassification rate up to 7 % for model 1 and 3 % for model 2 compared to the rate in
experiment 1. Meanwhile, testing the models by using datasets from wells in different
block increased the averaged misclassification rate up to 17% for model 1 and 15% for
model 2 compared to the results in experiment 1.
Table 5.11: Table of minimum, maximum, and averaged value of misclassification

error in each model and experiment
Minimum error(%) Maximum error(%) Averaged error(%)

Experiment
Model 1 Model 2 Model 1 Model 2 Model 1 Model2
Experiment 1 2.12 9.86 43.38 38.28 18.54 19.02

Experiment 2 7.57 6.39 52.26 48.56 24.74 21.93
Experiment 3 11.74 12.36 80.88 77.66 35.39 34.54
The increasing rates of misclassification in experiment 2 and 3 were due to the dif-
ference of GR data distribution between the model and the testing data. Several exam-
ples of experiment 1, 2, and 3 were taken from model of 17 12 " in Well 15/6-9 S and the
distribution between the model and testing dataset are presented in Fig. 5.8. A contrast
between the distribution of each dataset was improved from experiment 1 to experi-
ment 3. If we observed the peak of the probability density of each lithology, it could
be seen that the GR value from testing data in experiment 2 were shifted away from
the GR value of the model and created a gap of GR value. A greater gap of testing data
distribution was indicated in experiment 3.
Summing up the observations above, we believed that the increased of misclassifi-
cation rate was related to the uncertainty factors on GR data. And we were certain that
the further away the location of wells used as the testing data source, the uncertainty
was increased. We identified some of the possible source of uncertainty, they were
geological factors, borehole environment, and GR tool. The uncertainty from geologi-
cal factors includes the mineral content, lithology mixture, and depositional environ-
ments of the formations. However, this geological factor was believed to have a small
contribution to the classification in this study because the wells tested in experiment
2 and 3 located in the same sub-basin, Viking Graben, and GR values of one formation
are less likely to change in vertical direction.
62
(a) Experiment 1, with testing data from Well 15/6-9 S(2180 - 2785.5 m)
(b) Experiment 2, with testing data from Well 15/5-7 A
63
(c) Experiment 3, with testing data from Well 16/1-14in Block 16
Figure 5.8: Training and testing data distributions from different experiments for
group of 17 21 ". The testing data was from Well 15/6-9 S
Borehole environment and GR tools were considered as the most contributing fac-
tors to this misclassification. We dismissed the effect of hole size because the models
were tested according to the hole size. However, factors such as drilling fluid, the tool
position, hole cavity, cement type, and the tool size could contribute as uncertainty
factors, notably with the lack of correction from the source and data limitation.
All of these uncertainties rise through the imperfect knowledge. However, the effect
from these uncertainties could be minimized if we integrate data from other source.
One of the example is the usage of data from geological description in addition to GR
data for lithology classification. This example was applied in this study and the ap-
plication was presented by adding prior probability into the classification rule. The
detailed discussion of the results from this application is shown in the next section.
Effect of prior probability on misclassification error
Within this discussion point, we would discuss the result of validation from model 1
and 2, the way prior probability working in reducing misclassification rate, and condi-
tions that lead model 2 fails to reduce the misclassification rate. Referring to Table 5.11,
model 2 from experiment 2 and 3 produced less misclassification rate, with difference
around 3% and 1% respectively, compared to model 1. Meanwhile, model 2 from ex-
periment 1 produced error 0.5% higher compared to model 1.
Based on the number of cases, 4 out of 7 total cases (4/7) in experiment 1 showed
that model 2 yielded less misclassification rate compared to model 1. A similar event
also occurs in several cases from experiment 2 and 3, with the number of cases 7/13
and 11/18 respectively. This observation discovered that more than half of the cases in
each experiment proved that model 2 produced less misclassification rate.
64
The unique values of prior probability in model 2 indicate the true distribution of
lithologies of the training data. This was unmistakable because the prior probability
values were generated based on the number of observation of each lithology from the
geological description of training data. This was different from model 1 which assumed
that the lithologies have an equal distribution.
So, in what way that the true distribution affects the classification? By recalling and
modifying the equation from Eq. 2.37, we will get following equation
f 1 (x) p(1)
R1 : ≥1
f 2 (x) p(2)
(5.2)
f 1 (x) p(1)
R2 : <1
f 2 (x) p(2)
According to the equation above, the true distribution from the prior probability will
affect the value of probability density of the models for each lithology group. And as the
further effect, the data distribution of the model for each lithology would also change.
In order to get a better understanding of this explanation, we took an example of ex-
periment 1 from group of 17 21 " in Well 15/6-9 S.
In model 1, the values of prior probability of shale and not shale lithology were
equal, p(shale) = p(not-shale) = 0.5. By multiplying these values to the probability
density, the model would produce probability density plot shown in Fig. 5.9a. Refer-
ring to the plot and the classification rule, all GR values which fulfill the condition of
f 1 (x) p(1)
f 2 (x) p(2) ≥ 1 fall within the green area which indicates shale region, R(shale). Mean-
f (x) p(1)
while, all GR values which fulfill the condition of f 21 (x) p(2) < 1 fall within the blue area
which indicates not-shale region, R(not-shale).
In model 2, the values of prior probabilities were calculated and the results were
p(shale)=0.368 and p(not-shale)=0.632. By multiplying these values into the proba-
bility density of each lithology, we generated model 2 which shown in Fig. 5.9b. Be-
cause the value of p(shale) is smaller compared to the value of p(not-shale), the area
of R(shale) is also smaller than the area of R(not-shale). In addition, the distribution of
probability density of model 2, f(shale) and f(not-shale), resembled the distribution of
training data (Fig. 5.9c) much closer compared to model 1.
By observing the plots from both model 1 and 2, the red dashed line which is
formed at the intersection of f(shale) and f(not-shale) separated R(shale) and R(not-
shale) very clearly. According to this, we discovered that the x-value of this intersection
point acted as a discriminator between these two lithology groups. Meanwhile, the
y-value of this intersection point influenced the size of overlapped region.
In order to understand in what extent these values can reduce the misclassification
rate, we summarized the x- and y- values of the intersection points from the example
in Table 5.12. The table shows that model 2 approached the x-values of testing data
much closer than model 1. Meanwhile, the less y-value was given from model 2. From
this observation, we believed that the more identical the x-values between the model
and the training data, less misclassification rate would be yielded, especially by know-
ing that the x-value behave as discriminator between the lithology groups. Meanwhile,
the smaller the y-value, the smaller the overlapped region of the model thus the clas-
sification from the model would be more precise. Even though the difference of these
values between model 1 and 2 was rather small, the misclassification error reduced by
5%.
65
(a) Probability density of model 1
(b) Probability density of model 2
(c) Probability density of testing data
Figure 5.9: Plot of probability density for case in experiment 1: 17 21 " in Well 15/6-9 S
66
Table 5.12: The values of the intersection points of the models and the testing data
from experiment 1, group of 17 21 " in Well 15/6-9 S
Intersection points
Data
x-values (API) y-values
Model 1 112.8 0.0053

Model 2 114.3 0.0051
Testing data 120.01 0.0135
The classification by using model 2 did not always produce less misclassification
rate but instead, model 1 produced a better result. We sensed that the main reason of
this event was the difference of the shape of data distribution between the model and
the testing. This reason indeed is related to the reason causing high misclassification
rate from the previous discussion point. However, the impact of this reason for this
case was somewhat different. We observed that in some cases adding prior probability
increased the contrast of x-value between the model and the testing data which led
to misclassification rate increment. Two examples showing this event are shown in
Fig. 5.10 and Fig. 5.11.
Application of validation process in lithology prediction
The discussions of model validation above give insights of how GR data can be pro-
cessed for lithology classification. Within this discussion point, we also presented the
best way to apply the method in the practice.
The experiments results discovered that the classification by using training and
testing data which originated from the same source was the most optimum in min-
imizing the misclassification rate. According to this, we suggested that the training
data for this application will be generated from the drilled section which lithology type
of this section is acknowledged and confirmed by other interpretation methods.
However, this may come with limitation during drilling formation at the beginning
of a section because no data from drilled formation is available. Therefore, we recom-
mended the usage of data from the neighboring well as the training data during drilling
the top of a section. And in order to reduce the errors from uncertainty factors, it is re-
quired to select the neighboring well which fairly identic to the current drilled well.
Once the information from the drilled formation is considered adequate, the training
data can be switch and use data from the drilled formation.
Meanwhile, the testing data will be taken from the section that just been drilled.
The usage of prior probability was also recommended to improve the results and the
value of prior probability can be calculated from the prediction of geological descrip-
tion of the current well. Once the data is classified toward the training data, the clas-
sification results can be verified with the results from other interpretation method to
check any misclassification from the proposed method. In our vision, we sensed that
the application of this classification method in the field could improve the lithology
interpretation and optimize the drilling operation.
67
(a) Model 1
(b) Model 2
(c) Testing data
Figure 5.10: The f(shale) of the models is more skewed to the left compared to the
f(shale) of the testing data. Multiplying the probability density with prior probabilities
in model 2 shifted the x-value of intersection point to right side (52.73 GR API), further
away from the x-value in testing data (29.78 GR API)
68
(a) Model 1
(b) Model 2
(c) Testing data
Figure 5.11: The f(not-shale) in training data has bimodal distribution and f(shale) has
right-skewed distribution. Meanwhile, f(not-shale) of testing data has right-skewed
distribution and f(shale) is less skewed than f(shale) in training data. Adding prioir
probability shifted the intersection point to 92.3 GR API, further away from the
intersection in the testing data (68.3 GR API)
69
5.3. PRELIMINARY STUDY OF BIVARIATE ANALYSIS
5.3 Preliminary study of bivariate analysis

Bivariate analysis is a statistical analysis which involves two different variables. It is
usually performed to improve the univariate analysis and determine empirical rela-
tionship between the two variables. In this study, we carried out this bivariate analy-
sis as a preliminary exploration which still requires improvement. The motivation of
this analysis was to improve the lithology classification by adding another variable of
well logging, which was neutron log. In the previous chapter, the experiments results
showed that univariate analysis of GR was insufficient to classify lithology as indicated
with the overlapping distribution of the lithology groups. We selected neutron log as
the second variable due to the capability of neutron log in distinguishing lithology, as
explained in Chapter 2.1.2.
The analysis was performed by plotting GR and neutron data together in a scat-
ter plot, with GR as the x-axis and neutron as the y-axis. Adjacent to the scatter plot,
there are 2 plots of probability density which estimated by kernel estimator. The GR
probability density plot is located at the bottom side of scatter plot, while the neutron
probability density plot is located at the left side of the scatter plot. The data points in
scatter plot are marked according to its lithology type. The data for analysis was gener-
ated from wells in Block 16, Well 16/1-14, Well 16/2-7, and Well 16/2-13 A. The results
of the plots for these wells are shown in Appendix E.
Within this section, we would discuss the results of plots from two different groups:
8 2 " from Well 16/1-14 and 17 12 " from Well 16/2-7 (Fig. 5.12 and Fig. 5.13). Both fig-
1
ures show that the overlapping distribution of lithologies indicated from one variable
could be distinguished by another variable. In other words, the GR and neutron vari-
ables complemented each other. This is because GR and neutron tool have different
principal of measurement, thus each tool has different sensitivity in the particular rock
types. The results also showed better separation of lithologies and we expected that the
misclassification rate could be reduced.
70
Figure 5.12: The overlapped distribution of shale and sandstone lithology from GR
could be separated by neutron data. Shale is dominated with high neutron value,
while sandstone has small neutron value.
Figure 5.13: The overlapped region of chalk and carbonate lithologies from neutron
could be separated by using GR. Chalk has smaller GR value than to carbonate.
71
5.4. CONCLUDING REMARK
5.4 Concluding remark

• The kernel density estimator with Epanechnikov kernel function has been ap-
plied to approximate the probability density of GR data. It was shown that the
kernel estimator was able to return the non-parametric distribution of GR data.
Moreover, the probability density from kernel estimator were considered better
over the probability density from histogram.
• The validation was applied to assess the GR data in lithology classification. From
the investigation, the quality of classification depends on the distribution of
training and testing data.
• The prior probability values which extracted from geological description were
effective in reducing the misclassification rate.
• The classification method from this study can be applied for lithology interpre-
tation in the practice. The application was considered useful to reduce time of
lithology interpretation during drilling operation.
• The preliminary study of bivariate analysis proved that by using two variables
of GR and neutron, the lithology classification was improved. The most signifi-
cant result was indicated from the separation of overlapped distribution of one
variable.
72
Chapter 6
Conclusions
In this study, a quantitative analysis of GR data for lithology classification and predic-
tion has been performed. Prior to the main analysis of this study, data exploration and
kernel application on GR data were carried out. The results from these processes were
exploited for the main analysis which was validation as a part of an assessment of vari-
able GR for lithology classification.
The data exploration was executed by visualization of statistical graphs: histogram
and boxplot, and analysis of statistical descriptive. At the beginning of the process, GR
data were grouped based on the lithology type. As the results were investigated, the
discovery showed that GR data contain sub-groups which originated from variable of
hole size. This was indicated from the high variance of GR within the lithology groups
and the contrast of GR values between different hole sizes during visualization GR in
log trace. We sensed that the sub-groups were emerged due to uncorrected GR data.
In order to support the discoveries, we tested the hypotheses from data exploration
by using Kruskal-Wallis H test. The results of p-value and mean rank value from these
tests agreed with the discoveries from data exploration. The results also proved that
GR data depended on lithology and hole size variables. The correction on GR data was
not possible due to inadequate informations from the data source, thus the GR data for
further analysis remained to be grouped based on the hole size.
The kernel density estimator was applied in this study to approximate the proba-
bility density of GR data which further would be employed for lithology classification.
The kernel estimator was chosen because of the capability of kernel estimator to return
a continuous probability density and approach the non-parametric distribution of GR
data. The type of kernel function applied in this study is Epanechnikov function. The
results of probability density from kernel application were preferred over histogram
due to the discreteness of histogram. The results of the probability plots also showed
that the lithology group within GR data was more convenient to be grouped into shale
and not-shale lithology. This was due to sandstone, chalk, and carbonate lithologies
had similar GR value range which was smaller than the shale lithology. In addition to
that, error from misclassifying of these 3 lithologies could be prevented.
The lithology classification and validation were performed by using 2 different
models in 3 different experiments. The results showed that the misclassification rate
was reduced significantly in experiment 1 which model and testing data were taken
from the same data source. Meanwhile, model 2 which added prior probability value
within the classification rule had less misclassification rate compared to model 1 which
neglected the prior probability value.
73
Summing up the observations of the main analysis, the main cause of lithology
misclassification was due to the difference of data distribution between the model and
the testing data. The most significant difference in data distribution was indicated in
experiments which testing data was taken from different wells and/or different block.
We sensed that the difference in data distribution was caused by uncertainty factors.
However, the effect of uncertainty factors could be reduced or even dismissed if the
data quality could be improved.
Considering that data quality was beyond the scope of this study, we suggested to
perform the experiment by using a corrected data logging. We also suggested to carry
out a similar study which lithology was classified by using the relative value of GR log,
thus the necessity of correction could be neglected and GR can be analyzed without
grouping the data based on the hole size.
The application of the method proposed in this study has been explained. The ap-
plication would give opportunities to predict and classify the lithology from GR read-
ing. Since the prediction was still limited and only processed the reading from the tool
location inside the borehole, a further work to improve the prediction beyond the tool
was recommended. In addition, the prediction could be improved by using the pos-
terior probability calculation from Bayes formula. The prediction will be more precise
because this formula calculates the conditional probability which are useful for deci-
sion making (Ott and Longnecker, 2010).
The results from bivariate analysis proved that the lithology classification could be
improved by using 2 variables: GR and neutron logging. We recommended to continue
the study of bivariate analysis through implementation of bivariate kernel estimation
and validation. It was also possible to combine data from mud logging and well logging
within the bivariate analysis and study the the most optimum variables combination.
In addition to that, discriminate analysis was also suggested so that the classification
could be enhanced.
74
Appendix A
Acronyms
MWD Measurement while Drilling
LWD Logging while Drilling
WOB Weight on Bit
GR Gamma Ray
ROP Rate of Penetration
GNT Gamma Ray/Neutron Tool
SNP Sidewall Neutron Porosity Tool
CNL Compesanted Neutron Log
EDA Exploratory Data Analysis
KDE Kernel Density Estimation
IQR Interquartile Range
MSE Mean Squared Error
MISE Integrated Mean Squared Error
AMISE Assymptotic MISE
ECM Expected Cost of Misclassification
NCS Norwegian Continental Shelf
TVD True Vertical Depth
75
Appendix B
Geological Description
The geological description is summarized in a table and consists information of for-

mation characteristic, dominated lithology, and depositional environment. The geo-
logical description was provided in the iQx software.
Group Formation Lithology Description
Nordland Utsira The upper boundary of this formation contains clay-

stones, while the formation itself mostly contains
sandstone. The formation probably represents shal-
low marine shelf sandstones.
Skade The formation is dominated with sandstone and

the lower boundary is characterized with decreas-
ing gamma-ray into claystones overlying Hordaland
Group. The formation was deposited in open marine
Hordaland environment.
Grid The majority lithology comprised in this formation

is sandstone. The upper boundary usually contains
claystones from Hordaland Group.
Balder Laminated shales is indicated at the upper part of

formation and decreasing to lower boundary sepa-
rating Sele formation, often with glauconitic over-
lying sediments. The formation was deposited in a
deep marine setting.
Sele The upper boundary separating Balder formation

Rogaland has an improved reading in GR and the lower bound-
ary has more sandy composition. The formation was
deposited in a deep marine setting.
Lista In areas where Lista formation is overlain by Sele for-

mation, there is no distinct changes indicated.
76
APPENDIX B. GEOLOGICAL DESCRIPTION
Heimdal The formation has majority of sandstone lithology.

the upper boundary is usually defined by mixture of
shale from Lista formation. The formation was
deposited as submarine fans derived from sand
accumulations on the shallow shelf. The turbidity
currents formed shale layers partly in Heimdal
formation.
Våle This formation is dominated with shale lithology

and was deposited in a marine environment.
Rogaland
Ty Ty formation is dominated with sandstone. The
lower boundary is indicated with decreasing gamma
ray into Shetland Group. The depositional
environment of this formation is deep marine fan
system.
Ekofisk This formation is dominated with chalk lithology.

The depositional environment is open marine with
deposition of calcareous debris flow and turbidities.
Tor The formation has layer thin in the Norwegian

sector and consists of chalk lithology. The
depositional environment is similar with Ekofisk
Shetland formation.
Hod The boundary between Hod and Blodøks formation

is often indicated with change in gamma ray
readings which increase toward Blodøks formation.
The depositional environment is open marine with
deposition of cyclic pelagic carbonates and distal
turbidities.
Blodøks The upper boundary is characterized with

decreasing gamma ray due to more chalk indicated
in the Hod formation. The formation was deposited
in anoxic conditions.
Hidra The upper boundary between Hidra and Blodøks

formation is characterized with lithology changing
from clhalk lithology to mudstone. The gamma ray
reading is decreasing towards Hidra formation. The
depositional environment of this formation is an
open marine with perioditic origin.
77
Rødby The upper boundary of Rødby formation and
Shetland Group has characteristic of decreasing
gamma ray reading toward Rødby formation. The
dominated lithology in this formation is chalk. The
Rødby formation has depositional environment of
an open marine with reddish sediments and
oxygenated environment.
Sola The lower boundary is indicated with increasing

gamma-ray reading from sandy sediments into
Cromer Knoll shaly Sola formation. The depositional environment
of this formation is a marine with alternating anoxic
and oxic conditions.
Åsgard This lower boundary of this formation is indicated

with increasing gamma-ray towars underlying
sediments which consists of rich claystones and
shales. The formation was deposited in an open
marine with low-energy shelf environment.
Draupne The formation has a clear break in upper boundary

with high response gamma ray. The formation was
deposited in a marine environment with restricted
circulation in the bottom and anaerobic condition.
Any sandstone indicated in the formation was
originated from turbiditic.
Viking
Heather The formation is dominated with silty claystone and

was deposited in an open marine environment.
Anomalously high gamma ray reading is indicated
in upper boundary.
Sleipner Sleipner formation is dominated with shaly rock.

The upper part of this formation marks with
transition into shales of the Viking Group or the
sandstones of Hugin formation. The formation was
Vestland deposited in a continental fluviodeltaic sequence.
Hugin This formation is dominated with sandstone

lithology and has clear log break in the upper
boundary with decreasing gamam ray reading. The
depositional environment of this formation is
shallow marine with occasional influence of
continental fluviodeltaic conditions.
78
APPENDIX B. GEOLOGICAL DESCRIPTION
No group Skagerrak The formation was probably deposited in a

defined prograding system of alluvial fans and dominated
with sandstone lithology.
Dunlin Cook The upper and lower boundary of this formation are
indicated with decreasing gamma ray reading. The
sandstone in this formation represents redeposited
sands from the edge of the shelf.
Statfjord Statfjord The top of formation is on the contact with Dunlin

Group and often consists of calcareous sandstones
and dark shales and siltstone. The formation was
deposited at lower alluvial plain and braided stream
deposits.
79
Appendix C
ANOVA Table Results of Hypotheses Test
The ANOVA table showed the summary of hypothesis testing. The calculated variables
are described below.
Source indicates the source of the variation in data. There are 3 sources provided,
Groups, Error and Total. Groups indicates the data variation due to the factor of
interest, or variation in the populations being compared. Error means the vari-
ation within each groups being compared, while Total means the total variation
in population data.
df means the degrees of freedom in the source. The formulas to calculate the df shown
as
d fGr oups = k − 1 (C.1a)

d f Er r or = N − k (C.1b)
d f Tot al = N − 1 (C.1c)
, where k is the number of groups in the source and N is the number of measure-
ments in the source.
SS means the sum of squares in the source. SSGr oups calculates the sum of squares
between treatment groups, which formula is shown in Equation 2.12. SS Er r or
indicates the sum of squares within groups which following
N X
X k ¡ ¢2
SS Er r or = yi j − M j (C.2)
i =1 j =1
, where y is the data point or measurement in data source, y = {y 1 , y 2 , . . . , y N }.

SS Tot al is equal to the sum of SSGr oups and SS Er r or .
MS is the mean square of each source and calculated following

SSGr oups
M SGr oups = (C.3a)
d fGr oups
SS Er r or
M S Er r or = (C.3b)
d f Er r or
Chi-sq is the distribution which approximated by H test value as shown in Equa-

tion 2.14 - Equation 2.16.
80
APPENDIX C. ANOVA TABLE RESULTS OF HYPOTHESES TEST
Table C.1: ANOVA table of hypothesis testing #1 in Well 15/5-7 A
Source SS df MS Chi-sq Prob>Chi-sq

Groups 1.2382e+10 2 6.1912e+09 2.4091e+03 0
Error 2.7975e+10 7850 3.5637e+06
Total 4.0358e+10 7852
Table C.2: ANOVA table of hypothesis testing #1 in Well 15/6-11 S

Groups 1.7061e+10 3 5.6869e+09 3.4302e+03 0
Error 2.1356e+10 7721 2.7659e+06
Total 3.8416e+10 7724
Table C.3: ANOVA table of hypothesis testing #1 in Well 15/6-9 S

Groups 1.7038e+10 3 5.6793e+09 3.7934e+03 0
Error 1.5930e+10 7337 2.1711e+06
Total 3.2967e+10 7340
Table C.4: ANOVA table of hypothesis testing #2 for shale lithology in Well 15/5-7 A

Groups 2.406e+09 2 1.203e+09 2.423e+03 0
Error 1.019e+09 3448 2.957e+05
Total 3.425e+09 3450
Table C.5: ANOVA table of hypothesis testing #2 for sandstone lithology in Well 15/5-7
A

Groups 7.857e+08 2 3.929e+08 1.658e+03 0
Error 3.434e+08 2381 1.442e+05
Total 1.129e+09 2383
Table C.6: ANOVA table of hypothesis testing #2 for shale lithology in Well 15/6-11 S

Groups 1.373e+09 3 4.577e+08 1.644e+03 0
Error 1.269e+09 3161 4.014e+05
Total 2.642e+09 3164
81
Table C.7: ANOVA table of hypothesis testing #2 for sandstone lithology in Well
15/6-11 S

Groups 1.269e+09 2 6.346e+08 1.836e+03 0
Error 7.214e+08 2877 2.508e+05
Total 2879
Table C.8: ANOVA table of hypothesis testing #2 for shale lithology in Well 15/6-9 S

Groups 2.298e+09 2 1.149e+09 1.665e+03 0
Error 3.316e+09 4066 8.155e+05
Total 5.614e+09 4068
Table C.9: ANOVA table of hypothesis testing #2 for sandstone lithology in Well 15/6-9
S

Groups 1.707e+08 1 1.707e+08 6.009e+02 1.061e-132
Error 3.535e+08 1844 1.917e+05
Total 5.242e+08 1845
82
Appendix D
Results of GR probability density of two

categories: shale and non-shale
The results of probability density of GR estimated by kernel estimator for two cate-
gories shale and non-shale lithology are provided in this section. The results were gen-
erated from GR reading in each hole section of Well 15/5-7 A, Well 15/6-11 S, and Well
15/6-9 S.
D.1 Well 15/5-7 A
(a) Hole section 36"
(b) Hole section 26"
83
D.2. WELL 15/6-11 S
(c) Hole section 17 12 "
(d) Hole section 8 12 "
Figure D.1: Probability density results of each hole section in Well 15/5-7 A using
kernel estimator grouped into shale and non-shale lithology
D.2 Well 15/6-11 S
84
APPENDIX D. RESULTS OF GR PROBABILITY DENSITY OF TWO CATEGORIES:
SHALE AND NON-SHALE
(b) Hole section 26"
(d) Hole section 12 41 "
(e) Hole section 8 21 "
Figure D.2: Probability density results of each hole section in Well 15/6-11 S using
85
D.3. WELL 15/6-9 S
D.3 Well 15/6-9 S
(b) Hole section 17 12 "
Figure D.3: Probability density results of each hole section in Well 15/6-9 S using
86
Appendix E
Results of bivariate analysis of GR and

neutron
87
E.1. WELL 16/1-14
E.1 Well 16/1-14
(a)
(b)
Figure E.1: Scatter plot and probability density plots of GR and neutron for Well
16/1-14 in hole section: (a) 12 14 " and (b) 8 12 "
88
APPENDIX E. RESULTS OF BIVARIATE ANALYSIS OF GR AND NEUTRON
E.2 Well 16/2-7
(a)
(b)
89
E.2. WELL 16/2-7
(c)
16/2-7 in hole section: (a) 17 12 ", (b) 12 14 ", and (c) 8 12 "
90
APPENDIX E. RESULTS OF BIVARIATE ANALYSIS OF GR AND NEUTRON
E.3 Well 16/2-13 A
(a)
(b)
16/2-13 A in hole section: (a) 12 14 " and (b) 8 12 "
91
Bibliography
Bateman, R. M. (2012). Openhole log analysis and formation evaluation. Society of

Petroleum Engineers SPE, Richardson, Tex, 2nd ed. edition.
Bonner, S. et al. (1992). Logging While Drilling: A Three-Year Perspective. Oilfield

Review, 4(3).
Bourgoyne, A. T., Millheim, K. K., and Chenevert, M. E. (1985). Applied Drilling Engi-
neering. Society of Petroleum Engineers, Richardson.
Burke, J., Campbell, Jr., R., and Schmidt, A. (1969). The Litho Porosity Cross Plot: A
New Concept For Determining Porosity And Lithology From Logging Methods. In
SPWLA-1969-Y, SPWLA. Society of Petrophysicists and Well-Log Analysts.
Busch, J., Fortney, W., and Berry, L. (1987). Determination of Lithology From Well Logs
by Statistical Analysis. SPE-14301-PA.
Clavier, C., Hoyle, W., and Meunier, D. (1971). Quantitative Interpretation of Thermal
Neutron Decay Time Logs: Part I. Fundamentals and Techniques. SPE-2658-A-PA.
Clavier, C. and Rust, D. (1976). Mid Plot:a New Lithology Technique. SPWLA-1976-
vXVIIn6a2.
Delfiner, P., Peyret, O., and Serra, O. (1987). Automatic Determination of Lithology
From Well Logs. SPE-13290-PA.
Deveaux, R. D. (1999). Applied smoothing techniques for data analysis. Technometrics,

41(3):263–263.
Dewan, J. T. (1986). Open-Hole Nuclear Logging - State Of The Art. In SPWLA-1986-

MM, SPWLA. Society of Petrophysicists and Well-Log Analysts.
Ellis, D. V. and Singer, J. M. (2010). Well logging for earth scientists. Springer, Dordrecht.
Frigge, M., Hoaglin, D. C., and Iglewicz, B. (1989). Some Implementations of the Box-
plot. The American Statistician, 43(1):50.
Gardner, J. S. and Dumanoir, J. (1980). Litho-Density Log Interpretation. In SPWLA-

1980-N, SPWLA. Society of Petrophysicists and Well-Log Analysts.
Glover, D. (2001). Petrophysics MSc Course Notes. University of Aberdeen.
Hansen, B. E. (2009). Lecture Notes on Nonparametrics.
92
BIBLIOGRAPHY
Kruskal, W. H. and Wallis, W. A. (1952). Use of Ranks in One-Criterion Variance Analysis.

Journal of the American Statistical Association, 47(260):583–621.
Laosripaiboon, L., Saiwan, C., and Prurapark, R. (2015). Reservoir Characteristics In-
terpretation by Using Down-Hole Specific Energy With Down-Hole Torque and Drag.
In OTC-25890-MS, OTC. Offshore Technology Conference.
Larionov, V. (1969). Borehole Radiometry.
Mann, H. B. and Whitney, D. R. (1947). On a Test of Whether one of Two Random Vari-
ables is Stochastically Larger than the Other. The Annals of Mathematical Statistics,
18(1):50–60.
Mwenifumbo, C. (1993). Kernel Density Estimation In The Analysis And Of Borehole

Geophysical Data. SPWLA-1993-v34n5a3.
Ott, L. and Longnecker, M. (2010). An introduction to statistical methods and data anal-
ysis. Brooks/Cole Cengage Learning, Australia ; United States, 6th ed edition.
Poupon, A. and Gaymard, R. (1970). The Evaluation Of Clay Content From Logs. In
SPWLA-1970-G, SPWLA. Society of Petrophysicists and Well-Log Analysts.
Privitera, G. J. (2015). Statistics for the behavioral sciences. SAGE, Los Angeles, second
edition edition.
Provost, Jr., C. (1987). A Real-Time Normalized Rate of Penetration Aids in Lithology

and Pore Pressure Prediction. In SPE-16165-MS, SPE. Society of Petroleum Engi-
neers.
Schlumberger Educational Services (1989). Schlumberger: Cased Hole Log Interpreta-

tion Principles/Applications. Houston.
Schlumberger Wireline & Testing (1998). Log interpretation charts. Schlumberger Wire-
line & Testing, Sugar Land, Texas.
Scott, D. W., editor (1992). Multivariate Density Estimation. Wiley Series in Probability
and Statistics. John Wiley & Sons, Inc., Hoboken, NJ, USA.
Scott, D. W. (2004). Multivariate density estimation and visualization. Papers /

Humboldt-Universität Berlin, Center for Applied Statistics and Economics (CASE)
2004,16.
Serra, O., Delfiner, P., and Levert, J. C. (1985). Lithology Determination From Well-Logs:
Case Studies. In SPWLA-1985-WW, SPWLA. Society of Petrophysicists and Well-Log
Analysts.
Silverman, B. (1981). Density estimation for univariate and bivariate data. In Interpret-
ing multivariate data, pages 37–53. Wiley: Chichester, v edition.
Silverman, B. W. (1986). Density estimation for statistics and data analysis, volume 26.
CRC press.
Simonoff, J. S. (1996). Smoothing methods in statistics. Springer series in statistics.

Springer, New York.
93
BIBLIOGRAPHY
Stieber, S. (1970). Pulsed Neutron Capture Log Evaluation - Louisiana Gulf Coast. In
SPE-2961-MS, SPE. Society of Petroleum Engineers.
Tukey, J. W. (1977). Exploratory data analysis. Addison-Wesley series in behavioral

science. Addison-Wesley Pub. Co, Reading, Mass.
Wilson, R. (1955). Mud Analysis Logging And Its Use In Formation Evaluation. In SPE-
587-G, SPE. Society of Petroleum Engineers.
Ziaja, M. and Roegiers, J.-C. (1998). Lithology Diagnosis Based on the Measurements of
Drilling Forces and Moments at the Bit. In SPE-47799-MS, SPE. Society of Petroleum
Engineers.
94

Automatic Lithology Well Logging

Uploaded by

Copyright:

Available Formats

Automatic Lithology Well Logging

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Automatic Lithology Well Logging

Uploaded by

Copyright:

Available Formats

Automatic Lithology Prediction from Well

Logging Using Kernel Density Estimation

Anisa Noor Corina

Norwegian University of Science and Technology

Anisa Noor Corina

This thesis presents an automatic real-time analysis of lithology interpretation

3 Methodology and data 21

4 Data exploration on GR data 25

5 Univariate KDE analysis on GR data 43

C ANOVA Table Results of Hypotheses Test 80

D Results of GR probability density of two categories: shale and non-shale 83

E Results of bivariate analysis of GR and neutron 87

2.1 Example of time-based surface logging (Bourgoyne et al., 1985) . . . . . . . 5

3.1 Flowchart of methodology in this study . . . . . . . . . . . . . . . . . . . . . 22

4.1 Graphical description of GR data in Well 15/5-7 A which grouped based

2.1 Methods of calculating shale volume . . . . . . . . . . . . . . . . . . . . . . 7

4.1 Statistic description of GR data in Well 15/5-7 A grouped according to the

5.1 Confusion matrix table of 2 sub-population, π1 and π2 . . . . . . . . . . . . 52

C.1 ANOVA table of hypothesis testing #1 in Well 15/5-7 A . . . . . . . . . . . . 81

tool. However, this application can be beneficial to improve lithology interpretation

2.1 Well Logging

Figure 2.1: Example of time-based surface logging (Bourgoyne et al., 1985)

2.1.1 Gamma Ray Logging

Clavier et al. (1971) Vsh = 1.7 − [3.38 − (IGR + 0.7)2 ]0.5

Larionov (1969), for tertiary rock Vsh = 0.083 × [2(3.7058×IGR ) − 1]

Table 2.1: Methods of calculating shale volume

a set of correction charts is available to correct GR manually to environmental condi-

2.1.2 Neutron Log

2.2 Exploratory data analysis

An example is taken from a paper written by Scott (2004) to give an illustration of

Freedman-Diaconis. This method is suitable for data containing a large number of

Histogram is also referred as simple univariate density estimator to approach prob-

2.3 Hypothesis testing

Mann-Whitney U test: rank sum test

For any samples sizes, n a and n b , U parameter has identities:

Depending on n 1 , n 2 , and level of significance, critical value of U , Ucr i t , can be calcu-

Kruskal-Wallis measures a parameter SS bg (R) which is defined as the between-

and the computational formula is shown as

where N is the size of data population.

2.4 Kernel density estimation (KDE)

Simple density estimator

Recall forward approximation of density function,

(#{x i ≤ b j +1 } − #{x i ≤ b j })/n

Smooth univariate estimator

Recall central approximation of density function,

2.4.2 Properties of kernel density estimator

= Bias2 fˆ(x) + Var fˆ(x)

second derivative of density function f . The optimal bandwidth, h 0 , is selected by

and the minimum AMISE follow

Table 2.2: Various kernel functions and forms

Triangular k(u) = (1 − |u|)

(b) Triangular kernel

(d) Triweight kernel (f) Epanechnikov kernel

Figure 2.9: Distribution of K (u) in various kernel functions

Choosing the optimal bandwidth

h 0 = 1.059σn −1/5 (2.29)

Table 2.3: Classification probabilities table

The probability misclassifying an object as a belonging to population π2 when the ob-

and the probability misclassifying an object as a belonging to population π1 when the

2. Equal (or undefined) misclassification cost: c(1|2) = c(2|1). The classification