Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Classification of rock facies using deep convolutional

neural network
Mohammad Hassan Soleimani
Mohammad Ali Riahi (  mariahi@ut.ac.ir )
Majid Bagheri
Mehran Rahimi

Research Article

Keywords: well data, facies classification, deep learning, convolutional neural network, machine learning

Posted Date: June 16th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-3046109/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License. Read
Full License

Page 1/18
Abstract
Investigating rock facies is very important in the study of hydrocarbon reservoirs. The logs used in this
research are related to facies logs from nine wells from the gas reservoir located in Kansas, USA. The rock
facies identified in this study are based on a visual examination of a 4,149-foot core from 9 wells. The physical
and chemical properties of stones were determined using special tools. The well-log includes five wireline
logging curves (GR, resistivity log (RL), photoelectric effect (PE), difference neutron density porosity (DPHI) and
average neutron density porosity (PHIA)) and two geological limiting variables (non-marine index - Nautical
(NM_M) and relative position (RP)). Our goal in this paper is to develop an effective model based on deep
learning for geological facies classification in wells. The classification of facies is done by studying the
lithological characteristics of rocks, which are the characteristics of modern sediments accumulated in specific
physical and geographical conditions. This study presents a new 1D-CNN model trained on different
optimization algorithms. Using normal well-log data is one of the main advantages of the proposed model.
The proposed model is compared with the support vector machine model and the nearest neighbor model and
shows more accurate results compared to them. This model shows successful results in the study of well log
data and therefore it can be suggested as a suitable and effective approach for processing well log data
required for lithology.

1. Introduction
The approach of this paper is to compare a one-dimensional convolutional neural network (1D-CNN) for facies
classification with KNN and SVM methods to see if facies classification can be improved. The most popular
geophysical investigation of this well, which is carried out with the purpose of revealing oil in a geological
section, is logging. Well-logs are physical and chemical measurements of rocks that are recorded by lowering
specialized sensors into wells after drilling. They allow us to determine porosity, permeability, fluid
composition, information about oil and gas saturation, etc. Facies (a part of a layer that differs in lithological
composition from adjacent layers) have characteristic features that can be found in cores of reservoir rock
samples taken from wells. They are the basis for describing the reservoir and building its model. Experts
analyze them to determine the type and sequence of facies. This is tedious and time-consuming. The process
of drilling a well and obtaining a core for wellbore analysis is expensive. These costs increase significantly
with the increase in the number of wells. Therefore, it is necessary to develop methods for facies determination
using well-log data (Remembrance of Imamordiev and Lyudmila Sukhosta, 2019). Batch normalization layers
usually follow convolutional layers to normalize output. They allow to increase the training speed significantly
and make the network less sensitive to initialization (Ioffe and Szegedy, 2015). Cross-entropy loss can be
considered to determine the loss of the proposed 1D-CNN model (Ceci et al., 2017). The entire sequential
network (all weights and biases) can be optimized by an optimization algorithm using cross-entropy as a
training metric. Adagrad (Duchi et al., 2011), Adadelta (Zeiler, 2012), and Adamax (Kingma and Ba, 2015) are
used in this work. The conventional approach used for automatic facies classification either required the use
of core, petrophysical logs, and borehole images all together to obtain a reliable prediction (Basu et al., 2002);
or provided lithological subjective classifications based on the texture of the data (Chai et al., 2009; Linek et al.,
2007; Newberry et al., 2004; Hall et al., 1996). There has been a lot of excitement lately about big data and the
dire need for data scientists who are capable of extracting meaning from it. Meanwhile, geologists have been
doing science with big data for years without needing to brag about its volume. But now that large and
Page 2/18
complex data sets are widely available, many tools and techniques exist to analyze them. Many free and open-
source packages now exist that provide powerful additions to the geoscientist's toolbox, many of which were
previously only available on proprietary (and expensive) software platforms. A support vector machine (or
SVM) is a type of supervised learning algorithm that must be provided with training data to learn relationships
between measurements (or features) and classes to be assigned (Hal, 2016). K-N is at the opposite end of the
classification spectrum of parametric models (Hastie et al., 2001) and is simply a "match of likeness" where
unknown objects are matched with the most similar objects of the known class. Fuzzy logic addresses the
natural fuzziness of data to generate discrete class assignments when faced with fuzzy predictor variables
and fuzzy class boundaries (Segaf and Nebrija, 2003). Neural networks are non-linear statistical models where
the weights are iteratively adjusted to reduce the error in class prediction (Rahimi and Riahi, 2022). As
technology advances and network devices increase, more devices are connected to the network, resulting in
more data and information on the network, which emphasizes the importance of network security. Malicious
traffic should be detected in networks and machine learning or more precisely deep learning (DL), which is a
future approach that should be used for better detection (Singh et al., 2021). In general, this research aims to
evaluate the performance quality of the one-dimensional convolutional neural network (1D-CNN) method in
facies classification; for this purpose, it was compared with other methods.

2. Area of study
Kansas is located in the central part of the United States and is largely characterized by a series of gently
rolling uplands and broad, flat river valleys. The state is divided into three main physiographic regions: the Flint
Hills in the east, the Smoky Hills in the west, and the High Plains in the northwest. The geology of Kansas is
also diverse, with rocks ranging in age from Precambrian to Quaternary. The eastern part of the state is
underlain by sedimentary rocks of Paleozoic age, including limestone, shale, sandstone, and chert. These rocks
were deposited in a shallow sea that covered the region during the Paleozoic Era and contain fossils of marine
animals such as trilobites, brachiopods, and crinoids. In the central part of the state, the sedimentary rocks of
the Permian Period are well exposed. These rocks include red sandstones, shales, and limestones and are
known for their oil and gas production. The western part of the state is underlain by sedimentary rocks of
Cretaceous and Tertiary age, including sandstones, shales, and limestones. These rocks were deposited in a
shallow sea that covered the region during the Mesozoic and Cenozoic eras. Kansas also has significant
mineral resources, including petroleum, natural gas, coal, salt, gypsum, and lead-zinc ores. The state's
petroleum industry is centered on the Hugoton and Central Kansas Uplifts, where oil and gas are produced
from reservoirs in the sedimentary rocks. Coal mining was once an important industry in southeastern Kansas
but has declined in recent decades. Salt and gypsum are produced from underground mines in central Kansas
while lead and zinc are mined in the state’s southeast. In addition to its sedimentary rocks, Kansas also has a
few areas of Precambrian igneous and metamorphic rocks. These include the granite and gneiss of the
Precambrian basement rocks in the eastern part of the state and the rhyolites and basalts of the Cenozoic
volcanic rocks in the central part of the state Overall, the geology of Kansas reflects the complex history of the
North American continent, from its early Precambrian origins to its more recent sedimentary and tectonic
evolution.)McCauley et al., 1997). Fossil Collecting in the Cretaceous Niobrara Chalk: Kansas Geological
Survey, Open-file Report 97 − 62, 14 p.

Page 3/18
In the Panoma Field, facies-controlled petrophysical properties dictate gas saturations and accurate
discrimination of facies reduces error in predicted permeability and gas volume (Dubois et al., 2003). However,
developing robust geologic models for reservoir analysis of large heterogeneous reservoirs like that in the
Panoma Field is impractical by traditional methods. This study presents a possible solution to facies
prediction using neural network and Markov chain techniques that will make possible the construction of a
geologic and petrophysical model with sufficient detail to accurately represent the fine-scale vertical and lateral
heterogeneities, a requirement for accurate reservoir modeling of the entire field. Gas production in the Panoma
Field is from seven stacked fourth-order nonmarine-marine sequences in Council Grove Group (Permian,
Wolfcampian) in the Hugoton Embayment of the Anadarko Basin deposited in a proximal shallow position on
a gently dipping ramp. Broad facies belts migrated in response to cyclical sea level fluctuations resulting in a
predictable vertical succession of facies within each sequence: 1) nonmarine redbeds: siltstone and fine
sandstone or clay-rich siltstone and shale (coastal plain; two facies) deposited on an exposure surface
(unconformity) on the underlying marine carbonate rock unit; 2) transgressive marine shallow-water
carbonates with grain-supported textures (shoal and mound); 3) deeper water dark marine siltstones (marine
siltstone and shale) or silty carbonate mud and wackestones (normal marine carbonates); 4) agitated shallow
water packstones and grainstones (also categorized as shoal and mound) or quiet shallow water dolomitized
carbonate mudstones and peloidal wackestones (tidal flat and lagoon); and 5) tidal influenced laminated and
fenestral mudstones (also tidal flat and lagoon). Subaerial exposure is evidenced by well-developed root molds
and other weathering and pedogenic features.

2.1 well data


This categorization represents a combination of both lithofacies and depositional facies designations. The
research area is a hydrocarbon field located in southwestern Kansas USA known as the Hugoton Basin a larger
and deeper part of the northern shelf extension that lies beneath Oklahoma and Texas. Over millions of years,
thousands of feet of carbonate (limestone and dolomite) and shale deposits have accumulated. The rock
facies identified in this study are based on a visual examination of a 4 149-foot core from nine wells. The
physical and chemical properties of rocks were determined using special tools. The well-log includes five
wireline logging curves (GR resistivity log (RL) photoelectric effect (PE) difference neutron density porosity
(DPHI) and average neutron density porosity (PHIA)) and two geological limiting variables (non-marine index
Nautical (NM_M) and relative position (RP)). Digital measurements are recorded at half-foot (0.15 m) intervals
(Dubois et al. 2007). The sequence analyzed is represented by nonmarine sandstone coarse and fine siltstone
marine siltstone and shale mudstone wackestone dolomite grainstone-packstone and phylloid-algal
buffelstone.

3. Methodology
The rock facies core and associated well log variables are used as input data to train a Convolutional Neural
Network (CNN) model. This model is suitable for evaluating rock facies in wells without core data but with a
logarithmic curve. Experiments were conducted in parallel with model construction to verify its properties and
workflow. The general workflow of deep convolutional network classification is as follows: Fig. 1

1) Obtain well logs, training datasets, and test datasets to verify the model's performance.

Page 4/18
2) Preprocess the data.

3) Construct a deep CNN for facies classification, determining the number of neurons for each hidden layer.

4) Train the classifier on the well-log data, considering the original data to obtain the optimal solution.

5) Evaluate and apply the obtained classifier to an unknown logging dataset.

3.1 The general structure of the models used (SVMs, KNN,


1_DCNN)
For the input data, a vector is added to the initial layer, followed by several successive convolution and pooling
layers to capture propagation characteristics. Pooling layers are used to subsample and reduce the grid
dimension, aiding in controlling overfitting computations. Each convolutional layer has multiple one-
dimensional convolutional filters (kernels). The kernel size is a hyperparameter dependent on the data. Next,
the feature maps with filters of the same size as the first convolution layer are passed through the second
convolution layer using the ReLU function. The generated feature maps are then sent to the first layer of max
pooling, which reduces the output feature instances of the previous layer for better handling of high-
dimensional data.

Also known as KNN or k-NN, it is a non-parametric, supervised learning classifier that uses proximity to perform
classification or prediction regarding the grouping of an individual data point. The original SVM algorithm was
developed in 1964 by Vladimir Vapnik and Alexei Cherninikis at the Institute of Control Sciences of the Russian
Academy of Sciences. In 1992, Boser, Isabel Guyon, and Vapnik proposed a way for nonlinear classification by
introducing the kernel and incorporating it into SVM. Support vector machines (SVM) are a set of supervised
learning methods used for classification, regression, and outlier detection.

In a 1D-CNN, the input and output are one-dimensional vectors, and the convolution operation is performed
along the time axis. The filter slides over the input vector calculates the dot product between the filter
coefficients and the input values at each position, and produces an output value. The activation function is
then applied to the output value, generating the final output at that position. The weights and biases of the 1D-
CNN are learned through a training process using backpropagation and gradient descent.

The 1D-CNN model was created using the Python programming language in the Google Colab environment,
utilizing various libraries such as NumPy, TensorFlow, pandas, etc. Well-log data was employed to train the
proposed 1D-CNN model for facies classification. The considered dataset consists of data from nine wells,
comprising 4149 samples. To enhance classification accuracy, a pseudo-well named "Recruit F9" was
introduced (Dubois et al., 2007). The dataset includes seven main wells (SHRIMPLIN, SHANKLE, LUKEU,
NOLAN, Recruit F9, NEWBY, and CHURCHMAN BIBLE) with facies types labeled by core. Figure 3 illustrates the
dataset, and Table 1 presents its statistical description.

Page 5/18
Table 1
The statistical description of the data.
DEP GR DeltaPHI ILD_log10 PHIND PEFZ NM_M RELPOS

count 2783 2783 2783 2783 2783 2783 2783 2783

mean 2873.34 66.24 0.64 3.75 13.11 3.8 1.52 0.52

std 138.34 31.61 0.24 5.04 7.38 0.89 0.49 0.28

min 2573.5 13.25 -0.025 -21.83 0.55 0.2 1 0.1

25% 2767.5 46.08 0.49 1.3 8.16 3.2 1 0.27

50% 2895 65.54 0.62 3.58 11.9 3.6 2 0.52

75% 2987 80.71 0.81 6.5 16.14 4.4 2 0.77

max 3122.5 361.15 1.48 18.5 84 8 2 1

Table 2

The total number of each facies in the total


data is presented in
facies_type The number of each facies

SS 259

CSiS 738

FSiS 615

SiSh 148

MS 217

WS 462

D 98

PS 498

BS 161

The feature vector is constructed using the following physical properties: GR, RL, PE, DPHI, PHIA, and two
geological limiting variables (NM_M and RP). Measured properties and their transformations are potential
elements of feature vectors. The dataset includes the following types of facies: non-marine sandstone (NS),
non-marine coarse silt (NCS), non-marine fine silt (NFS), marine silt and shale (MSS), mudstone (limestone)
(M), limestone (W), dolomite (D), aggregate-granule (limestone) (P_G), and P_AB (limestone). Two facies are of
continental origin named NCS and NFS, and six facies are of marine origin (MSS, M, W, D, P_G, and P_AB). In
non-marine areas, NCS usually has a higher permeability than NFS for a given porosity. An input color map for
nine rocks is shown in Fig. 4 to create a lithological description of the interpreted well.
Page 6/18
The standardization method was applied to the data, resulting in a mean and variance of zero. The PE feature
is not available for all wells, only for 3232 out of 4149 samples. From a geological perspective, the boundary
of facies is sometimes unclear, and it can exhibit transitions. The lists of facies labels and their nearest
neighbors are presented in Table 3.

Table 3
Lists of the facies labels and their
nearest neighbors.
Facies Label Adjacent facies

1 SS 2

2 CSiS 1,3

3 FSiS 2

4 SiSh 5

5 MS 4.6

6 WS 5.7

7 D 6.8

8 PS 6.7,9

9 BS 7,8

4. Results
The proposed model is trained on different optimization methods such as Adagrad, Adadelta, and Adamax.
Evaluation of the proposed model for facial classification is presented in Table 4. In comparison to
optimization methods, 1D-CNN (Adamax) is the most stable (with the increase of iteration number). Thus, a
model trained with Adagrad fails in this experiment. Its training and validation accuracies are 56% and
58%respectively. However, the gap for 1D-CNN(Adamax) is 14%, in contrast to the other two models (6% for 1D-
CNN(Adadelta) and 2% for 1D-CNN (Adagrad)). The use of Adamax results in the highest validation and
training accuracies.

Table 4
Evaluation of the proposed model for facies classification.
Model Training Validation Validation Training Adjacent facies
loss loss accuracy (%) accuracy (%) value
score

1D-CNN 1.096 1.046 58% 56% 90%


(Adagrad)

1D-CNN 1.506 1.42 47% 41% 84%


(Adadelta)

1D-CNN 0.367 1.0902 71% 85% 92%


(Adamax)
Page 7/18
The proposed CNN model demonstrates the relationship between the iteration number and different
optimizers. As the number of iterations increases, the model's performance improves, indicating high accuracy.
Among the three optimizers, 1D-CNN (Adamax) exhibits the most stable behavior with an increasing number of
iterations. Overall, the 1D-CNN (Adamax) and 1D-CNN (Adagrad) outperform the KNN and SVM models, as
shown in Table 5.

Table 5
Comparison of the proposed model and KNN,
SVM models for facies classification.
Model Adjacent facies value
score

1D-CNN (Adagrad) 90%

1D-CNN (Adadelta) 84%

1D-CNN (Adamax) 92%

KNN 87%

SVM 86%

The accuracy of both training and validation datasets using 1D-CNN (Adagrad) is low but showed a higher
result than K-NN and SVM. In general, the model, trained with the Adamax optimizer, surpasses the models
trained by Adagrad and Adadelta. The figure showcasing the prediction ability and accuracy of the model
based on the loss function trend, when three optimization methods are applied, can be found in Fig. 6.

Next, we will examine the comparison of the confusion matrix for all five methods depicted in Fig. 7. When
discussing the "Classification" of a "Data Set" using classification methods, the objective is to achieve the
highest precision and accuracy possible, particularly in accurately identifying samples related to specific
categories.

In summary, the accuracy of the validation data for the Adagrad and Adadelta optimizers is close to and
inferior to that of Adamax. As a result, Adamax outperforms Adadelta and Adagrad in this test. A set of
randomly selected wells that do not have labels is used to test the trained model. The obtained results can be
estimated from the well-log data. Therefore, zoning based on well-logs is a very important step in wellbore data
preprocessing. Running the models on the test data showed that, as we concluded, the 1-D-CNN model
(Adamax) has the best performance in facies classification, and this result can be seen in Fig. 8.

Page 8/18
Table 6
Comparison of models using Evaluation metric f-measure
Method Evaluation
metrics
Facies SS CSiS FSiS SiSh MS WS D PS BS macro
avg

SVM F-measure 89 80 78 70 58 65 62 71 85 73
(%)

KNN F-measure 29 70 75 75 55 56 60 73 75 63
(%)

1D-CNN F-measure 73 63 65 60 9 46 0 55 27 44
(Adagrad) (%)

1D-CNN F-measure 0 51 19 0 0 5 0 51 0 14
(Adadelta) (%)

1D-CNN F-measure 92 75 74 78 53 62 80 72 77 74
(Adamax) (%)

In general, according to Fig. 8 and Tables 4, 5, and 6, it can be concluded that the SVM and K-NN models
performed relatively well. Still, in the optimization methods of our proposed model, the model 1D-CNN
(Adamax) Compared to other models, has higher accuracy for Adjacent facies classification.

Conclusion
Numerous studies have addressed the challenge of facies classification using well logs, with machine learning
methods emerging as effective solutions. This paper proposed a CNN-based approach for facies classification
based on borehole measurements. The model was evaluated using three different optimizers: Adagrad,
Adadelta, and Adamax. Key features considered in the model included limiting variables such as PE, GR, RL,
DPHI, PHIA, and geological information. Results demonstrated that the Adamax optimizer yielded the best
prediction performance for the network response. Moreover, the 1D-CNN model outperformed SVM and KNN,
highlighting its superiority in facies classification. The proposed method exhibited significant results for facies
of both marine origin (dolomite, algal phylloid feldspar, and granulite) and continental origin (coarse siltstone
and sandstone). The 1D-CNN (Adamax) model showed substantial statistical improvements in facies and
adjacent facies classification, making it suitable for identifying lithology in complex geological structures. This
approach can be considered a preliminary stage in the quantitative petrophysical analysis of wells. Evaluation
and analysis of the proposed deep learning model can provide valuable insights for future research and facies
recognition. Overall, 1D-CNN proves to be a powerful tool for classification tasks, especially for sequential
data. However, designing an effective 1D-CNN architecture requires careful consideration of the specific task,
characteristics of the input data, and appropriate preprocessing and meta-parameter tuning.

Declarations
Acknowledgments
Page 9/18
The authors express their gratitude to the Research Council of the University of Tehran for their support.

References
1. Basu, T., 2002. Automated facies estimation. from integration of core, petrophysical logs, and borehole
images AAPG Annual Meeting (2002), pp. 1-7.
2. Bohling, G.C., Dubois, M.K., 2003. Integrated Application of Neural Network and Markov Chain Techniques
to Prediction of Lithofacies from Well Logs (Kansas Geological Survey Open File Report 2003-50). Kansas
Geological Survey Lawrence, Kansas, USA.
3. Ceci, M., Hollmén, J., Todorovski, L., Vens, C., Dzeroski, S., 2017. Machine learning and knowledge
discovery in databases - European conference, ECML PKDD 2017, Skopje, Macedonia, September 18-22.
4. Chai, H., Li, N., Xiao, C., Liu, X., Li, D., Wang, C., Wu, D., 2009. Automatic discrimination of sedimentary
facies and lithologies in reef-bank reservoirs using borehole image logs Appl.Geophys., 6 (2009), pp. 17-29
5. Cortes, C., Vapnik, V., 1995. Support-vector networks - Machine learning– Springer.
6. Cover, T., Hart, P., 1967. Nearest neighbor pattern classification - IEEE Transactions on information theory.
Ieeexplore.ieee.org. DOI: 10.1109/TIT.1967.1053964
7. Duchi J., 2011. Adaptive subgradient methods for online learning and stochastic optimization J. Mach.
Learn. Res., 12 (2011), pp. 2121-2159
8. Dubois, M. K., Bohling, G.C., Chakrabarti, S., 2007. Comparison of four approaches to a rock facies
classification problem.Computers & Geosciences 33, 599–617.
9. Dubois, M.K., Byrnes, A.P., Bohling, G.C., Seals, S.C., Doveton, J.H., 2003. Statistically-based lithofacies
predictions for 3-D reservoir modeling: an example from the Panoma (Council Grove) Field, Hugoton
Embayment, southwest Kansas, KGS Open-file Report 2003-30).
10. Hastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning: Data Mining, Inference,
and Prediction. Springer, New York, 533pp.
11. Hall. B., 2016. Facies classification using machine learning - The Leading Edge. library.seg.org.
doi.org/10.1190/tle35100906.
12. Hall, J., Ponzi, M., Gonfalini, M., Maletti, G., 1996. Automatic extraction and characterisation of geological
features and textures from borehole images and core photographs SPWLA 37th Annual Logging
Symposium, Society of Petrophysicists and Well-Log Analysts.
13. Imamordiev.Y., Sukhosta, L., 2019. Lithological facies classification using deep convolutional neural
network . DOI: doi.org/10.1016/j.petrol.2018.11.023
14. Ioffe, S., Szegedy, C., 2015. Batch normalization: accelerating deep network training by reducing internal
covariate shift32nd International Conference on Machine Learning.
15. McCauley, J., Buchanan, Sawin, R., 1997. Fossil Collecting in the Cretaceous Niobrara Chalk: Kansas
Geological Survey, Open-file Report 97-62, 14 p
16. Newberry, B.M., Hansen, S.M., Perrett, T.T., 2004. A Method for Analyzing Textural Changes within Clastic
Environments Utilizing Electrical Borehole Images
17. Linek, M., Jungmann, M., Berlage, T., Pechnig, R., Clauser, C., 2007. Rock classification based on resistivity
patterns in electrical borehole wall images J. Geophys. Eng., 4 (2007), p. 171
Page 10/18
18. Rahimi, M., Riahi, M, A., 2022. Reservoir facies classification based on random forest and geostatistics
methods in an offshore oilfield, Journal of Applied Geophysics, ISSN 0926-9851,
https://doi.org/10.1016/j.jappgeo.2022.104640.
19. Saggaf, M.M., Nebrija, E L., 2003. A fuzzy logic approach for the estimation of facies from wire-line logs.
American Association of Petroleum Geologists Bulletin 87,7, 1223–1240.
20. Singh, K., Mahajan, A., Mansotra, V., 2021. 1D-CNN-based Model for Classification and Analysis of
Network Attacks International Journal of Advanced Computer Science and Applications, DOI:
10.14569/IJACSA.2021.0121169
21. Zeiler, M.D., 2012. Adadelta an Adaptive Learning Rate Method arXiv: 1212.5701v1 [cs.LG]

Figures

Figure 1

The Panoma Field is located in the Hugoton Embayment of the Anadarko Basin in the southwest (Kansas
Geological Survey Open File Report 2003).

Page 11/18
Figure 2

Flowchart of the proposed approach.

Page 12/18
Figure 3

Surface and log curve of seven main wells a, b, c, d, e, f, g respectively (LUKEU, SHRIMPLIN, NOLAN, NEWBY,
SHANKLE CHURCHMAN BIBLE and Recruit F9).

Page 13/18
Figure 4

Distribution of facies for lithology description of the wellbore.

Page 14/18
Figure 5

Presents the cross vector plots and lithology distributions for data sets from well-logs

Page 15/18
Figure 6

Loss function curves and facies classification accuracy for training and validation datasets for 1D-CNN
(Adagrad),1D-CNN, (Adadelta), 1D-CNN (Adamax), KNN, and SVM.

Page 16/18
Figure 7

Comparison of data confusion matrix for 5 models used in the test.((a) 1D-CNN (Adamax),(b)1D-CNN,
(Adadelta), (c)1D-CNN (Adagrad),(d) KNN,(c) SVM

Page 17/18
Figure 8

Schematic comparison of learning models and the proposed model on test data for facies classification.

Page 18/18

You might also like