Journal of Applied Geophysics 194 (2021) 104443

Contents lists available at ScienceDirect

Journal of Applied Geophysics

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

Volcanic lithology identification based on parameter-optimized GBDT

algorithm: A case study in the Jilin Oilfield, Songliao Basin, NE China
Zhichao Yu a, Zhizhang Wang a, *, Fancheng Zeng b, Peng Song b, Bestman Adjei Baffour a,
Peng Wang c, Weifang Wang a, Ling Li a
a
College of Geosciences, China University of Petroleum, Beijing 102249, China
b
PetroChina Jilin Oil Field E&P Research Institute, Songyuan, Jilin 138000, China
c
Sinopec Petroleum Exploration and Production Research Institute, Beijing 10083, China

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

Keywords: The reservoir rocks in the volcanic strata of the Jilin oil field are characterized by great complexity and diversity
Volcanic in composition and structure of lithology. To enhance the rate of lithology identification in subsurface is very
Lithology identification laborious. However, lithology identification is often ignored in quantitative studies, though it is the basis for
Songliao Basin
reservoir characterization. In this paper, an ensemble learning algorithm named gradient boosting decision tree
Ensemble learning algorithm
(GBDT) was used to establish the classification model for the volcanic lithology identification of the Lower
GBDT
Cretaceous Yingcheng Formation in the Songliao Basin, NE China. At the same time, support vector machine
(SVM), logistic regression (LR) and decision tree (DT) classification models were also adopted in contrast with
the classification accuracy of GBDT model. Subsequently, the optimal key parameters for each model were
determined by employing validation curves and GridSearchCv. These results indicate that the GBDT model is
superior to the single classifier and can accurately distinguish the lithologic interface of breccia tuff and rhyolite.
Moreover, it also has better recognition ability for thin layer. It was concluded that the ensemble learning al­
gorithm GBDT has significantly enhanced the accuracy of lithology identification and can be used as a lithologic
identification technology.

1. Introduction produce up to 49.5 × 104 cubic meters of gas per day (m3/d) through the
rhyolite reservoir (Zhang et al., 2008; Wang et al., 2015). China has also
According to history of petroleum, the major targets of petroleum initiated many volcanic hydrocarbon explorations at depths greater than
exploration and exploitation around the globe are often clastic and 3000 m (Zishu and Wu, 1994; Mao et al., 2015; Feng, 2008). China's
carbonate reservoirs. Less concentration is given to the deeply buried volcanic hydrocarbon exploration target has the characteristics of larger
volcanic reservoirs due to their complicated lithologies and lithofacies size and deeper-layer formations compared with other analogous vol­
characteristics (He et al., 2020). Since the first discovery of volcanic canic oil and gas fields throughout the world (Jia et al., 2016). Deep-
hydrocarbon reservoir in San Joaquin basin, California, USA in 1887, layered formations, especially volcanic strata, represent the major tar­
significant advances have been made in volcanic hydrocarbon explora­ gets of exploration in the Songliao Basin. Several huge gas fields with
tion. More than 300 volcanic or volcanic-related reservoirs have been reserves of over 100 billion cubic meters have been found, including
found worldwide, of which 169 volcanic hydrocarbon reserves have Changling, Yingtai, and Wangfu gas fields (Zhang et al., 2015; Zhang
been proven (Petford and Mccaffrey, 2003). Statistically, there exist et al., 2017). As the fourth strategic energy succession region of 100
abundant oil and gas resources in global volcanic reservoirs, containing billion cubic meters of natural gas exploration of Jilin oilfield, industrial
a total of 65.5 × 108 tons of proven oil reserves and 36 × 108 tons of gas gas flow has been obtained from wells in the strata of Huoshiling for­
(Schutter, 2003). An example is the Cristales oil field of the North Cuba mation, Shahezi formation and Yingcheng formation in Dehui fault
basin where more than 3425 tons of oil per day (t/d) were successfully depression (Fig. 1a, b). The main formation of this gas field is Cretaceous
extracted from depths exceeding 2000 m (Zou et al., 2008). Addition­ in age, buried at a depth of more than 3000 m. However, the compli­
ally, in the Yoshii-Kashiwazaki gas field in the Niigata Basin, single wells cated geological conditions of volcanic gas reservoir have resulted in

* Corresponding author.
E-mail address: [email protected] (Z. Wang).

https://doi.org/10.1016/j.jappgeo.2021.104443
Received 9 December 2020; Received in revised form 3 July 2021; Accepted 24 August 2021
Available online 28 August 2021
0926-9851/© 2021 Elsevier B.V. All rights reserved.
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 1. Location of the Dehui Fault Depression and division of its tectonic units. (a) Tectonic components of the Songliao Basin, including two uplifts, two de­
pressions, one plunge and one slope area. The Dehui Fault Depression is located in the Southeast Uplift of the Songliao Basin. (b) The Baojia Sag, located in the north-
eastern part of the Dehui Fault Depression of the Southeast Uplift(modified from Jilin Oilfied). (c) The depth-domain map of Cretaceous strata in the study area.

great complexity and diversity in composition and structure of volcanic various applications in different research areas, but they all have their
lithology and lithofacies. own limitations. Supported vector machine algorithm is difficult to
In recent years, researchers have reached broad agreement about the implement for large-scale training samples and neural network is easy to
physical properties and reservoir space of volcanic reservoirs in different fall into local optimum (Alpaydin, 2014; LeCun et al., 2015). Logistic
regions(Sun et al., 2019; Feng et al., 2018; Gong et al., 2017). However, regression and decision tree are easy to under fit, resulting in low clas­
for volcanic oil reservoirs, no practical mature methods or techniques to sification accuracy (Guoyin et al., 2018; Guo and Liu, 2016; Camila
effectively identify the volcanic lithology are available. In the research et al., 2018).
of igneous reservoir, lithology identification is the basis of reservoir The goal of this study is to distinguish the acidic volcanic rocks with
characterization (Ye et al., 2017; Han et al., 2018). Different types of similar composition but different structures in deep buried volcanic
volcanic lithology have distinct petrological characteristics and mineral strata within the working area of Baojia sag, Dehui fault depression
assemblage types, and different logging numerical characteristics and (Fig. 1b). Based on the conventional logging data of the working area,
pattern combination relationship of logging curves. It is very chal­ this paper innovatively introduces the ensemble learning algorithm
lenging to improve the rate of lithology identification in igneous reser­ named gradient boosting decision tree (GBDT) into the study of lithol­
voirs. Core analysis is the most direct and effective method for ogy identification. Experiments show that the GBDT algorithm has
identifying the lithology of volcanic rocks, but due to the high cost of higher classification performance than the traditional classification
coring, it is almost impossible to take cores in every single well. It is, model and can be used as a lithologic identification technology. The
therefore, crucial to make full use of conventional logging data to results allow us to improve our analysis of volcanic reservoir intervals
identify the lithology of volcanic rocks. The lithology in the study area is and subsequent identification of volcanic lithofacies.
predominantly acidic volcanic rocks, therefore, the change in compo­
sition is relatively small, while variation of structure exhibits different 2. Geological setting
logging response of resistivity (RLLD, RLLS), density (DEN), acoustic
slowness (AC), and compensated neutron log (CNL) values. A more The Dehui Fault Depression is 4053 km2 in size and is located in the
traditional method of identifying volcanic lithology based on well log middle of southeast uplift of Songliao Basin. The Dehui Fault Depression
data is the use of cross plots (Zhang et al., 2017). At present, many is a secondary tectonic unit of Songliao Basin, which is east of the
machine learning methods have been introduced into volcanic lithology Nongan Uplift, south of the Wangfu Fault Depression, west of the Jiutai
identification, including neural network, support vector machine (SVM), Uplift, and north of the Huaide Uplift and the Lishu Fault Depression
logistic regression (LR) and decision tree (DT). These methods have (Fig. 1a). The Dehui Fault Depression can be further divided into seven

2
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 2. Stratigraphic column of the Jilin Oilfield.

sub structural units, namely the Nongan graben, the Huajia sag, the construction accompanied with the strong tectonic movement (Hui­
Baojia sag, the Helong sag, the Lanjia sag, the Nongan South sag and the guang et al., 2011),thus resulting in forming volcanic structural traps in
Longwang sag. The study area, which measures 300 km2, is located in the local area, which provide favorable reservoir for hydrocarbon
the Baojia Sag (Fig. 1b). enrichment of Yingcheng formation.
Within this gas field, hydrocarbon reservoirs have been identified in
Cretaceous and Jurassic strata. Natural gas production from Cretaceous 3. Data and method
reservoirs accounts for a large amount of total gas production (Libin
et al., 2006). The Cretaceous strata can be divided into two series (Upper The study area consists of the Baojia sag of the Jilin Oilfield (Figs.1b
and Lower Cretaceous) and nine formations (Fig. 2). From bottom to top, and 1c). To date, a total of 26 wells have been drilled in the Cretaceous
these Cretaceous strata are divided into the Shahezi Formation (K1sh), strata, including 18 vertical wells, and 8 inclined wells, with depths
Yingcheng Formation (K1y), Denglouku Formation (K1d), Quantou exceeding 3000 m. Core samples (including thin sections) and imaging
Formation (K1q), Qingshankou Formation(K2qn), Yaojia Formation logging data (mostly in the form of formation microscanner images
(K2y), Nenjiang Formation(K2n), Sifangtai Formation(K2s) and Min­ (FMI)) provided direct evidence for volcanic rock identification. Hence,
gshui Formation(K2m) (Yang et al., 2019; Jing and Liande, 2016). conventional logging data, which are calibrated by cores and FMI, could
The Baojia Sag experienced multiple stages of tectonic movements in serve as sufficient information for establishing machine learning model
the early Yingcheng formation(Shuangfang et al., 2010). Meanwhile, for identification of volcanic rocks.
multi-phase volcanic eruption has generated large sets of volcanic Ensemble learning is a new machine learning paradigm, which

Fig. 3. Workflow of GBDT algorithm.

3
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 4. Characteristics of volcanic reservoir cores: (a) Rhyolite, marked by vesicular structure, well DS63, 2394.60 m. (b) Dacite core interval, a series of steep
fractures were encountered, well DS7, 2083.00–2085.70 m. (c) Tuff, steep fracture filling calcite, well DS83, 4086.08 m. (d) Tuffite, well DS80, 2562.00 m. (e)
Ignimbrite, characterized by welded structure, with coarse grain size, well DS80, 2582.50 m. (f) Breccia tuff, finer grain size than ignimbrite, well DS80, 2543.00 m
(g) Diabase, fine to medium grain size, well DS80, 4158.00 m.

constructs multiple learners to solve the same problem (Avnimelech and [ ]F(x)=Fm− 1(x)
Intrator, 1999; Elghazel and Aussem, 2015; Miyoshi et al., 2006). By rim = −
∂L(yiF(xi) )
f or i = 1, 2, …, n. (2)
referring to the gradient descent method, its underlying principle which ∂F(xi)
is training the newly added weak classifier according to the negative n
gradient information of the loss function of the current model is applied (b) In accordance with the training set{(xi, rim)} , a base lear­
i=1
(Fig. 3; Li et al., 2018; Sakhnovich, 2007). Subsequently, the trained nerhm(x)is constructed to fit the pseudo residual.
weak classifier will be appended to the existing model. The GBDT al­ (c) The multiplierγm is calculated by the following one-dimensional
gorithm can be explained as the adoption of a decision tree as the weak optimization problem:
classifier in a gradient boosting algorithm (Liao et al., 2016; Jin Yuan
∑
n
et al., 2018; Xin et al., 2019). The workflow of GBDT algorithm is as γ m = arg min L(yiFm− 1 (xi) + γhm (xi) ) (3)
follows: γ i=1
(1) Initializing the model with constantγ0
(d) The model is then updated
∑
n
F0 (x) = argmin L(yiγ0) (1) Fm (x) = Fm− 1 (x) + γm hm (x) (4)
γ0 i=1
(3) OutputFM(x)stands for the prediction of a strong classifier
(2) For m from 1 to M: composed of a series of weak decision tree models.
(a) The negative gradient of loss function is used to approximate the
value of residual in the current modelFm − 1(x):

4
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 5. Thin-section photomicrographs: (a)

Spherulite rhyolite, developing vesicular
structure, well DS110, 3220.00 m. (b)
Dacite, micro-fractures filled by organic
matter, carbonate metasomatism were dis­
played locally, well DS17, 2237.34 m. (c)
Tuff, chloritization and carbonate meta­
somatism can be observed, well DS104,
3060.00 m. (d) Tuffite, no porosity were
observed under microscope, carbonate
metasomatism and micro-fractures filled by
organic matter were recorded, well DS33,
2897.60 m. (e) Breccia tuff, tuff fillings were
obviously dissolved, well DS80, 2740.60 m.
(f) Ignimbrite, under plane polarized light,
dissolved microfractures and dissolved pores
(generated by devitrification of tuff, seen in
Fig. 7g) were encountered, well DS80,
2522.00 m. (g) Ignimbrite, under perpen­
dicular polarized light, devitrification of tuff
and carbonate metasomatism were observed,
well DS80, 2522.00 m. (h) Diabase,the
chloritization in the feldspar was developed
and micro pores were observed, well DS80,
4158.00 m.

4. Application and results volcanic rocks within the borehole, is an effective tool for us to better
understand the development of volcanic reservoirs. Through the above
4.1. Rock type and characteristics dataset, there are mainly three categories (volcanic lava, pyroclastic
rock and intrusive rock) and seven kinds (rhyolite, dacite, tuff, tuffite,
Detailed lithology analysis was performed on the studied wells on breccia tuff, ignimbrite and diabase) of volcanic rock types in the study
whole cores and core plugs of the Yingcheng Formation. Accordingly, area.
these core samples were analyzed with casting and conventional thin Volcanic lava mainly consists of rhyolite and dacite. Core from (a)
section. Additionally, imaging logging, which can obtain high- depth of 2394.6–2394.8 m in Well DS63 in the Yingcheng Formation
resolution images (as data are collected by vertically scanning the for­ represents a typical rhyolite development section, the core sample is
mation at 2.5 mm intervals) and identify structure characteristics of generally greyish-white and features the development of vesicular

5
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Table 1
Well log responses of different volcanic lithology.
Volcanic lithology GR AC RHOB CNL RLLD RLLS
(API) (μs/ft) (g/cm3) (%) (Ω⋅M) (Ω⋅M)

Rhyolite 128.4–217.8 53.0–59.3 2.06–2.66 1.05–13.02 96.1–1990.6 73.1–1686.8

172.8 55.8 2.59 3.28 874.6 248.4
Dacite 130.8–201.1 54.2–63.7 2.50–2.67 3.02–8.57 35.8–332.7 43.5–610.3
153.5 57.9 2.61 5.22 181.3 251.2
Tuff 80.3–321.3 50.2–73.7 2.04–2.71 3.54–20.05 44.4–1986.2 41.2–1981.5
159.7 57.3 2.54 8.75 612.1 586.4
Tuffite 107.7–266.0 51.3–74.8 2.11–2.64 5.86–38.14 22.9–1961.5 27.4–1625.5
167.2 57.6 2.53 14.0 346.6 314.6
Breccia tuff 76.8–324.1 51.8–67.1 2.02–2.70 4.14–23.14 30.6–1796.8 35.2–1852.0
167.2 57.3 2.57 10.47 345.4 334.3
Ignimbrite 35.3–99.7 56.3–83.9 2.18–2.69 5.87–38.01 8.6–733.1 7.8–523.4
606 71.9 2.54 23.46 48.2 36.9
Diabase 22.7–84.6 48.7–69.7 2.12–2.91 0.14–30.99 10.3–1948.4 7.9–1667.1
39.7 55.4 2.71 11.00 257.4 149.9
Minimum-Maximum
Average

Fig. 6. FMI images of different volcanic lithology: (a) Rhyolite, showing the development of flow structure, well DS83,4341.00–4343.00 m. (b)Tuffite, featured by
layering development, well DS80,2925.00–2927.00 m. (c)Tuff with tuffaceous structure, well DS83,4138.00–4140.00 m. (d) Breccia tuff, porphyritic structure, well
DS83,4288.00–4290.00 m. (e) Ignimbrite, marked by blocky structure, well DS80,2581.00–2583.00 m. (f) Diabase, blocky structure with well-developed fracture,
well DS80,4158.00–4160.00 m.

6
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 7. The multi-curve cross plot of volcanic lithology.

structure (Fig. 4a). It shows a very characteristic flow structure (Fig. 6a) Pyroclastic rock comprises of tuff, tuffite, breccia tuff and ignimbrite.
and is marked by vesicular characteristic under microscope (Fig. 5a). Their conventional well logging values are recoded in Table 1. Inter­
Conventional logging was characterized by high natural gamma ray section of two kinds of conventional logging data shows that there are
radiation (>150API), low acoustic slowness, high density, low neutron overlapping zones over different lithology (Fig. 7). The cores spanning a
porosity, low resistivity and moderate-to-low compensated neutron log depth of 2500 to 4200 m were obtained from Well DS83 and DS80 in the
values (Table 1). The dacite core was obtained from a depth of 2083.00 Yingcheng Formation. Tuff and tuffite have greyish-white and grey color
to 2085.70 m in Well DS7 and the core section contains a series of steep (Fig. 4c, d), with grain size less than 2 mm, and the tuff core has been
fractures (Fig. 4b). Additionally, micro-fractures filled by organic matter partially cemented by calcite (Fig. 4c) while the tuffite has the charac­
were also recorded during the thin section studies (Fig. 5b). Compared teristic of layering development (Fig. 4d). In contrast, breccia tuff and
with rhyolite, which are relatively denser, the dacite interval exhibited ignimbrite have greenish grey or dark color and breccia makes up 15%
lower resistivity (RLLD, RLLS), lower density (DEN), lower natural gamma portion of the core sample (Fig. 4e, f). Petrographic studies on cores and
ray radiation (GR), higher acoustic slowness (AC), and higher compen­ thin section samples obtained from the Yingcheng Formation indicated
sated neutron log (CNL) values (Table 1). that the Yingcheng Formation has been subjected to various diagenetic

Fig. 8. Sketch of 10-folder cross validation.

7
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 9. Validation curves for key parameters of different classifiers.

8
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 10. Learning curve of different classifiers.

alterations. Dissolution is one of the most important diagenetic pro­ reliability of the proposed method. The methods of DT, SVM, LR, and
cesses in terms of porosity generation in the Yingcheng Formation. It is GBDT algorithms were used respectively to construct the model.
recorded in most thin sections in the pyroclastic rock and is marked by
feldspar dissolution facies (Fig. 5e, f). Moreover, the Yingcheng For­ 4.3. Parameter optimization
mation has been greatly affected by metasomatism during the burial
within the area. Carbonate metasomatism are the main features 4.3.1. Cross validation
observed in thin sections of this formation (Fig. 5g). The pyroclastic Cross validation is a common approach in machine learning to build
rocks appear as distinct structure on FMI image. Tuffite, tuff, breccia tuff models and verify model parameters. It can be explained as follows: the
and ignimbrite are characterized by lamellar, tuffaceous, porphyritic, training data set were grouped into training and validation data sets.
and blocky structure, respectively (Fig. 6b, c, d, e). The training set was used to train the classifier while the validation set
Diabase is the most common intrusive rock in the study area, which was employed to test the model trained by the training set (Krogh and
belongs to basic hypabyssal intrusive rock with fine to medium grain Vedelsby, 1995). In this paper, 10-folder cross validation (Fig. 8) was
size through core observation (Fig. 4g). Also, the feldspar is alternated adopted to conduct the experiment. The original training data were
by chlorite and micro pores were observed under microscope (Fig. 5h). grouped into 10 sections (generally equally), a validation set was set for
The diabase is imaged as blocky structure on FMI (Fig. 6f). each subset of the data, and the rest of the 9 groups were taken as the
training set. Consequently, we can get 10 models. Average classification
4.2. Model establishment of lithology identification accuracy of the validation set of the 10 models is regarded as the per­
formance index of the classifier under the 10-folder cross validation. 10-
The fundamental data of lithology analysis is various logging curves folder cross validation can effectively avoid the occurrence of over
related to lithology, therefore, conventional logging data, which are fitting and under fitting. Furthermore, the final result is more convincing
calibrated by cores and FMI, can serve as training data to identify li­ in the sense that almost all samples in each round have been used for
thology. On the basis of logging response mechanism of petrophysics, 6 training the model.
conventional logs (GR, AC, RLLD, RLLS, DEN, CNL), which are sensitive to
lithology, were selected to be the original input features of the model, 4.3.2. Validation curves
whereas the output is volcanic lithology. Before the experiment, the The validation curve determines whether the model is over-fitting or
training data were preprocessed to improve the quality of the data: the under-fitting by virtue of the values of some hyperparameters by
abnormal data caused by human or external factors were eliminated. drawing the curve of a single hyperparameter with training score and
Subsequently, the data were normalized from 0 to 1, avoiding the in­ cross validation score (Airola et al., 2011). In instances where the
fluence of dimension on prediction results. These ‘calibrated’ data were training score and the cross validation score are very low, the classifier
collected from Well DS80, Well DS83, Well DS7 and Well DS63. A total may not be suitable due to possible occurrence of under-fitting. Over-
of 9724 data points were obtained from the Yingcheng formation. 6806 fitting may also occur if the training score is high and the validation
(approximately 70%) data points were randomly selected as training score is low. The performance curves of vital hyperparameters for
samples and the rest served as test samples to check the efficacy and different model are drawn in Fig. 9.

9
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 11. Confusion matrix of f1 score over different classifiers.

4.3.3. Grid search ordinarily designated as positive class while other classes are designated
By selecting a reasonable parameter range, the optimal parameter set as negative class. The prediction results of the classifier on the test set
of the model can be sought out automatically by using Grid search (Ji are only true or false. Confusion matrix is a visual tool in supervised
et al., 2008). The essential idea is to divide the parameters to be opti­ learning, which is mainly employed to compare the prediction results
mized into grids in a certain space, searching for the optimal parameters and genuine results (Mitchell, 2003; Yang et al., 2019). Each row in the
by traversing all the intersections in the grid. In accordance with the matrix represents the predicted label, and each column represents the
given model, cross validation can be automatically carried out. Addi­ true label. The confusion matrix of classification results is shown in
tionally, the validation curve can help specify the search scope, thus Table 3.
considerably increasing the efficiency. Table 2 shows the parameters However, accuracy is not the only index used in unbalanced binary
requiring adjustment of each model and corresponding optimal classification. PrecisionPr, recallReandf1score are also employed as
parameters. important indexes to evaluate the performance of the algorithm. They
are defined as follows:
4.4. Model evaluation Pr = NTP /(NTP + NFP ) (5)

There are many evaluation indexes in machine learning over classi­ Re = NTP /(NTP + NFN ) (6)
fication problems, such as learning curve, F1 score, ROC curve. Fig. 10
shows the learning curve of the four classifiers. The result shows that f 1 = 2(1/Re + 1/Pr)− 1
=2
PrRe
(7)
SVM has the best performance in a single classifier, while LR and DT Pr + Re
have the problem of underfitting. Meanwhile, the GBDT model out­ Where precisionPr represents the true positive proportion of the
performs the single classifier and the score can reach approximately 0.9. samples predicted as positive (i.e. how many of the found were correct
With regard to binary classification, the concerned classes are

10
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 12. Lithology identification results of well DS83.

hits). Recall literally is how many of the true positives were recalled Fig. 12 shows the result of volcanic lithology identification of well
(found), i.e. how many of the correct hits were found.f1 score is the DS83 from a depth of 4086 to 4396 m. Identification results of different
harmonic average of precision and recall. It is commonly used as eval­ classification models are contrasted with the logging profile (which has
uation metrics in many cases because there is a reciprocal relationship been accurately calibrated by FMI) of this well. It can be seen that the
between precision and recall. A stable algorithm should simultaneously GBDT model has the best performance on lithology identification. It has
maximize both precision and recall. Blindly increasing either precision a satisfactory application on breccia tuff, tuff and rhyolite, which can
or recall may be of no significance since f1 score will be high under accurately distinguish their lithologic interface. Nevertheless, the tuff
circumstance of either high precision or recall. and tuffite are easily misjudged and intersected. The confusion matrix
Tables 4 and 5 show the precision, recall, andf1score for each type of also indicates that the two lithologies tend to be obscure, which is
lithology on the test set over different classifiers. The result showed that mainly due to the little difference in mineral composition contents with
SVM has the best performance in a single classifier, and the average distinct origin. Additionally, thin layers with small thickness can be
value off1score reached 81.2%, while LR and DT can only reach 72.8% identified by the GBDT model. Considering the breccia tuff interval
and 70.2%, respectively. Evidently, GBDT has significantly enhanced developed within a buried depth of 4225.1–4226.7 m, only the recog­
the efficiency of the classifier by increasing the f1 score up to 87.1%. nition results of GBDT model is consistent with the lithologic interpre­
tation results of logging. Fig. 13 shows the result of volcanic lithology
identification of well DS107 from a depth of 2800 to 2950 m. It can be
4.5. Results
seen that the GBDT and SVM model has the best performance on li­
thology identification. It has a satisfactory application on breccia tuff
Fig. 11 shows the confusion matrix of f1 score of 7 kinds of lithology
and ignimbrite, while DT and LR model tend to misjudge the tuff and
over different classifiers on test samples. Among the single classifier, the
breccia tuff.
f1 score of rhyolite and dacite are relatively high (exceeding 90%), while
tuffite has the lowest, with 41%, 53% and 22% in the DT, SVM and LR,
5. Discussion
respectively. Evidently, the model constructed by GBDT algorithm
performs superior to the three single classifiers, the f1 score of tuffite
Volcanic reservoirs in deep buried Cretaceous strata are heteroge­
increased to 63%. The average accuracy of each lithology can reach up
neous, making effective exploration relatively challenging. Lithology
to 85.6%.

11
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Fig. 13. Lithology identification results of well DS107.

Table 2 Table 3
Critical hyperparameters and optimal parameter values for each classification Confusion matrix of classification results.
model. True label Predict label
Classification Optimized parameter Search Optimal
Positive Negative
model range parameter
Positive NTP NFN
DT Feature selection criterion Gini/ Entropy
Negative NFP NTN
Entropy
Maximum depth of the 1– 20 10
tree
SVM Parameter γ of kernel 1– 50 19
avoid the occurrence of over-fitting and under-fitting, 10-folder cross
function validation was adopted to conduct the experiment so that all samples in
Penalty C 10− 3– 102 82.86 each round could be used for training the model. It is critical and
LR Regularization strategy L1/L2 L1 challenging to achieve the best accuracy without over-fitting and under-
Penalty C 10− 3– 102 79.06
fitting. Validation curves and GridSearchCv were employed to deter­
GBDT Number of estimators 1– 100 56
Learning rate 0.01– 1 0.23 mine the optimal key parameters for each model to find the balance
Subsample 0– 1 0.16 between under-fitting and over-fitting. These results indicated that SVM
Maximum depth of the 1– 20 7 and DT outperformed LR in the single classifier. Logging curves are
tree highly nonlinear related. A strong classifier such as SVM and DT can
better adapt to the irregular data while LR is a linear model. Therefore,
identification is the groundwork in research of this kind of reservoir. they can achieve better classification results than LR. The GBDT model is
Therefore, it is crucial to undertake related work to enhance the accu­ superior to the single classifier and can accurately distinguish the lith­
racy of lithology identification (Zhang et al., 2015). In this study, we ologic interface of breccia tuff and rhyolite. Moreover, it also has better
presented a new ensemble learning algorithm (namely GBDT) to address recognition ability for thin layer. Rhyolite and dacite have large quantity
this problem. Firstly, conventional logging data, which are calibrated by of samples, the number of which is approximately five times the tuffite.
cores and FMI, can serve as training data to help establish the model. To This makes the classification result of rhyolite obviously surpass that of

12
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Table 4
Precision, recall and f1 score for 10-fold-cross validation over DTC and GBDT classifier.
Lithology Sample DTC classifier Sample GBDT classifier
number number
Pr Re f1 Pr Re f1

rhyolite 462 0.94 0.58 0.72 462 0.98 0.92 0.95

dacite 664 0.75 0.97 0.84 664 0.98 1.00 0.99
tuff 202 0.49 0.53 0.51 202 0.73 0.83 0.78
tuffite 85 0.31 0.62 0.41 85 0.86 0.49 0.63
breccia tuff 268 0.66 0.38 0.48 268 0.66 0.86 0.75
ignimbrite 299 0.98 0.96 0.97 299 1 1 1
diabase 802 0.99 0.99 0.99 802 1 1 1

Table 5
Precision, recall and f1score for 10-fold-cross validation over SVM and LR classifier.
Lithology Sample SVM classifier Sample LR classifier
number number
Pr Re f1 Pr Re f1

rhyolite 462 0.99 0.91 0.95 462 0.98 0.89 0.93

dacite 664 0.99 0.99 0.99 664 0.96 0.99 0.97
tuff 202 0.59 0.55 0.57 202 0.45 0.49 0.47
tuffite 85 0.49 0.58 0.53 85 0.50 0.14 0.22
breccia tuff 268 0.61 0.68 0.64 268 0.51 0.61 0.56
ignimbrite 299 1.00 1.00 1.00 299 0.94 0.97 0.96
diabase 802 1.00 1.00 1.00 802 0.99 0.99 0.99

the tuffite in the single classifier. However, for this kind of unbalanced of f1score reached 81.1%, while LR and DT reached 71.5% and
data, the GBDT algorithm can still determine the features of samples 70.2%, respectively. GBDT significantly enhanced the efficiency
fairly well. It can adaptively adjust the weight distribution of samples, of the classifier and increased thef1 score up to 87.1%.
thereby making it possible to deal with samples that are hard to train
with.
The approach presented in this study could be used to improve the Declaration of Competing Interest
accuracy of volcanic lithology identification. However, the tuff and
tuffite are usually misjudged on account of their similar mineral The authors declare that they have no known competing financial
composition. Future work focused, in particular, on solving the problem interests or personal relationships that could have appeared to influence
of similar logging response with different lithology is recommended. the work reported in this paper.
Great deal of additional scientific research could be undertaken to
determine other features which can distinguish them in origin. Acknowledgements

6. Conclusions This research was performed on Python 3.6 using machine learning
package scikit-learn. This study was supported by the Strategic Priority
(1) Multi-phase volcanic eruption has generated large sets of volca­ Research Program of Jilin Oilfield (Grant No. 2019-FW-027). We are
nic constructions in the Yingcheng Formation, making it rela­ deeply grateful to the Exploration and Development Research Institute
tively hard to identify volcanic lithology due to its strong of Jilin Oilfield, CNPC, for providing research data and publication
heterogeneity. By using a variety of data, including cores, casting permission. The editor and anonymous reviewers are greatly appreci­
and conventional thin sections and imaging logging data, we ated for theirthorough and critical reviews and suggestions, which have
were able to divide the volcanic lithology into three categories significantly improved the quality of this paper.
(volcanic lava, pyroclastic rock and intrusive rock) and eight
kinds (rhyolite, dacite, tuff, tuffite, breccia tuff, ignimbrite, (and) References
diabase) in the Yingcheng Formation within the study area.
(2) In this paper, ensemble learning algorithm GBDT was used to Airola, A., Pahikkala, T., Waegeman, W., et al., 2011. An experimental comparison of
establish a classification model for volcanic lithology with strong cross-validation techniques for estimating the area under the ROC curve[J].
Computat. Stat. Data Analysis 55 (4), 1828–1844.
heterogeneity. The results showed that the GBDT model out­ Alpaydin, E., 2014. Introduction to Machine Learning. The MIT Press.
performs the single classifier. It can accurately distinguish the Avnimelech, R., Intrator, N., 1999. Boosted mixture of experts: an ensemble learning
lithologic interface of breccia tuff and rhyolite, and also has scheme[J]. Neural Comput. 11 (2), 483–498.
Camila, M.S., Leonardo, G.F., Egberto, P., Leonardo Costa, O., 2018. Machine learning
better recognition ability for thin layer. Although the ensemble approaches for petrographic classification of carbonate-siliciclastic rocks using well
learning algorithm GBDT has significantly enhanced the accuracy logs and textural information. J. Appl. Geophys. 155, 217–225.
of lithology identification, there still exist issues in distinguishing Elghazel, H., Aussem, A., 2015. Unsupervised feature selection with ensemble learning
[J]. Mach. Learn. 98 (1–2), 157–180.
the different lithology with similar well logging response. Feng, Z.Q., 2008. Volcanic rocks as prolific gas reservoir: a case study from the Qingshen
(3) To enhance the generalization ability and avoid under-fitting and gas field in the Songliao Basin, NE China[J]. Mar. Pet. Geol. 25 (2008), 416–432.
over-fitting, parameter optimization becomes crucial and critical. Feng, J.W., Ren, Q.Q., Xu, K., 2018. Quantitative prediction of fracture distribution using
geomechanical method within Kuqa Depression, Tarim Basin, NW China. J. Petrol.
Experiments were conducted on how to select the optimal pa­
Sci. Eng. 162, 22–34.
rameters for each model. 10-folder cross validation curves and Gong, L., Gao, S., Fu, X.F., Chen, S.M., Lyu, B.Y., Yao, J.Q., 2017. Fracture characteristics
GridSearchCv were employed to determine the optimal key pa­ and their effects on hydrocarbon migration and accumulation in tight volcanic
rameters for each model to find the balance between under-fitting reservoirs: a case study of the Xujiaweizi fault depression, Songliao Basin, China.
Interpretation 5 (4), 57–70.
and over-fitting. Eventually, these resultsrevealed that SVM has Guo, Y., Liu, Y., 2016. Deep learning for visual understanding: a review.
the best performance in a single classifier, and the average value Neurocomputing 187 (C), 27–48.

13
Z. Yu et al. Journal of Applied Geophysics 194 (2021) 104443

Guoyin, Z., Zhizhang, W., Huaji, L., Yanan, S., Wei, C., 2018. Permeability prediction of Petford, N., Mccaffrey, K.J.W., 2003. Hydrocarbons in Crystalline Rocks [M]. The
isolated channel sands using machine learning. J. Appl. Geophys. 159, 605–615. Geological Society of London, London.
Han, Y., Yuan, C., Fan, Y., 2018. Identification of igneous reservoir lithology based on Sakhnovich, A., 2007. Nonisospectral integrable nonlinear equations with external
empirical mode decomposition and energy entropy classification: a case study of potentials and their GBDT solutions[J]. J. Phys. A Math. Theor. 41(15).
Carboniferous igneous reservoir in Chunfeng oilfield. Oil Gas Geol. 39 (4), 759–765 Schutter, S.R., 2003. Occurrences of hydrocarbons in and around igneous rocks[J].
(in Chinese with English abstract). Hydrocarb. Crystalline Rocks 214 (1), 35–68.
He, H., Li, S., Liu, C., Kong, C.X., Jiang, Q.P., Chang, T.Q., 2020. Characteristics and Shuangfang, Lu, Hui, Sun, Weiming, Wang, et al., 2010. Key factors controlling the
quantitative evaluation of volcanic effective reservoirs:A case study from Junggar accumulation of volcanic gas reservoirs in the deep part of southern Songliao Basin
Basin, China. J. Pet. Sci. Eng. 195, 107723. [J]. J. Daqing Petrol. Instit. 34 (5), 42–47 (in Chinese with English abstract).
Huiguang, Li, Haitao, Xue, Wenbiao, Huang, et al., 2011. Hydrocarbon source rock Sun, H.T., Zhong, D.K., Zhan, W.J., 2019. Reservoir characteristics in the cretaceous
exploration potential of deep layer in Dehui fault depression[J]. Sci. Technol. Eng. volcanic rocks of Songliao Basin, China: a case of dynamics and evolution of the
11 (27), 6578–6582 (in Chinese with English abstract). volcano-porosity and diagenesis. Energy Explor. Exploit. 37 (2), 607–625.
Ji, W., Shuli, K., Chuanning, T., et al., 2008. Meta-prediction of phosphorylation sites Wang, Luo, Jianghai, Li, Yongmin, Shi, et al., 2015. Review and prospect of global
with weighted voting and restricted grid search parameter selection[J]. Nucleic volcanic reservoirs[J]. Geol. China 42 (5), 1610–1620 (in Chinese with English
Acids Res. 4, 4. abstract).
Jia, H., Ji, H., Wang, L., et al., 2016. Tectono-sedimentary and hydrocarbon potential Xin, Q., Hao Jue, H., Qing, Y., et al., 2019. Prediction of Temperature of Asphalt
analysis of rift-related successions in the Dehui Depression, Songliao Basin, Pavement Surface Based on APRIORI-GBDT Algorithm[J]. J. Highway Transpor. Res.
Northeastern China[J]. Mar. Pet. Geol. 76, 262–278 (in Chinese with English Develop. 36 (5), 1–10.
abstract). Yang, X., Wang, Z., Zhou, Z., et al., 2019. Lithology classification of acidic volcanic rocks
Jin Yuan, L., Yong, D., Tao, L.I., 2018. Classification of Flight Delay Based-on GBDT[J]. based on parameter-optimized AdaBoost algorithm[J]. Acta Pet. Sin. 40 (4),
Mathemat. Pract. Theory 48 (4), 1–7. 457–467.
Jing, Zhao, Liande, Bai, 2016. Main controlling factors of high-quality volcanic reservoir Ye, T., Wei, A., Deng, H., 2017. Study on volcanic lithology identification methods based
in southern Songliao basin[J]. Special Oil Gas Reser. 23 (3), 52–56 (in Chinese with on the data of conventional well logging data: a case from Mesozoic volcanic rocks in
English abstract). Bohai bay area. Prog. Geophys. 32 (4), 1842–1848.
Krogh, A., Vedelsby, J., 1995. Neural network ensembles, cross validation, and active Zhang, J., Qin, L., Zhang, Z., 2008. Depositional fades, diagenesis and their impact on the
Learning[J]. Adv. Neural Inf. Proces. Syst. 7 (10), 231–238. reservoir quality of Silurian sandstones from Tazhong area in Central Tarim Basin,
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436. western China[J]. J. Asian Earth Sci. 33 (1–2), 42–60.
Li, L.J., Yu, Y., Bai, S.S., et al., 2018. Towards effective network intrusion detection: a Zhang, D., Zou, N., Jiang, Y., 2015. Logging identification method of volcanic rock
hybrid model index and GBDT with PSO [J]. J. Sens. 6, 1–9. lithology: a case study from volcanic rock in Junggar Basin. Lithologic Reserv 27 (1),
Liao, Z., Huang, Y., Yue, X., et al., 2016. In silico prediction of gamma-aminobutyric acid 108–114.
type-a receptors using novel machine-learning-based SVM and GBDT Approaches[J]. Zhang, L., Zhang, G., Qi, Y., 2017. Lithology identification of carboniferous volcanic rock
Biomed. Res. Int. 2016 (6), 1–12. with logging data in Xiquan Area. Junggar Basin. Xinjiang Petrol. Geol 38 (04),
Libin, Zhao, Zhilong, Huang, Ma, Yujie, et al., 2006. A study on geochemical character 427–431 (in Chinese with English abstract).
and origin of deep natural gas in Dehui fault depression of the southern Songliao Zishu, Zhang, Wu, Banghui, 1994. Investigation of the research status and exploration
basin. Nat. Gas Geosci. 17 (2), 176–182 (in Chinese with English abstract). technology at home and abroad about volcanic reservoir[J]. Natural Gas Explorat.
Mao, Z.G., Zhu, R.K., Luo, J.L., Wang, J.H., Du, Z.H., Su, L., Zhang, S.M., 2015. Reservoir Develop. 16 (1), 1–26 (in Chinese with English abstract).
characteristics, formation mechanisms and petroleum exploration potential of Zou, C.N., Zhao, W.Z., Jia, C.Z., et al., 2008. Formation and distribution of volcanic
volcanic rocks in China. Pet. Sci. 12, 54–66. hydrocarbon reservoirs in sedimentary basins of China[J]. Pet. Explor. Dev. 35 (3),
Mitchell, Tom M., 2003. Machine learning[M]. McGraw-Hill. 257–271. London: Geological Society, London, Special Publications, 2003, 214(1):
Miyoshi, S., Uezu, T., Okada, M., 2006. Statistical mechanics of time domain ensemble 35–68(in Chinese with English abstract).
learning[J]. J. Phys. Soc. Jpn. 75 (8), 2652–2674.

14