An intelligent technique for the characterization of coal microscopic images using ensemble learning

Abstract

Coal is a primary natural resource of fuel that is efficiently used for electricity generation, steel or iron production, and household usage. Characterization is needed for industries to understand the quality of coal before shipping. Currently, industries follow chemical, microscopical, and machine-based analysis as the gold standard for coal characterization. These conventional analyses of coal are an indispensable method over the years and have tested by high skilled petrologists. Though, these types of optical or machine-dependent recognition of coal category samples are quite slow, expensive, and restricted by subjective analyses with less accuracy. The main aim of this research is to propose an accurate, time, and cost-effective machine learning-based automated characterization system by analyzing coal color and textural features. This paper comes up with a quantitative approach toward the characterization of dissimilar types of coal for better utilization in industries. The proposed ensemble learning coal characterization method provides an accuracy of around 97% and takes less computational time than conventional methods. Hence, the proposed automated coal characterization system provides support to industries in the development of computer-aided assessment of coal category samples.

Keywords

Coal HSV GLCM image processing machine learning

1 Introduction

Coal is a natural fossil fuel source of energy for electricity generation, steel or iron production, and household practices. The characterization of coal quality plays a vital role in the suitable exploitation of coal. The quality of coal is usually followed by accomplishing arbitrary quality tests on the ultimate finished coal wedges before shipping to industries. Currently, industries evaluate the grade of coal using various conventional methods, namely chemical analysis, microscopical, and machine-based analysis established from prevailing expertise that is validated with ordinary samples [1]. Characterization using these conventional procedures are time-consuming, expensive, and also dependence on expert petrologists that may give subjectivity in the result. Chemical analysis results loss of concentrated coal during the process of raw ore crushing. It also causes severe air pollution because of the emission of chemical gases and the crushing of coal blocks into fine particles.

Hence, an automated coal characterization system would be preferred for industries which is much more intelligent than the traditional one and also environment-friendly. Generally, coal images can be categorized according to their structural, color, and textural characteristics. This research comes with an idea to classify the coal quality corresponding to its color and textural features utilizing image processing, pattern recognition techniques, and machine learning models.

2 Related works

In recent years, with the popularization of computer and high-performance imaging equipment, the method of image-based characterization is easy to be realized. It can be seen that microscopic images color, structures, and texture provides excellent differences in the analysis of coal, and these features have been frequently used to characterize them. Till date, the countless researcher has suggested the coal characterization techniques. Hong et al. [2] proposed a coal gauge sorting system using a convolutional neural network. Hao et al. [3] discussed the image-based fractal characterization of microfracture structure in coal using CT images. Qinghe et al. [4] studied pores in pulverized coal using the image analysis method. The feature engineering methods, such as the extraction of feature texture based on the gray level symbiotic matrix, [5, 6] is used to identify the coal gangue. Zhang et al. [7] suggested a genetic algorithm based on SVM for the prediction of ash content in coarse coal by image analysis. Mejia et al. [8] proposed the automated maceral characterization using histogram analysis of the color features of the images. Van et al. [9] modeled the dissemination of numerous coal constituents in the feature space utilizing a simple nearest neighbor and the hybrid model to categorize every pixel in the image. Jing et al. [10] discussed the coal analysis of topological and geometrical characteristics of the fracture system using X-ray images. Li et al. [11] suggested the quantitative characterization and visualization of coal fracture using micro CT scanning. Ren et al. [12] discussed the fractal characterization and pore structures of synclines coal. Joana et al. [13] discussed the geochemical and petrographical characterization of coal using the genetic geodynamic model. Chaves et al. [14 –16] proposed different methods for char characterization using image processing. There are many other methods introduced in literature for coal characterization [17 –19].

The existing schemes are satisfactory but unable to extract image features like color and textural features with better accuracy for coal quality characterization using an image processing method. Hence, in this paper, we intended at the development of a computer-based coal quality characterization system utilizing image processing and machine learning methods for subtyping the coal categories as per classes CL₁, CL_2, and CL_3, which are further named as good, average and poor quality of coal. The main objective of the proposed work is to develop a machine learning system using an ensemble-based algorithm for automated coal characterization. The motivation behind this research is to overcome the quality examination challenges confronted by the manufacturing and mining industries by offering the computer-based technique. The proposed scheme improves the results that can be acquired by examining the textural and color features of coal samples. Such computerized systems guarantee the reliability, cost-effectiveness, exactness, consistency in outcomes, and easiness. This study demonstrates the comparative performance analysis of the proposed model (ensemble of classifiers (EoC)) and standard classifiers (i.e., decision tree, linear discriminant, and support vector machine) for coal characterization. Consistent characterization results have been accomplished utilizing the noticeable textural, color features, and the ensemble of classifiers (EoC).

3 Ensemble learning-based coal characterization

The ensemble learning-based characterization of coal types from the optical microscopic images is precisely presented in Fig. 1.

Fig. 1

Ensemble coal characterization system.

These methods of automated coal characterization process have been concisely presented in the following sections.

3.1 Image capturing and preprocessing

The different types of coal sample images are collected from the CSIR-National Metallurgical Laboratory (NML), Jamshedpur, India. A Leica DM4500P optical light microscope has been used for capturing of coal images of size 1024*1024 pixels. It is assembled by monochrome camera and oil immersion lenses utilizing the green incident light. It is imitated under the polished surface. Here, the 500x magnification was utilized to capture the images. The captured image of high resolution provides good discrimination between different grades of coal. The overall one hundred twenty images have been captured from three different kinds of specified coal samples (i.e., CL₁, CL_2, and CL₃). All images have been manually labelled, i.e., assigned to one of three categories of coal, by petrologists, according to its characteristics like carbon content, ash content etc. The microscopic pictures of different coal samples are shown in Fig. 2.

Fig. 2

Different types of coal microscopic images.

The original coal images are comparatively bigger, so to characterize it appropriately, the method of sub-imaging is required. The chosen region of interest (ROI) should comprise repeated patterns of features. It is required because every coal sub-image sample in the original coal image has to be calculated for discriminating them into three classes of coal characteristics. The bounding-box [20] technique is used to crop the sub-image from the original sample image. The rectangular bounding-box method is used here. The rectangular area coordinates can be calculated by doing the mean of each pixel coordinates in that particular ROI area. Then, we provide the bounding box information like the top-left corner point and size of the crop window to crop the images. This method outcomes in sub-images comprising a distinctive solitary pattern with the hypothesis that there does not have any touching ranges or areas. The sub imaging stages are demonstrated in Fig. 3. A total of three hundred sixty sub-images containing a single coal texture pattern has been cropped and used as an input image for the coal characterization process.

Fig. 3

Sub imaging of original sample image.

Some of the noises appeared during the image capturing and cropping process. So, to avoid this problem, pre-processing techniques are used that are achieved by selective filtering and contrast enhancement methods [21].

3.2 HSV color and GLCM textural features extraction

The process of feature extractions is beneficial when the image is more extensive. In this procedure, some features of the sample are extracted with the help of standard feature descriptors. Three types of features are extracted for coal image-based characterization: textural, color, and morphological features. In this paper, color and textural features have been used for automated coal characterization.

Color is measured as a significant feature descriptor for the characterization of coal categories. If the specific area is exaggerated, then the coal category area transforms the color efficiently. Color histogram technique [22] is widely used in extracting color features. The restriction in using this method is that it does not consider the spatial arrangement of color. To overcome this issue, the HSV color spaces [23] have been used here for feature extraction. HSV comprises of an H-hue, S-saturation, and V-value (brightness). Hue correlates precisely to the idea of tone, whereas saturation relates to the concept of tint, and value tells about the intensity. The hue, brightness, and saturation (HSV) are the color model considered here as a feature vector. Figure 4 shows the HSV color space representation of three different types of coal samples.

Fig. 4

HSV color space representation of coal samples: (a) CL₁, (b) CL₂ and (c) CL₃.

The gray-level co-occurrence matrix (GLCM) is used for describing the luminance variation and spatial correlation of images [24]. It is a method of extricating Haralick’s textural features. This dual probability takes the structure of a square array GL_d, with two-dimensional matrix, rows, and columns, which is equivalent to several discrete gray levels (intensities) in the image being inspected. It can be represented as: ${GL}_{d} = [x, y] = n_{xy}$ (1)

Here, n_xy is occurrences number of pixels (x, y) present at a distance d in an image. The GL_d co-occurrence matrix is having dimension n×n, where, n is a number of grays level in the image. In this research, five features of GLCM (i.e., correlation, contrast, energy, entropy, and homogeneity) were extracted for each type of coals and are used as feature vectors.

Altogether, these three color and five textural feature vectors are utilized in calculating the moment values (i.e., mean, skewness variance, and kurtosis) for each of the sub-imaged samples for feature extraction and then, provided as an input to the classifier to evaluate the accuracy of the machine learning models for coal quality characterization. There is a total of seventeen features that have been extracted, out of which five are textural features, and thirteen are color features.

3.3 Normalization of featured data and features selection

Before feature selection and classification, the normalization of the featured data set, which has a diverse range of values and also, to evaluate the characterization ability of each feature or group of features in the labeled classes, is one of the mandatory steps for automated coal characterization. Hence, every input value has to be dispensed discretely and for every feature (X_i) in training set, mean (α), and standard deviation (γ) are calculated. The arithmetical expression to compute the normalized features is [25]: $Y = \frac{(X_{i} - α)}{γ}$ (2) where Y is transformed value and provides new values of the normalized feature with a mean value of zero and a standard deviation of value one.

Classifier performance is affected by selected features, so it is essential to select appropriate features before classification. Hence, this process is known as feature selection. One of the statistical methods used for feature selection is ANOVA [26] testing. The Analysis of Variance, popularly known as the ANOVA, can be used in cases where there are more than two groups to select the appropriate features. A total of ten features has been selected using the ANOVA method.

3.4 Classification

In a pattern recognition problem, classification is primarily utilized for distributing the feature space into numerous classes or categories depending upon similar features. Depending upon a number of categories, each feature vector is assigned a category label, which is a pre-defined integer number and is based upon the classifier outcome. Every classifier must be designed in such a way that the use of a group of inputs deliver the anticipated set of outcomes. The aggregate estimated data are partitioned into testing and training data set. The training data is utilized for updating the weights, and this process of network training is known as learning. The testing data has been utilized for validating the performance of the classifier. In this research, we suggest the utilization of the ensemble of the classifier for labeling each coal sub-image as CL₁, CL₂, CL₃ categories based on a group of estimated features. The performance of extracted color and textural features in classifications has also been examined with three other traditional classifiers, i.e., decision tree (DTC), linear discriminant (LD), and support vector machine (SVM). The appropriate parameter tuning has been done for every classifier to accomplish the optimal accuracy, and similar training and test data sets have been utilized for all while calculating their distinct classifier performance. Decision tree (DTC) [27] classification method has been used effectively for widespread varieties of the classification problem. The ‘tree’ consists of internal nodes, leaves, and the root node. Internal nodes split the dataset while leaves are called terminal nodes comprises of similar classes. The configuration of DTC includes fine tree type, with a hundred maximum splits using Gini’s diversity index. The main limitation with DTC is overfitting and also delivers less information on the connection between predictors and response. Linear Discriminant (LD) [28] classifier is a method utilized for data categorization and dimensionality deduction. The primary purpose of LD is to reserve most of the class discrimination facts without missing any information. The configuration of LD includes a full covariance structure. The small sample size issues are the main disadvantage of LD. Support vector machine (SVM) [29] used to examine patterns and data for adequate classification. SVM is a non-probabilistic classifier used to categorize new samples into corresponding similar categories using the hyperplane. The configuration of SVM includes a radial basis function as a kernel function with a five-kernel scale. The main drawback of SVM is that it has several vital constraints that need to be set appropriately to accomplish the optimal classification outcomes for any specified problems.

To overcome these issues, multi classifier methods or ensemble methods have been suggested here for better coal characterization. Multi classifier methods are more superior than single classifier equivalents because of several reasons: it diminishes the possibility of poor selection; complicated decision boundary has been learned efficiently; and beneficial with different features.

3.4.1 Ensemble classification for automated coal characterization

It is extremely required to sustain a high classification and less error percentage in an automated coal characterization system. Though, it is tough for a single machine learning classifier to accomplish this for a complex coal image characterization problem. To overcome these issues, an ensemble-of-classifiers (EoC) [30] based method for the classifications of extracted coal textural and color features are examined here. An ensemble classifier method, also called a multi classifier method, made by merging many miscellaneous classifiers. Miscellany may be accomplished by utilizing fully diverse groups of classifiers and training data set for every classifier. The concept behind this is that every member of the ensemble classifier will produce diverse decision boundaries and creates diverse error, and appropriate grouping of classifiers will decrease the overall error. Bagging is utilized here to promote diversity to train every member of the ensemble classifier utilizing arbitrarily depicted a subset of the training set. Class labels have been produced by distinct classifiers that are united utilizing majority votes, and class label selected by maximum classifiers is the ultimate ensemble classifier conclusion. The architecture of an ensemble of classifiers (EoC) for coal characterization is depicted in Fig. 5.

Fig. 5

Ensemble classifier architecture for coal characterization.

Here, in the primary phase, DTC has been used as source classifier, and flexible styles of DTC gained utilizing diverse parameters setting has been used to form an ensemble. After that, in the secondary phase, groups of classifiers having diverse topologies have been utilized for constructing the ensemble. Generally, the combination or ensemble of 3 distinct classifiers, i.e., DTC, LD, and SVM, has been discovered to execute best with available data and utilized for final coal characterization. Hence, only the investigational outcomes of the secondary phase of execution of an ensemble of classifiers with diverse topologies are shown in Table 1. The algorithm for ensemble classification for an automated coal characterization has been briefly discussed in few steps.

Table 1

Classification accuracy with standard classifier

Classifiers	Fold I	Fold II	Fold III	Fold IV	Fold V	Characterization accuracy %
DTC	72.34	83.01	82.62	88.62	88.62	83.04
LD	82.22	82.22	86.35	86.35	87.12	84.85
SVM	89.68	92.90	89.33	91.62	92.36	91.17
EoC	96.66	97.21	96.12	97.35	96.22	96.71

Algorithm:

Choose one classifier as source classifier.

Select a group of three or five classifiers having different topologies to construct an ensemble.

Train the data.

The majority of voting gives a final class label.

3.5 Cross-validation

The cross-validation technique [31] is used when there is not sufficient data present to generate the validation set. In this piece of research, we have used the k-fold cross-validation resampling process. In this process, the original sample set is divided via stratified split into i equivalent size disjoint subsets S1, S2, ... ... . Si, and after that i, runs of the training sample tests are executed. Taking into consideration the value of i as five, the entire sample sets are arbitrarily separated into five portions, through which all category is signified in a similar range as original sample sets. Hence, the process is executing an entire of five epochs having a different grouping of testing and training samples. Ultimately, the five operation approximations have been averaged to provide on the whole approximation of classifier recital in terms of computational time and accuracy.

4 Results

The projected idea has been executed using MATLAB 2015, and the investigational model has been processed utilizing an Intel(R) Core (TM) i7 with 3.20Ghz PC speed, through 64-bit, RAM of 24GB executed on the professional operating system of Windows 10. Coal quality is characterized based upon color and textural features present in the coal samples. The data are balanced. The measured feature value analysis demonstrates that the coal samples are distinguishable and appropriate classifier with better accuracy is needed for characterization. Considering this, an ensemble classifier system has been suggested for the coal quality characterization. Table 1 shows the average performance of classifiers based on accuracy utilizing five-fold cross-validation.

Here, the best overall accuracy of around 97% is achieved with the proposed ensemble classifier method for coal sub-image samples with a five-fold cross-validation method. In EoC, we can see that the accuracy is more than 90 % in each five-fold constantly, and the average of every fold was discovered to be higher than all other classifiers. The computational time (in a sec) that consists of training, as well as the testing stage, has been recorded for each of the mentioned classifiers and shown in Fig. 6.

Fig. 6

The computational time for characterization.

The average computation time for EoC was discovered to be a little bit higher than that of DTC, LD, and SVM but still less than the manual methods of industries. Hence, the recommended system could support in real mining, steel, and manufacturing industries for the screening of coal samples.

5 Model prediction

The prediction of the model in classification problem is one of the mandatory steps to recognize the feasibility of the proposed method. This can be done in various ways. In this paper, we use scatter graphical plot analysis, confusion matrix, and ROC curve for the prediction of the proposed model.

5.1 Scatter plot

The scatter plot [32] graphical analysis observes which variables separate the class colors most clearly. Here, blue points represent class CL₁, red points CL₂, and yellow CL_3, respectively. It has been observed that CL₁ is easily separable from the other two classes with all four classifiers. The CL₂ and CL₃ classes are much closer together in all classifier measurements. Here,×symbol indicates misclassified points. The blue points (CL₁) are all correctly classified, but some of the other two categories of coal are misclassified. Figure 7 shows the different classifier scatter plots.

Fig. 7

Scatter plot: (a) DTC, (b) LD, (c) SVM and (d) EoC.

5.2 Confusion matrix

The CM (confusion matrix) [33] chart has been used to recognize how the currently chosen classifier executed in every class. Here, rows illustrate correct classes, while columns illustrate the predicted classes. Also, diagonal cells illustrate where correct classes and predicted classes counterpart. If compartments are green, then the classifier has been performed excellently and categorized as the interpretations of this particular true class appropriately. The true positive rates for accurately categorized points displayed in the green cell in the True Positive Rate column, while the false-negative rate for wrongly categorized points displayed in the light orange cell in the False Negative Rate column. Figure 8 shows the confusion matrix obtained by different classifiers.

Fig. 8

Confusion Matrix: (a) DTC, (b) LD, (c) SVM and (d) EoC.

5.3 ROC curve

The ROC curve [34] demonstrates the false-positive rates versus true positive rates for the chosen trained classifier. An accurate outcome having no misclassified object becomes 90 degrees at the top left of the graph. A bad outcome shows a line or streak at 45 degrees. The AOC (Area Under Curve) number is a degree of the overall excellence of the classifier. Here, EoC shows the best result with AOC value 1. Figure 9 shows the ROC of different classifiers.

Fig. 9

ROC curve: (a) DTC, (b) LD, (c) SVM and (d) EoC.

This prediction of above methods predicted that the machine learning-based automated coal characterization shows better performance concerning the accuracy and computational time than existing conventional methods. So, we can say that the proposed intelligent characterization technique of coal microscopic images using ensemble learning can be feasible and recommended to industries for coal characterization.

6 Conclusion

Coal is the frequently used fossil fuel for steel, power industries, and also for household practices. It has unique properties and behaviors, which makes it necessary to characterize for quality inspection before shipping to industries. Petrologist’s screening of coal samples is always a slow and subjective process. The underlying theme of this research is to develop an automated screening method using an ensemble classifier system for the quality characterization of coal. The proposed multi classifier system has an accuracy of around 97% and also, its computational time is faster than conventional methods. Hence, the proposed system could be used in industries for the preliminary screening of coal samples and assists in the development of an automated coal characterization system.

Footnotes

Acknowledgment

The council of scientific and industrial research (CSIR) India, supports the research as mentioned above under Direct SRF. We would also like to thanks Dr. Vivek Mishra, CSIR-NML, Jamshedpur, India, and Dr. Subrajeet Mohapatra, BIT Mesra, Ranchi, India for his valuable comments and guidance. The authors would also like to thanks to anonymous reviewers for their valuable suggestions and comments.

References

Gupta

, Advanced coal characterization: A review, Energy & Fuels21 (2007), 451–460.

Hong

, Zheng

, Zhu

, Pan

and Zhou

, Automatic recognition of coal and gangue based on convolution neural network. https://arxiv.org/abs/1712.00720, 2017.

Hao

, Zhou

Yingfang

, Yao

Yanbin

and Wu

Kejian

, Imaged based fractal characterization of micro-fracture structure in coal, Fuel239 (2019), 53–62.

Niu

Qinghe

, Pan

Jienan

, Jin

, Wang

Haichao

, Li

Meng

, Ji

Zhongmin

, Wang

Kai

and Wang

Zhenzhi

, Fractal study of adsorption-pores in pulverized coals with various metamorphism degrees using N2 adsorption, X-ray scattering and image analysis methods, Journal of Petroleum Science and Engineering176 (2019), 584–593.

Liang

, Cheng

and Ma

, Identification of coal and gangue by self-organizing competitive neural network and SVM, International Conference on Intelligent Human-Machine Systems and Cybernetics. IEEE (2010), 41–45.

Min

H.E.

, Wang

P.P.

and Jiang

H.H.

, Recognition of coal and stone based on SVM and texture, Computer Engineering & Design33(3) (2012), 1117–1121.

Zhang

, Yang

, Wang

, Dou

and Xia

, Ash content prediction of coarse coal by image analysis and GA-SVM, Powder Technology268 (2014), 429–435.

Mejia

J.R.A.

, Mattosb

and Torres

C.O.

, Automated coal petrography for macerals characterization using histograms technique, in: Proc. of SPIE, 8th Iberoamerican Optics Meeting and 11th Latin American Meeting on Optics, Lasers, and Applications, 2013, 8785.

Van Vuuren

P.A.

, et al., Using visual texture analysis to classify raw coal components, International Conference on Systems, Signals and Image Processing (IWSSIP), 2015, 212–215.

10.

Jing

, Armstrong

Ryan T.

, Ramandi

Hamed Lamei

and Mostaghimi

Peyman

, DigiCoal: A computational package for characterisation of coal cores, Journal of Petroleum Science and Engineering176 (2019), 775–791.

11.

Yingying

, Cui

Hongqing

, Zhang

Ping

, Wang

Dengke

and Wei

Jianping

, Three-dimensional visualization and quantitative characterization of coal fracture dynamic evolution under uniaxial and triaxial compression based on μCT scanning, Fuel, 2019, https://doi.org/10.1016/j.fuel.2019.116568.

12.

Ren

Pengfei

, Xu

Hao

, Tang

Dazhen

, Li

Yukui

, Chen

Zhenlong

, Sun

Changhua

, Zhang

Fulin

, Chen

Shida

, Xin

Fudong

and Cao

Likun

, Pore structure and fractal characterization of main coal-bearing synclines in western Guizhou, China, Journal of Natural Gas Science and Engineering63 (2019), 58–69.

13.

Ribeiro

Joana

, Machado

Gil

, Moreira

Noel

, Suárez-Ruiz

Isabel

and Flores

Deolinda

, Petrographic and geochemical characterization of coal from Santa Susana Basin, Portugal, International Journal of Coal Geology203 (2019), 36–51.

14.

Chaves

, Robles

L.F.

, Bernal

, Alegre

and Trujillo

, Automatic characterisation of chars from the combustion of pulverised coals using machine vision. Ptec, 2018, doi: 10.1016/j.powtec.2018.06.035

15.

Chaves

, Trucco

, Barraza

and Trujillo

, An image processing system for char combustion reactivity characterisation, Computers in Industry106 (2019), 60–70.

16.

Chaves

, Trujillo

and Barraza

, Global and local features for char image classification, in: Artificial Computation in Biology and Medicine of Lecture Notes in Computer Science, Springer, Cham9107 (2015), 98–107.

17.

Mishra

Vivek

, Bhowmick

Trisrota

, Chakravarty

Sanchita

, Varma

Atul Kumar

and Sharma

Mamta

, Influence of coal quality on combustion behaviour and mineral phases transformations, Fuel186 (2016), 443–455.

18.

Alpana , Mohapatra

, Automated coal characterization using computational intelligence and image analysis techniques, Communication, Control and Intelligent Systems (CCIS), 2015, pp. 176–180, DOI: 10.1109/CCIn-telS.2015.7437903.

19.

Alpana , Mohapatra

, A Textural Characterization of Coal SEM Images Using Functional Link Artificial Neural Network. In: Raman B., Kumar S., Roy P., Sen D. (eds) Proceedings of International Conference on Computer Vision and Image Processing. Advances in Intelligent Systems and Computing, vol 459, 2017, Springer, Singapore, https://doi.org/10.1007/978-981-10-2104-6_11

20.

Russ

J.C.

, The Image Processing Handbook, 5th ed., Taylor and Francis, 2007.

21.

Srivastava

H.B.

, Kumar

, Verma

H.K.

and Sundaram

S.S.

, Sundaram, Image pre-processing algorithms for detection of small/point airborne targets, Defense Science Journal59(2) (2009), 166–174.

22.

Alamdar

and Keyvanpour

M.R.

, A new color feature extraction method based on dynamic color distribution entropy of neighborhoods, IJCSI Int. J. Comput. Sci. Issues 8, 2011.

23.

Alpana , Mohapatra

, Machine learning approach for automated coal characterization using scanned electron microscopic images, Computers in Industry75 (2016), 35–45.

24.

Claudio

, Pablo

and Estevez

, Ore grade estimation by feature selection and voting using boundary detection in digital image analysis, Int. J. Miner. Process.101 (2011), 28–36.

25.

Samarsinghe

, Neural networks for applied sciences and engineering, Auerbach, 2007.

26.

Gun

A.M.

, Dasgupta

and Gupta

M.K.

, Fundamental of Statistics, 5th ed., World Press, 2005.

27.

Otukei

J.R.

and Blaschke

, Land cover change assessment using decision trees, support vector machines and maximum likelihood classification algorithms, Int. J. Appl. Earth Obs. Geoinf.12 (2010), 27–31.

28.

Hemjot , Sharma.

, A refinement: Better classification of images using LDA in contrast with SURF and SVM for CBIR system, International Journal of Computer Applications (0975–8887)117(16) (2015), 31–33.

29.

Zhou

, Tang

, Yang

, Yan

, Lu

and Cen

, Support vector machine based online coal identification through advanced flame monitoring, Fuel117 (2014), 944–951.

30.

Mohapatra

, Patra

and Sanghmitra

, An ensemble classifier system for early diagnosis of acute lymphoblastic leukemia in blood microscopic images, Neural Comput & Applic, 2013, doi: 10.1007/s00521-013-1438-3.

31.

Kohavi

, A study of cross-validation and bootstrap for accuracy estimation and model selection, in: Proceedings of the 14th International Joint Conference on Artificial Intelligence117 (1995), 1137–1143.

32.

Cliche

, Rosenberg

, Madeka

and Yee

, Scatteract: Automated extraction of data from scatter plots. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, Springer (2017), pp. 135–150. doi:10.1007/978-3-319-71249-9_9.3

33.

Maria

N.J.R.

and Pankaj

, Performance Analysis of Text Classification Algorithms using Confusion Matrix, International Journal of Engineering and Technical Research (IJETR) ISSN: 2321-0869 (O) 2454-4698 (P), Volume-6, Issue-4, December, 2016.

34.

Sonego

, Kocsar

and Pongor

, ROC analysis: Applications to the classification of biological sequences and 3D structures, Briefings in Bioinformatics9(3) (2008), 198–209, https://doi.org/10.1093/bib/bbm064.