Support vector machine as an alternative method for lithology classification of crystalline rocks

• Published in: Journal of geophysics and engineering 2017-03, Vol.14 (2), p.341-349
• Main Authors: Deng, Chengxiang, Pan, Heping, Fang, Sinan, Konaté, Ahmed Amara, Qin, Ruidong
• Format: Article
• Language: English
• Subjects: Borderline-SMOTE; CCSD-MH; imbalanced data sets; lithology classification; SVM
• ISSN: 1742-2132; 1742-2140
• DOI: 10.1088/1742-2140/aa5b5b

With the expansion of machine learning algorithms, automatic lithology classification that uses well logging data is becoming significant in formation evaluation and reservoir characterization. In fact, the complicated composition and structural variations of metamorphic rocks result in more nonlinear features in well logging data and elevate requirements to algorithms. Herein, the application of the support vector machine (SVM) in classifying crystalline rocks from Chinese Continental Scientific Drilling Main Hole (CCSD-MH) data was reported. We found that the SVM performs poorly on the lithology classification of crystalline rocks when training samples are imbalanced. The fact is that training samples are generally limited and imbalanced as cores cannot be obtained balanced and at 100 percent. In this paper, we introduced the synthetic minority over-sampling technique (SMOTE) and Borderline-SMOTE to deal with imbalanced data. After experiments generating different quantities of training samples by SMOTE and Borderline-SMOTE, the most suitable classifier was selected to overcome the disadvantage of the SVM. Then, the popular supervised classifier back-propagation neural networks (BPNN), which has been proved competent for lithology classification of crystalline rocks in previous studies, was compared to evaluate the performance of the SVM. Results show that Borderline-SMOTE can improve the SVM with substantially increased accuracy even for minority classes in a reasonable manner, while the SVM outperforms BPNN in aspects of lithology prediction and CCSD-MH data generalization. We demonstrate the potential of the SVM as an alternative to current methods for lithology identification of crystalline rocks.