A machine learning-based approach for constructing a 3D apparent geological model using multi-resistivity data

• Published in: Geoscience letters Vol. 11; no. 1; pp. 54 - 23
• Main Authors: Puntu, Jordi Mahardika, Chang, Ping-Yu, Amania, Haiyina Hasbia, Lin, Ding-Jiun, Suryantara, M. Syahdan Akbar, Tsai, Jui-Pin, Yu, Hwa-Lung, Chang, Liang-Cheng, Zeng, Jun-Ru, Kassie, Lingerew Nebere
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.12.2024; Springer Nature B.V; SpringerOpen
• Subjects: 3D apparent geological model; 3D resistivity model; Algorithms; Alluvial fans; Atmospheric Sciences; Biogeosciences; Boreholes; Choushui river alluvial fan; Clay; Coastal zone; Datasets; Decision trees; Earth and Environmental Science; Earth Sciences; Electrical resistivity; Geology; Geophysics/Geodesy; Gravel; Learning algorithms; Machine learning; Matrix methods; Modelling; Oceanography; Performance evaluation; Planetology; Python; Python-based modeling; Radial basis function; Research Letter; Sand; Statistical methods; Statistical models; Supervised learning; Supervised machine learning; Support vector machines; Python-based modeling; Choushui river alluvial fan; 3D resistivity model; Supervised machine learning; 3D apparent geological model
• ISSN: 2196-4092; 2196-4092
• DOI: 10.1186/s40562-024-00368-0

This study presents a comprehensive approach for constructing a 3D Apparent Geological Model (AGM) by integrating multi-resistivity data using statistical methods, supervised machine learning (SML), and Python-based modeling techniques. Demonstrated through a case study in the Choushui River Alluvial Fan (CRAF) in Taiwan, the methodology enhances data coverage significantly, from 62 to 386 points, by incorporating resistivity data sets from Vertical Electrical Sounding (VES), Transient Electromagnetic (TEM), and borehole information. A key contribution of this work is the rigorous harmonization of these data sets, ensuring consistent resistivity values across different methods before constructing the 3D resistivity model, addressing a gap in previous studies that typically handled these data sets separately, either building models individually or comparing results side-by-side without fully integrating the data. Furthermore, python-based modeling and radial basis function interpolation were employed to construct the 3D resistivity model for greater flexibility and effectiveness than conventional software. Subsequently, this model was transformed into a 3D AGM using the SML technique. Four algorithms, namely, random forest (RF), decision tree (DT), support vector machine (SVM), and extreme gradient boosting (XGBoost) were implemented. Following evaluation via confusion matrix analysis, evaluation metrics, and examination of receiver operating characteristics curve, it emerged that the RF algorithm exhibits superior performance when applied to our multi-resistivity data set. The results from the 3D AGM unveil distinct resistivity anomalies correlated with sediment types. The clay layer exhibited low resistivity (≤ 59.98 Ωm), while the sand layer displayed medium resistivity (59.98 <  ρ  < 136.14 Ωm), and the gravel layer is characterized by high resistivity ( ≥ 136.14 Ωm). Notably, in the proximal fan, gravel layers predominate, whereas the middle fan primarily consists of sandy clay layers. Conversely, the distal fan, located in the western coastal area, predominantly comprises clayey sand. To conclude, the findings of this study provide valuable insights for researchers to construct the 3D AGM from the resistivity data, applicable not only to the CRAF but also to other target areas.