Comparative stratigraphic classification using conventional hand sketching, 3D geological, and AI-based techniques
Journal
Results in Engineering
Journal Volume
30
Start Page
111103
ISSN
25901230
Date Issued
2026-06
Author(s)
Abstract
Accurate stratigraphic classification is essential in geotechnical engineering to ensure safe and efficient design solutions. Geological modeling typically involves interpreting subsurface features from sparse borehole data, which is challenging because the data are multivariate, incomplete, and often uncertain. This study compares three approaches for developing soil profiles in the eastern Taipei Basin: (1) a conventional hand sketching method, (2) a kriging-based three dimensional (3D) geological model, and (3) an Artificial Interlligance (AI)-based model using shallow Machine Learning (ML) classifiers, with Random Forest (RF) selected as the best model. The AI-based models were trained using top and bottom elevations and plan coordinates to predict soil profiles at the study site. RF hyperparameters were refined using Randomized Search Cross Validation (CV) and Bayesian optimization, and predictive uncertainty was quantified using Shannon entropy. Model performance was assessed using (i) classification metrics and (ii) stratigraphic-profile comparison metrics that capture stratum accuracy, thickness agreement, and sequence similarity using ten selected boreholes. Three scenarios with varying numbers of boreholes were conducted to evaluate the accuracy of the models. The AI-based method achieved an average accuracy of 0.84, outperforming conventional hand sketching (0.79 average accuracy) and 3D geological modeling (0.81 average accuracy). Under reduced borehole density (40% data removal), the proposed method maintained an accuracy of 80%, whereas the other methods showed a significant decline. The results of the evaluation demonstrate that the AI-based model provides a robust, data-driven, and objective framework for stratigraphic classification, enhancing reproducibility and enabling quantifiable uncertainty estimation.
Subjects
3-D geological model
Conventional hand sketching method
Machine learning
Random forest
Stratigraphic classification
Subsurface profile
Publisher
Elsevier B.V.
Type
journal article