MACHINE LEARNING-BASED ESTIMATION OF THE GEOLOGICAL STRENGTH INDEX FROM ROCK MASS IMAGES: AN EXPLAINABLE AI APPROACH

Abstract

The Geological Strength Index (GSI) is widely employed to characterize jointed rock masses and provide essential input parameters for the design and numerical modelling of rock engineering structures. However, conventional chart-based GSI assessment relies heavily on visual judgment, making the estimated values susceptible to subjectivity and inter-practitioner variability. This study proposes a data-driven, image-based framework for automated and objective estimation of GSI from rock-face photographs. A dataset comprising 200 rock-mass images, covering a GSI range of approximately 15–85, was standardized to 256 × 256 pixels. A total of 59 colour, texture, edge, and fractal descriptors were extracted to quantitatively capture rock-mass characteristics. A two-stage Pearson correlation-based feature-selection strategy reduced these descriptors to approximately 19 non-redundant features, which were subsequently used to develop Support Vector Regression (SVR), Random Forest (RF), and CatBoost models. Hyperparameters were optimized using grid search with five-fold cross-validation. Evaluation on an independent test set of 40 images demonstrated that SVR achieved the highest predictive performance, with an R² of 0.655, RMSE of 12.56, and MAE of 9.22, followed closely by CatBoost (R² = 0.650) and RF (R² = 0.600). Model interpretability through SHAP and permutation-importance analyses consistently identified the standard deviation of GLCM correlation as the most influential predictor, complemented by edge density and colour-based descriptors. These dominant image features are physically consistent with the blockiness and surface-condition components underlying conventional GSI assessment. The proposed framework provides a reproducible and interpretable image-based approach for reducing subjectivity in GSI estimation and demonstrates the potential of machine learning for quantitative rock-mass characterization.