Subsurface Lithological Characterization Via Machine Learning-assisted Electrical Resistivity and SPT-N Modeling: A Case Study from Sabah, Malaysia

Recently, significant progress has been made in establishing relationships between geophysical and geotechnical datasets to evaluate subsurface geological characteristics, especially to address engineering infrastructure failures. This research, therefore, explores the efficiency of machine learning (ML) combined with descriptive statistics in optimizing geophysical and geotechnical datasets from electrical resistivity tomography (ERT) and standard penetration tests (SPT-N) for subsurface lithological characterization of the Kabota-Tawau area in Sabah, Malaysia. As the first report of such analysis in this area, the study aims to address infrastructure design challenges posed by the increasing needs of the growing population. The derived ERT models, coupled with k-means clustering and regression results for modeled resistivity–SPT-N data at varying depths, clearly depict the main lithologies of the study area. The lithological units include the topsoil, weathered soil units, highly weathered/fractured units, and relatively weathered/fractured units. The use of regression analysis enabled the identification of new statistical correlations between resistivity and SPT-N in the area. These correlations accurately predicted outcomes with a 77% accuracy rate, supported by highly efficient performance values from the descriptive statistics. The predictions were based on resistivity values specific to different lithologies, which were determined to be < 150 Ωm. This progress is useful for forecasting SPT-N and reducing survey expenses, especially in extensive regions earmarked for infrastructure projects. Despite the study area’s generally low resistivities, associated with low load-bearing capacity, the study suggests modifications for the placement of medium to heavy infrastructure weights in highly to relatively weathered units. However, piling to the fresh bedrock is advised for high-rise buildings of super weights. Based on the research findings, the established ML-assisted methodological approach can be applied to other terrains with similar geology, particularly for early-stage subsurface characterization.