Random forest-based lithology identification using geophysical well logging data along a borehole in a coastal aquifer in Suruga Bay, Japan

Abstract

Abstract Machine learning techniques have attracted attention for lithology identification, which is fundamental for understanding geological structures. However, the efficacy of machine learning-based lithology identification, specifically in deep (several hundred meters) coastal aquifers composed predominantly of unconsolidated sediments, remains underexplored. Therefore, this study developed machine learning models based on decision trees, random forest (RF), logistic regression, and multilayer perceptron (MLP) algorithms that classify lithofacies using geophysical well logging data collected along a borehole located in Suruga Bay, Shizuoka Prefecture, Japan. Two to nine input features were selected for the seven cases to classify four lithofacies: “cobbles and gravel”, “sand with gravel”, “silt with gravel”, and “basaltic lava”. The models showed accuracy ranging from 73.9% to 90.4%, with the RF- and MLP-based models achieving accuracy of 90.4% and 90.1%, respectively, when trained using all available input features. This indicates their effectiveness for lithology identification in the target borehole. Feature importance analysis using the SHapley Additive exPlanations method suggested that natural gamma ray, widely used in groundwater exploration, served as a key input feature in many of the analyzed cases. The incorporation of other input features (i.e., electrical resistivity, P-wave velocity, density, and NMR-derived parameters) generally improved the prediction performance. The results demonstrate the applicability of machine learning algorithms for lithology identification in a coastal aquifer several hundred meters thick, consisting of unconsolidated alluvial sediments and lava flows. These findings suggest that machine-learning techniques have the potential to reduce the time, labor, and costs associated with hydrogeological surveys and groundwater exploration.