Use of Tree-Based Machine Learning Methods for Stratigraphic Classification in 3D Geological Modelling

Abstract

The need for underground development has been expanded due to rapid urbanization and to reduce the negative impact on city living. In recent years, 3D geological modelling has been widely used by engineers and geo-scientists for desk study and has shown its capability of integrating geological and geotechnical information for better usage in building and civil engineering projects. To provide techniques and instruction(?), a 3D Geo-data Modelling and Management System (GeM2S) has been established to better understand the subsurface conditions in Singapore. However, the geological stratum between the existing boreholes is often not investigated, which brings the possibility of vital errors in underground design due to inaccurate interpretation of the ground conditions. It is desirable to utilize advanced machine learning methods to predict and update the 3D geo-models. Machine learning is regarded as a subset in the field of artificial intelligence, which has shown its rapid development recently. However, the accuracy of the model is one of the major concerns when the techniques are applied. In this study, four machine learning models are proposed to provide the solutions for stratigraphic classification, namely, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (Catboost). The borehole data are located in the Bishan planning region in Singapore. The voxel-based input data for the models are spatial coordinates (X, Y, Z) and ground surface elevation. The prediction results demonstrate that LightGBM with optimization have produced the highest performance in this multi-classification problems. Finally, based on the prediction results of LightGBM, the voxel-based 3D geological models and the selected cross-sections are built for further analysis and comparison.