Enhancing geological data interpolation by integrating convolutional neural networks (CNN) with the Ordinary Kriging (OK) method: a case study of Wuhan

Rapid urbanization has driven a significant increase in demand for underground infrastructure, including transportation systems, utilities and sewage networks. However, the lack of a comprehensive, reliable subsurface geological model has posed challenges for engineers and urban planners. This work aims to address this issue by improving urban planning and risk assessment through access to a robust geological model. We develop predictive geological models by integrating a modified Ordinary Kriging (OK) approach to build a layered geological model that accounts for subsurface unconformity. To address the uneven spatial distribution of lithological drill data, we developed a framework using Convolutional Neural Networks (CNNs) to predict geological features in data-sparse regions. The results enhance urban planning by providing a clearer understanding of subsurface conditions and reducing risks associated with underground infrastructure projects. A dataset of 2340 lithological boreholes covering 16 different geological strata was cleaned and integrated. These models support underground infrastructure development by reducing uncertainty.