Transparent City-Scale 3D Engineering Geological Modeling Using Deep Learning with Bayesian Local Refinement

Abstract Urban expansion creates complex underground transportation systems, buried lifeline systems, and structural foundations. Reliable and transparent city-scale three-dimensional (3D) geological stratigraphic information is essential for urban construction, risk assessment, and underground facilities management. A key challenge is to leverage large, heterogeneous geological data to achieve city-scale geological modeling while allowing efficient project-scale refinement. This study proposes a deep learning-empowered, three-stage framework with Bayesian local updating to develop the first 10 m-resolution city-scale 3D engineering geological model for Hong Kong, using a multisource and multimodal database containing more than 100,000 boreholes. It predicts key geological interfaces and assembles them into a 3D model, explicitly accounting for spatial correlation, coastline boundary constraints, and fault-related discontinuities. It achieves Pearson correlation coefficient > 0.95 for stratigraphic interface depth prediction and macro F 1 score > 0.9 for stratigraphic condition classification, outperforming conventional methods. The model reveals pronounced city-scale stratigraphic variability. At the project scale, the city-scale model is used as a baseline prior and updated with site-specific data to generate a high-resolution geological model, capturing more stratigraphic units and finer stratigraphic details while reducing the uncertainty and outperforming models without prior information. The proposed framework supports city-scale strategic planning, hazard evaluation, and project-level engineering design.