Three-dimensional geotechnical-layer mapping in Seoul using borehole database and deep neural network-based model

Despite the increase in demand for three-dimensional (3D) subsurface geology maps, the attainment of reliable geotechnical characterization of specific regions and development of corresponding optimization models remains a challenging task. This can be attributed to the large volume of geotechnical survey data and uncertainties associated with the use of relevant stochastic and geostatistical approaches. This paper presents the design of an optimum deep neural network (DNN)-based learning model for reliable classification of geotechnical layers of the 3D underground space in Seoul, South Korea. A large-scale borehole database was created to this end using geotechnical-layer depths and 3D spatial coordinates of a borehole dataset for Seoul. It is based on stepwise outlier detection and has been preprocessed via correction and normalization of missing values. The best-fitting model was obtained by optimizing the hyperparameters of a DNN-based classifier, and its performance was evaluated in terms of precision and accuracy. Subsequently, a 3D grid network was established to facilitate local geotechnical-layer classification, and its performance was evaluated by applying the proposed DNN-based model for each unit lattice. The accuracy of the resulting 3D geotechnical-layer map was validated via a comparison against the corresponding thematically classified maps that depict the two-dimensional (2D) density of gravitational anomalies, 2D topsoil properties, and 3D geotechnical-layer classification based on geostatistical conditional simulations.