Stochastic Modelling of Subsurface Stratigraphy from Sparse Measurements and Augmented Training Images

Exploration and utilization of urban underground space require a sound understanding of subsurface stratigraphy. Image-based machine learning algorithms are appealing to engineering practitioners as they can effectively leverage valuable geological knowledge reflected in training images for stochastic simulations. However, in practice, only limited training images are available, which may not exhaust all the potential stratigraphic patterns at the site of interest. To explicitly tackle this dilemma, in this study, a generative adversarial network (GAN) is employed to generate multiple random image samples from a single training image. The compatibility of generated image samples with site-specific data is ranked based on total information entropy, and the selected image samples can be used to derive a robustness index for adaptive specification of the optimal next sampling location. The performance of the method is demonstrated using real examples from a Hong Kong reclamation site. Results indicate that the proposed framework can efficiently generate multiple image samples with plausible geological patterns, and the adaptively selected training images can improve the stochastic prediction performance.