An ensemble learning paradigm for subsurface stratigraphy from sparse measurements and augmented training images

Research output: Journal Publications and ReviewsRGC 21 - Publication in refereed journalpeer-review

View graph of relations

Author(s)

  • Chao Shi
  • Yu Wang
  • Hao-Qing Yang

Detail(s)

Original languageEnglish
Article number105972
Journal / PublicationTunnelling and Underground Space Technology
Volume153
Online published23 Jul 2024
Publication statusPublished - Nov 2024

Abstract

The performance of computer vision-based techniques for stratigraphic modeling relies heavily on qualified training images to capture the complex stratigraphic connectivity. In geotechnical engineering, only limited training images are available for a specific site. Stochastic simulation modelling based on limited training data may be biased as the collected images that reflect prior geological knowledge may not encompass all potential stratigraphic patterns. Therefore, it is crucial to establish a high-quality, domain-specific training image database for effective stratigraphic modelling. In this study, an ensemble learning paradigm is proposed to tackle this issue and develop subsurface geological cross-sections from sparse data by reconstruction and redistribution of stratigraphic statistics revealed from limited training images. A domain-specific training image database is first established using generative adversarial networks (GAN) that enable the generation of arbitrary sized image samples from a single training image. Subsequently, multiple qualified image samples that are compatible with site-specific data are adaptively selected and utilized for the ensemble learning of geological cross-sections. The performance of the proposed framework is demonstrated using real geological cross-sections collected from a reclamation site and a tunnelling project in Hong Kong. The results indicate that the proposed method can effectively generate diverse image samples that encompass stratigraphic features beyond those reflected in a single training image. More importantly, the ensemble learning framework can capture the complex spatial stratigraphic connectivity of soil layers with enhanced prediction accuracy. © 2024 Elsevier Ltd.

Research Area(s)

  • Data augmentation, Domain-specific database, Generative modelling, Prior geological knowledge