Dual-Modal Feature Learning Approach for Probabilistic Semisupervised Interpretation of Subsurface Stratigraphy from Sparse Boreholes and Cone Penetration Tests

Probabilistic interpretation of subsurface stratigraphy from sparse boreholes and cone penetration tests (CPTs) remains a critical yet nontrivial task in geotechnical site characterization. The challenge primarily arises from three key factors: (1) the modality gap between categorical borehole data and continuous CPT soundings; (2) the noisy and oscillatory nature of CPT profiles; and (3) the spatial variability of geological settings. This study proposes a dual-modal feature learning approach for semisupervised learning of subsurface stratigraphy from sparse site-specific data with quantified uncertainty. First, a nonparametric spatial interpolator is employed to stochastically interpret sparse boreholes into categorical feature profiles, offering enriched prior spatial stratigraphic context around CPTs. Each CPT profile is segmented into multiple unlabeled segments using change point detection, and a semisupervised learning approach was developed to sequentially filter and map stratigraphic patterns extracted from the prior categorical feature profiles onto the unlabeled CPT segments in a physics-informed manner. Subsequently, the obtained stratified CPTs are integrated with sparse boreholes to generate the most probable geological cross section with quantified stratigraphic uncertainty. Application to a complex real-world reclamation site demonstrated that the approach can not only accurately stratify noisy and oscillatory CPT data with site-specific soil types but also capture intricate geological variations without relying on predefined spatial correlation functions.