A novel MRF-based approach is developed for stochastic stratigraphic simulation.
Bayesian machine learning is implemented for inferring modeling parameters.
A new technique called DANN-KHMD is developed to sample initial stratigraphic profile.
Both synthetic and real-world cases are studied to demonstrate the performance.
The proposed approach outperforms three existing techniques.
Stratigraphic modeling with quantified uncertainty is an open question in engineering geology. In this study, a novel stratigraphic stochastic simulation approach is developed by integrating a Markov random field (MRF) model and a discriminant adaptive nearest neighbor-based k-harmonic mean distance (DANN-KHMD) classifier into a Bayesian framework. The DANN-KHMD classifier is effective for extracting anisotropic patterns from sparse and heterogeneous spatial categorical data such as borehole logs. The MRF parameters can be initially estimated roughly or customized (if site-specific knowledge is available). Later these parameters can be updated and regularized in an unsupervised manner with constraints from site exploration results in a Bayesian manner. Throughout the learning process, both the soil profile and the MRF parameters are updated in a probabilistic manner. The advantages of the proposed approach can be summarized into four points: 1) inferring stratigraphic profile and associated uncertainty in an automatic and fully unsupervised manner; 2) reasonable initial stratigraphic configurations can be sampled and hence lower the computational cost; 3) both stratigraphic uncertainty and model uncertainty are taken into consideration throughout the inferential process; 4) relying on no training stratigraphy images. To illustrate the effectiveness of the developed approach, two synthetic cases and three real-world cases are studied and the advantages of the new approach over existing approaches are demonstrated.