Geological Uncertainty Quantification Using Image Warping and Bayesian Machine Learning

Hui Wang, Xingxing Wei8th 2022 · DOI

Abstract

This paper introduces a novel "warp-and-simulate" framework designed to quantify geological uncertainty in complex, non-stationary stratigraphic fields. By integrating Thin Plate Spline (TPS) image warping with Bayesian Machine Learning, the method transforms distorted soil layers into a stationary "standard space" for simplified analysis. Using a synthetic dataset with sparse borehole information, the approach demonstrates high accuracy in stratigraphic prediction and provides a robust measure of uncertainty through information entropy.

Introduction

In geotechnical engineering, accurate subsurface characterization is critical for safety and design. However, natural geological processes like tectonic shifts and irregular deposition create "non-stationary" fields—complex, wavy, or distorted strata. Traditional stochastic models, such as Markov Random Fields (MRFs), often struggle with these complexities because they typically assume stationarity, where spatial patterns are uniform across an area. Furthermore, borehole data is usually sparse, making it difficult to quantify the uncertainty inherent in geological models. This research aims to bridge this gap by creating a model that handles non-linear geological distortions while providing a rigorous measure of uncertainty.

Methodology

The proposed multi-step pipeline essentially "straightens" the earth before performing spatial analysis. The process is summarized in the following flowchart:

Figure 1: Flowchart of the proposed multi-step geological modeling and uncertainty quantification pipeline.
  1. Image Warping (Thin Plate Splines): Using known borehole data and geological prior knowledge, a Thin Plate Spline (TPS) algorithm transforms a distorted, non-stationary geological field into a simplified, stationary "standard space" where layers appear flat.
  2. DANN-KHMD Classifier: To address sparse data, a Discriminant Adaptive Nearest Neighbor-based K-Harmonic Mean Distance classifier is used to provide initial soil label estimates for unknown pixels based on the nearest borehole data.
  3. Bayesian Markov Random Field (MRF): In the simplified standard space, an MRF model characterizes spatial correlations. A Bayesian framework using Markov Chain Monte Carlo (MCMC) simulation infers "granularity coefficients" ($\beta$), describing how soil types cluster.
  4. Inverse Warping: Once the simulation is complete in the standard space, the results are warped back into the original, distorted coordinate system to produce the final stratigraphic map.
  5. Uncertainty Quantification: The model calculates Information Entropy (IE) for every pixel. High entropy indicates areas where the model is less certain of the soil type, typically occurring at layer boundaries.

The Warp-and-Simulate Concept

The core innovation of this research is the "Standard Space" shortcut. Rather than developing complex mathematical models to describe wavy soil, the framework changes the coordinate system to make the soil appear simple. Standard simulations are then run, and the results are bent back to the original configuration.

Figure 2: The deformation process of soil profiles, illustrating the transformation between stationary and non-stationary spaces.

By using the DANN-KHMD classifier, the model adaptively changes its metric based on local data density, which is crucial when borehole data is sparse. The resulting Information Entropy maps provide a heat map of doubt, allowing engineers to identify where additional site investigations (e.g., more boreholes) are most needed.

Results and Discussion

The study validated the approach using a synthetic 100x100 pixel soil profile grid representing three soil labels (A, B, and C) arranged in five sinusoidal layers. Four virtual boreholes were extracted, representing only 4% of the total pixels.

  • High Accuracy: Despite the limited data, the Maximum A Posteriori (MAP) estimation achieved an accuracy of 89.76% compared to the ground truth.
  • Robust Convergence: The MCMC simulation for granularity coefficients showed stable convergence after approximately 1,200 iterations.
  • Boundary Sensitivity: Information Entropy maps successfully identified that uncertainty is highest at the interfaces between soil layers, aligning with geological intuition.

Conclusion

The "warp-and-simulate" framework offers a powerful tool for geotechnical engineers to map complex strata and quantify uncertainty. While the synthetic case is promising, the technique's dependency on prior knowledge for setting control points remains a limitation. Future work will focus on scaling the logic to 3D volumes and validating the approach against complex, real-world site investigation data.

Summary updated

BibTeX
@inproceedings{wang2022geological,
  author = {Hui Wang and Xingxing Wei},
  title = {{Geological Uncertainty Quantification Using Image Warping and Bayesian Machine Learning}},
  booktitle = {8th},
  year = {2022},
  doi = {10.3850/978-981-18-5182-7\_00-05-014.xml},
}
Figure