open access

https://hdl.handle.net/2115/100476

Integrated Stratigraphic and Soil Property Modeling Using Potts Model and Gaussian Process Regression

英語

ground investigation

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

This study examines the applicability of a Potts-model-based layered Gaussian process regression (GPR) framework for three-dimensional subsurface characterization. Ground properties are notoriously challenging to map due to their sparse, uncertain, and spatially variable nature (the MUSIC-3X paradigm). The proposed framework first utilizes an anisotropic Potts-model segmentation to assign discrete stratigraphic labels to grid cells, then estimates continuous geotechnical properties within each assigned label using GPR. Its performance is benchmarked against three reference approaches: a naive layer-mean model, a fifth-order sparse polynomial model, and an unlayered GPR model. Two datasets are considered: the S-VG5 benchmark dataset based on synthetic cone penetration test data with complex heterogeneity, and the Naruohama boring dataset, evaluated via leave-one-source-out cross-validation utilizing SPT N-values. In the data-rich S-VG5 benchmark problem, the layered GPR achieved the best prediction accuracy (RMSE = 1.340), substantially outperforming the unlayered GPR (2.338) and the polynomial regression (2.787). However, in the highly sparse Naruohama field dataset, the naive model yielded the smallest mean cross-validation error. This phenomenon highlights a fundamental bias-variance tradeoff under extreme data starvation: while layered GPR strictly honors geological boundaries, severe intra-layer data sparsity drastically inflates predictive variance, whereas simple mean-based models remain robust. These findings indicate that Potts-model-based layered GPR represents a theoretical optimum when sufficient observations are available within stratigraphic units, yet simpler, high-bias estimators remain highly competitive under the severe sparsity typical of practical engineering conditions.

会議発表論文

Version of Record