Probabilistic soil property prediction using multi-fidelity Gaussian process regression and multiple imputation by chained equations

This study proposes a comprehensive framework to address two key benchmark problems in probabilistic site characterization, using a large-scale geotechnical dataset from an offshore airport. The first problem focuses on predicting the spatial variation of undrained shear strength. For this, we introduce a multi-fidelity Gaussian Process Regression (MF-GPR) approach that integrates sparse, high-fidelity site-specific measurements with abundant but less precise low-fidelity regional data (Big Indirect Data). This demonstrates a strategy to enhance predictive models by systematically leveraging information across different scales. The second problem tackles the estimation of missing mechanical parameters. We employ Multiple Imputation by Chained Equations (MICE), an iterative method that effectively handles complex missing data patterns by capitalizing on the statistical correlations between different soil properties. By applying these two methodologies, this work presents a robust baseline for the GEOAI benchmarks, showcasing how advanced statistical techniques can maximize data utility, handle data heterogeneity and sparsity, and pave the way for more reliable geotechnical site characterization.