Soil Stratification and Spatial Variability Estimated Using Sparse Modeling and Gaussian Random Field Theory

Abstract

We propose a method for simultaneously estimating the trend and random components of soil properties using the least absolute shrinkage and selection operator (LASSO) for sparse modeling without assuming any basis functions and Gaussian random field theory, respectively. The uncorrelated observation noise is also estimated at the same time. A two-step algorithm is introduced to avoid the shrinkage problem of LASSO. Numerical examples with random realizations show that the method avoids shrinkage. The proposed method requires four parameters, namely, the variances of the random component and observation error, autocovariance distance, and regularization parameter, for one-dimensional problems. We propose a method that uses Akaike’s information criterion or Bayesian information criterion and particle swarm optimization to determine these four parameters. It is shown that the detection ratio of the layer boundary is determined by the number of observation data and the difference between trend values. The proposed method is applied to actual cone penetration test data to estimate the trend component of the soil property.