MS-13-083-cd banner.jpg Study on Autocorrelation Model for Spatial Distribution of Soil Properties using Gaussian Process Regression Y. Tomizawa1, I. Yoshida2 and Y.Otake3 1School of Integrative Science and Engineering, Tokyo City University, Japan 2Department of Urban and Civil Engineering, Tokyo City University, Japan 3Department of Civil Environmental Engineering, Tohoku University, Japan ABSTRACT

The spatial distribution of soil properties is often modeled by dividing it into trend and random components. Gaussian process regression with multiple stochastic fields is used to separate and model them in this study. The random component is important for understanding the characteristics of local variability in the spatial distribution. Autocorrelation models for the random components are examined based on 1D and 3D cone penetration test data. We compared various autocorrelation functions such as Gaussian, Markovian, Binary Noise, Whittle-Matérn (WM), and Gaussian-Markovian geometric mean (GMGM) model in terms of Akaike and Bayesian information criterion, AIC and BIC. As the result of comparison of the information criteria, GMGM and WM model are selected for the model site. GMGM is a very simple model and requires only exponential function, while WM requires special functions such as gamma and Bessel functions. From the viewpoint of computational cost, GMGM has advantage compared with WM. Spatial distribution of trend and random components in 1D and 3D are estimated by using GMGM model for the autocorrelation function of the random component.

Keywords: Soil Properties, Gaussian Process Regression, Random Components, Autocorrelation Function.

 pdflogo.jpg MS-13-083