Site Characterization Model Using Artificial Neural Network and Kriging

In this paper, the problem of site characterization is treated as a task of function approximation of the large existing data from standard penetration tests (SPTs) in three-dimensional subsurface of Bangalore, India. More than 2,700 field SPT values $\left(N\right)$ has been collected from 766 boreholes spread over an area of $220{\text{-km}}^2$ area in Bangalore, India. To get $N$ corrected value $\left(N_c\right)$ , $N$ values have been corrected for different parameters such as overburden stress, size of borehole, type of sampler, length of connected rod. In three-dimensional analysis, the function $N_c=N_c\left(X,Y,Z\right)$ , where $X$ , $Y$ , and $Z$ are the coordinates of a point corresponds to $N_c$ value, is to be approximated with which $N_c$ value at any half-space point in Bangalore, India can be determined. An attempt has been made to develop artificial neural network (ANN) model using multilayer perceptrons that are trained with Levenberg-Marquardt back-propagation algorithm. Also, a geostatistical model based on ordinary kriging technique has been adopted. The knowledge of the semivariogram of the $N_c$ values is used in the ordinary kriging method to predict the $N_c$ values at any point in the subsurface of Bangalore, India where field measurements are not available. The results obtained show that ANN model is fairly accurate in predicting $N_c$ values. In case of ordinary kriging, a new type of cross-validation analysis shows that it is a robust model for prediction of $N_c$ values. A comparison between the ANN and geostatistical model demonstrates that the ANN model is superior to Geostatistical model in predicting $N_c$ values in the subsurface of Bangalore, India.