Application of Artificial Intelligence Techniques in Calibration of Ground Penetrating Radar

Abstract

Ground penetrating radar(GPR),an instrument that works by the principle of electromagnetic wave propagation,is of great use for in-situ ground analysis. Its potential uses in the field of geotechnical engineering are not fully explored with sufficient vigour,probably due to economic reasons and highly heterogeneous nature of soil. Taking this into account, if the instrument can be calibrated specifically for geotechnical purposes, it can reduce the number of experimental analyses that needs to be carried on soil and can also reduce the necessity of undisturbed samples of soil to a good extend. The present study discusses the calibration of GPR, through clay content of soil with some of its properties like particle density, bulk density, porosity, void ratio with the dielectric permittivity and peak wave frequency which are obtained from GPR data. Based on the database available in literature, the detection of clay content in three different grades of soil A1,A2 and A3, classified as per American Association of State Highway and Transportation Officials(AASHTO) are done using some recently developed AI techniques, multigene genetic programming (MGGP) and feature selection using NSGA-II, multivariate adaptive regression splines (MARS), Gaussian process regression (GPR)and Support Vector Machine (SVM). The predictive capabilities of these methods are discussed through coefficient of correlation(R2), Nash Sutecliffe coefficient of efficiency (E) root mean square
error(RMSE), average absolute error (AAE) and maximum absolute error(MAE). Based on above statistical parameters, the methods were ranked, for best fit, minimum error and least deviation of mean of actual to predicted values from 1. A combined rank was taken based on these three. It was observed that NSGA –II algorithm trained using ANN is showing better capability for predicting the clay content, with R2 value of 0.97 for both test and train data the least RMSE of 0.77 and 0.65 for test and train data, respectively.Though,the number of data points are limited, the AI methods used could predict the values with very good accuracy.