Application Of Bayesian Inversion Of Electromagnetic Induction Data For Uxo Discrimination

Abstract

This paper presents the results of a study applying the Bayesian inversion approach to electromagnetic induction (EMI) data, for applications such as UXO discrimination. The cases investigated feature prominent impediments to simpler treatment, namely high signal clutter and multiple objects sensed simultaneously. The fundamental feature of Bayesian inversion is rational incorporation of prior information into a stochastic inference algorithm, to reach the most robust posterior probability of the model identity based on measured data. In UXO detection and classification, the model is a set of parameters corresponding to a particular object in a particular disposition. Prior information about the target or setting and the randomness of noise from different sources warrant the application of a Bayesian approach in this particular inverse problem. Broadband EMI responses at different locations, in terms of scattered magnetic field components in-phase and out-of-phase with the transmitted primary field, form the data vector. Undoubtedly, EMI measurements are contaminated with errors in sensor positioning, truncation errors in sensors and computers, metallic clutter items, ambient radio interference, etc. The prior information derives from sampling excavation at a particular site, soil information, historical information on use of a site, archival knowledge on different object types, forward modeling results for a particular type of UXO, and other pertinent information one can collect for a given UXO cleanup project. Compared to deterministic inversion algorithms, Bayesian inversion should be more advantageous for dealing with an inverse or inference problem when data are contaminated by random errors, as long as one can justify characterizing the prior information statistically. Two kinds of problem were solved here using the Bayesian approach: (1) data contaminated with random noise and (2) data for cases in which more than one UXO-sized object is in the sensor's field of view at the same time. For the first case, we applied the inversion algorithms on 100 sets of synthetic data; results were compared with that from simple least squares (SLS) algorithm. Comparison shows that Bayesian approach can give more accurate results, given that we can provide reasonable prior information and statistics on the noise. For the second case, we measured data for two cylinders at different distances from one another, with signals overlapping to one degree or another. Results show that in most cases the signatures of each individual contributing target can be extracted.