Modified Bayesian algorithm implemented in compressive sensing applied to spatially sampled GPR measurement under high clutter conditions

Abstract

This paper investigates the implementation of compressive sensing (CS) for stepped frequency continuous wave ground penetrating radar (SFCW GPR) imaging system. Previous works in this field mostly focus on reducing the frequency samples in the measurement. While this approach enables faster scanning speed, we consider reducing spatial sampling is more efficient in reducing the data acquisition time in the GPR survey over a very large area. In this study we propose a data acquisition method and CS algorithm. A two-step sampling scheme is presented. In the first step, the spatial sampling was directly conducted in the data acquisition process. In the second step, the frequency sampling was conducted offline during the signal processing. Full frequency information was used in pre-processing to suppress the noise and clutter in the experiment data. To solve the sparse-data problem, some CS algorithms are compared and a modified Bayesian approach based on fast relevance vector machine (RVM) is proposed. The performance of the proposed CS-GPR system is analyzed using the real experimental GPR data set which contains non ideal conditions, e.g. high level clutter, not truly sparse targets, and inaccurate estimation of wave velocity in the medium. Using the proposed data acquisition and CS algorithm, even with these non ideal conditions, CS can give clear and stable results with high probability detection of the target.