The use of ground penetrating radar (GPR) for detecting the presence of near-surface interfaces is a scenario of special interest to the underground coal mining industry. The problem is difficult to solve in practice because the radar echo from the near-surface interface is often dominated by unwanted components such as antenna crosstalk and ringing, ground-bounce effects, clutter, and severe attenuation. These nuisance components are also highly sensitive to subtle variations in ground conditions, rendering the application of standard signal pre-processing techniques such as background subtraction largely ineffective in the unsupervised case. As a solution to this detection problem, we develop a novel pattern recognition-based algorithm which utilizes a neural network to classify features derived from the bispectrum of 1D early time radar data. The binary classifier is used to decide between two key cases, namely whether an interface is within, for example, 5 cm of the surface or not. This go/no-go detection capability is highly valuable for underground coal mining operations, such as longwall mining, where the need to leave a remnant coal section is essential for geological stability. The classifier was trained and tested using real GPR data with ground truth measurements. The real data was acquired from a testbed with coalclay, coal-shale and shale-clay interfaces, which represents a test mine site. We show that, unlike traditional second order correlation based methods such as matched filtering which can fail even in known conditions, the new method reliably allows the detection of interfaces using GPR to be applied in the near-surface region. In this work, we are not addressing the problem of depth estimation, rather confining ourselves to detecting an interface within a particular depth range.
KEY WORDSGPR, bispectrum, interface detection, horizon control, coal mining.