The detection of coal/shale interfaces close to the surface using ground penetrating radar (GPR) is difficult because the echo is drowned in unwanted signal components. These components are susceptible to changes in amplitude and cannot be removed well by standard preprocessing techniques. Second order correlation based methods such as matched filtering fail to locate the echo reliably under these conditions. Features that extract signal shape information based on the bispectrum can provide desirable immunity from clutter and noise. These features are investigated for this task using signals obtained from finite difference time domain simulations of a typical coal seam. It is shown that a pattern recognition approach using features derived from the bispectrum allows the detection of layer interfaces using GPR to be extended reliably to the near surface region.