Optimal GPR bandwidth for time-frequency landmine discrimination

In this work we investigate which bandwidth of a ground penetrating radar (GPR) is optimal for time-frequency landmine discrimination. We extracted three time-frequency features of the early-time target response from the Wigner distribution. The features were found to be relatively invariant to target depth for a data acquired with a stepped-frequency ultra-wideband GPR. The frequency sweep was from 0.3 GHz up to 6 GHz. The features allowed discrimination of two different low-metal landmines from a mine-like stone. The results were visualized in the three-dimensional feature space where each point related to a certain target represents a certain GPR scenario. For a number of scenarios we obtained two separated clusters for the landmines and the stone respectively. Numerically the quality of target discrimination can be evaluated with the Mahalanobis distance which estimates the separation between such feature clusters accounting for their size. Here we use the Mahalanobis distance as a criterion of optimality for the GPR bandwidth. Having obtained good results for the large data bandwidth, we reduce it by digital filtering with a small step in changing the cut-off frequencies, then extract the features and compute the Mahalanobis distance between the landmines and the stone. Its maximal value defines the optimal GPR lower and upper frequencies.