If training data appears to be mislabeled, should we relabel it? Improving supervised learning algorithms for threat detection in ground penetrating radar data