A comparison of segmentation techniques for target extraction in ground penetrating radar data

In a typical GPR survey, a great amount of data is collected, and only a small percentage of this amount represent useful data (i.e. target data) whereas the majority of the data is considered non-useful. The first of the post-processing stages, which relies heavily on a skilled operator, involves pointing out the areas that may contain targets and suppressing others. Consequently, this process consumes considerable amounts of time and effort, apart from the fact that the existence of the human factor at this critical stage invariably introduces inconsistency and error into the interpretation. In this paper, three techniques for automatic GPR data segmentation are discussed and compared. The techniques rely on the computation of certain features from which a neural network can then arrive at a decision to classify the data segments in question as targets or otherwise. The first technique is based on extracting three statistical features from A-scan segments while the second technique computes statistical features from B-scan regions to produce the discrimination bases for the neural classifier to distinguish targets from non-targets. In the third technique, some regional properties of B-scan segments accompanied by the statistical features are used to achieve discrimination not only between targets and non- targets, rather between hyperbolic-shaped and non-hyperbolic-shaped targets as well. All three techniques were on different types of GPR data collected from a variety of sites, and they proved to be very efficient to form a robust automatic technique for data reduction and segmentation.