GPR expert

The approach used by the RMA to all three experts (MD, GPR and IR) is to extract from their data information about shape of objects [17].

In the case of GPR, on each of C-slices simple preprocessing (including edge detection) is performed and important edges are extracted. These edges from all C-slices are then put together either on the same 2D image (on fig.(21)³ edges from the same depth, i.e. same C-slice, are presented by same colour) or on the 3D image, taking into account which C-slice presents which depth.

  

Figure 21: From left to right, two C-slices followed by a 2-D image of edges from all C-slices, where colour is a label of depth, followed by a 3-D image of edges

If the A-scan envelops are detected using the Hilbert, as previously explained, the results are much better (RMA work) as it can be seen on fig.(22).

  

Figure 22: From left to right: view of a mine, 3-D GPR image of the same mine

In order to recover the correct 3-D shape of buried objects, RMA ([14]) has developed algorithms based on the convolution, by modelling the behaviour of the GPR in the time domain. The developed algorithms are faster than the classical migration methods and provide very good results as it can be seen on fig.(23).

  

Figure 23: From left to right: raw GPR 3-D image of a PMN mine, restored 3-D image of a PMN mine and restored 3-D image of a 10cm barbed wire with three barbes

The VUB has developed a classification method for ultra-sonic and GPR signals [11]. The method first consists in extracting features (the time signal itself, its Fourier transform, its auto-correlation, its wavelet coefficients, its Wigner-Ville transform and the derived scattergrams of the latter). After a selection of the most discriminant features, a supervised parametric approach based on the Bayes optimal classifier and which requires a training phase, is used. Two classification methods have been tried : the first one only implements one classifier making use of all the selected features, the second one uses a hierarchically organized multi-classifier system, combining the conditional probabilities computed by specialized classifiers either by averaging them or by multiplying them. The obtained results are described on fig.(24). The GPR data of these tests were provided by the DETEC laboratory of the ``Ecole polytechnique fédérale de Lausanne'' (Suisse).

  

Figure 24: GPR: Classification of a stone and of a PFM-1 mine in the ground from C-scans