This article focuses on the possible drawbacks and pitfalls in the GPR data interpretation process commonly followed by most GPR practitioners in archaeological prospection. Standard processing techniques aim to remove some noise, enhance reflections of the subsurface. Next, one has to calculate the instantaneous envelope and produce C-scans which are 2D amplitude maps showing high reflectivity surfaces. These amplitude maps are mainly used for data interpretation and provide a good insight into the subsurface but cannot fully describe it. The main limitations are discussed while studies aiming to overcome them are reviewed. These studies involve integrated interpretation approaches using both B-scans and C-scans, attribute analysis, fusion approaches, and recent attempts to automatically interpret C-scans using Deep Learning (DL) algorithms. To contribute to the automatic interpretation of GPR data using DL, an application of Convolutional Neural Networks (CNNs) to classify GPR data is also presented and discussed.

Ground-penetrating Radar (GPR) sections commonly suffer from strong scattered energy and weak reflectors with distorted lateral continuity. This is mainly due to the gradual variation of moisture with depth, dense lateral sampling of common-offset GPR traces (which are considered as zero-offset data), along with the small wavelength of the electromagnetic waves that is comparable to the size of the shallow subsurface dielectric heterogeneities. Focusing of the diffractions requires efficient migration that, in the presence of highly heterogeneous subsurface formations, can be improved by a detailed migration velocity model. Such a velocity model is difficult to develop because the common-offset antenna array is mostly used for its reduced time and cost in the data acquisition and processing stages. In such cases, migration processes are based on limited information from velocity analysis of clear diffractions, cores, or other ground truth knowledge, often leading to insufficient imaging. We have developed a methodology to obtain GPR sections with focused diffractions that is based on multipath summation, using weighted stacking (summation) of constant-velocity migrated sections over a predefined velocity range. The success of this method depends on the assignment of an appropriate weight, for each constant-velocity migrated section to contribute to the final stack, and the optimal width of the velocity range used. Additionally, we develop a postmultipath summation processing step, which consists of time-varying spectral whitening, to deal with the decrease of the dominant frequency due to attenuation effects and the additional degraded resolution expected by the constant migration summed images. This imaging strategy leads to GPR sections with sufficiently focused diffractions, enhancing the lateral and the temporal resolution, without the need to explicitly build a migration velocity model.

Several researchers have utilized multipath summation to manage the common problem of scattered energy within GPR sections. Such energy results in degrading the lateral resolution and continuity of reflectors. If detailed velocity models are known, then it is fairly easy to focus the scattered energy by means of conventional migration methods. However, this is rarely the case in GPR sections, as the common-offset antenna array is mostly used, and therefore cannot provide velocity models. This gives an important advantage for the multipath summation method, which has proved to be successful in focusing such diffractions, without the need to build a detailed migration velocity field model. This multipath summation method is based on stacking (summation) of constant velocity migrated sections (weighted or not) over a predefined velocity range. The main drawback of this technique is the high computational cost and the need for user interference to select the appropriate stacking weights. We developed an improved implementation of the weighted multipath summation method that reduces both the computational cost, and the user interference in stacking weights selections. This data adaptive methodology can expedite the migration process, suppress the need for a detailed velocity model, and reduce the user subjectivity. Moreover, a data adaptive spectral scaling scheme was developed. This is applied on the output of the multipath summation process to reduce the expected blurriness in the resulting GPR sections.