Uso de redes neurais aplicadas para a criação de seções GPR preditivas

Abstract

During with Ground Penetrating Radar (GPR) data acquisition imaging underground targets, some operational difficulties may arise (rugged topography, anthropogenic or other nature barriers) that prevent the proper spatial positioning of the GPR lines, for example, in a survey according to a predefined regular mesh. These limitations (total number, positioning and length of GPR lines) can compromise the perfect imaging of the target and, consequently, make it difficult to reconstruct geological or engineering features of interest. The objective of this work is to develop a methodology allowing the creation of artificial GPR sections between pre-existing GPR lines, thus improving data density, using Deep Convolutional Neural Networks (DCNNs) techniques of Style Transfer and Frame Interpolation. Given a set of three consecutive GPR sections, parallel and equidistant from each other (named A, B and C), the so-called “predictive GPR sections” were created, whose intermediate reflections must honor the geometries found in sections A and C, and that have in the GPR B sections the parameters that allow the comparison, validation and quantification of the quality of the result obtained. The methodology was tested for three distinct geological contexts which consisted of: (i) a colony of montiform microbialites found in carbonate rocks in the Irecê Basin, Salitre Formation, located at Fazenda Arrecife (Bahia); (ii) sandy siliciclastic rocks of Aeolian origin that occur in the Parnaíba Basin, belonging to the Pedra de Fogo Formation, located in the Serra das Araras (Piauí), which exhibit tabular and wedge-shaped depositional geometries; (iii) and, in a Quaternary washover deposit that show wedge geometry, located on the shore of Lagoa Mirim, Vila Taim (Rio Grande do Sul). Predictive GPR sections were created to the three different geological scenarios studied. With the delimitation of the original and predictive geometries, it was possible to compare the differences based on their overlapping and, thus, calculate the respective standard deviations and mean misfits. The comparison between the techniques reveals that the geometries of the microbialites obtained with Frame Interpolation and the two techniques association exhibited differences greater than that of Style Transfer, being up to 0.015 and 0.009 greater, respectively, for the standard deviation and mean misfit. For aeolian rocks they were 0.020 of standard deviation and 0.017 of mean misfit greater for two techniques association than for Style Transfer. And finally, in the washover deposit, the differences between the two parameters were not greater than 0.004. Exclusively in this deposit, the predictive geometries identified in both techniques produced equivalent results. Therefore, the predictive geometries obtained show that the Style Transfer technique was the one that presented the best precision and accuracy results for microbialites and aeolian rocks, while in the washover deposits the two techniques association reproduced results as good as the Style Transfer. Only for the microbialites was it possible to recognize the predictive depositional geometries when using the Frame Interpolation technique.