Diffraction pattern recognition using deep semantic segmentation

Diffraction imaging can help better understand small-scale geological structures. Due to their often-weak signal, in order image them, it is necessary separate diffraction signals from the rest of wavefield. Many different methods have been developed for wavefield separation, and newest trend includes application artificial neural networks deep learning. Available case studies with a deep-learning approach separation show good results when applied synthetic sedimentary setting datasets where are either strong or pronounced characteristics. Examples, however, missing crystalline hardrock settings signal-to-noise ratio by far lower usually within complex reflectivity medium, steep tails incomplete. In this study, we showcase semantic segmentation model on seismic, real ground-penetrating radar, seismic datasets. Synthetic sections were generated using random noise levels coherent resembling pattern interfering tails. For GPR dataset, successfully delineated, although some locations reflections picked up because similar pixel values as apex diffractions. As through number approaches, able completely delineate single several inlines that was massive sulphide body. The algorithm also enabled us recognize an incomplete diffraction, at edge cube, which never labelled. This originated outside volume may be target future mineral exploration programmes, thanks providing possibility.