Diffraction Recognition Using Deep Learning

Diffractions are important features in seismic data that carry information on small-scale geological entities. Because of their low-energy signal compared to the reflected waves, in order to image them, diffraction signal needs to be separated from the rest of the wavefield including ambient noise. Several methods have been developed for diffraction signal separation, and the newest trends include deep learning algorithms. Using deep learning, as a tool for diffraction recognition and separation has been promising particularly in sedimentary settings. In this study we use deep learning algorithm for diffraction classification in hardrock settings dominated with much lower signal-to-noise ratio and much higher background velocity. Staring with simple synthetic seismic sections and in order to make more complex case, we add more features such as different noise levels and reflection features with different dip directions. Results show good potential, taking into account a limited amount of training data. We also employed the developed algorithm on a real GPR dataset with strong diffractions to showcase the workflow. Our longer-term plan is to build a workflow enabling deep learning algorithms for hardrock seismic datasets in order to automatize feature extractions including diffractions.