Extraction of diffraction events from seismic data using deep learning based approach

Abstract

Diffractions carry information that can help imaging of small-scale heterogeneities smaller than the seismic wavelength. Extracting diffraction events is key step because the amplitude is weaker than that of overlapped reflection events. Recently, deep learning (DL) based approach has been used as a powerful tool for diffraction separation. However, most DL approaches only identify the locations of diffractions, separation of diffractions were inaccurate. In this work, we proposed DL based diffraction extraction method which preserves the amplitude and phase characteristics of diffraction. Owing to the systematic generation of training dataset using t-SNE analysis, we can extract faint diffractions and diffraction tails overlapped by strong reflection events. In addition, we clearly demonstrated the effect of training dataset on the DL performance. Since the extracted diffractions by our method preserve the amplitude and phase, they can be used for velocity model building and high-resolution imaging with diffractions.