DiffractioNet: Deep-learning seismic and GPR diffraction separation

• DOI: 10.1190/geo2024-0279.1
• Publication Date: 2025-02-13T15:39:19Z
• Authors: Alexander Bauer; Benjamin Schwarz; Jan Walda; Dirk Gajewski
• Keywords: Ground-Penetrating Radar; Separation (statistics)

The separation of the diffracted wavefield is a notorious challenge in both seismic and ground-penetrating-radar (GPR) subsurface imaging. Over last decade, numerous studies have attempted to address this with various deterministic schemes. While each these schemes has specific advantages disadvantages, they all require an adaptation corresponding processing parameters for application, especially when crossing scales between electromagnetic measurements. In recent years, convolutional neural networks (CNNs) emerged as powerful tools data analysis. However, their performance strongly depends on training labels, generation which can be complex time-consuming process. study, we introduce DiffractioNet, deep-learning framework diffraction separation. Our approach based automated synthetic patches labels reflected wavefields. We augment dataset reference field results from coherent subtraction, well-established method With combined dataset, train CNN decompose any input or GPR into trained DiffractioNet provides solution efficient on-the-fly that does not parameter adaptation. demonstrate by applying unseen data.