Paper submitted to EAGE Annual 2026

Summary

Diffractions are generated when a wave interacts with small subsurface features such as a tip, point, or edge discontinuity. To image diffractions is to image those fine details that naturally occur in the Earth.  This property finds immediate application in the detection and mapping of point diffracting subsurface boulders, especially those buried tens to hundreds of meters beneath the sea floor, which pose substantial risk to the installation of wind turbines (Wenau et al., 2022).  Extracting the diffractions from seismic data is notoriously difficult because diffractions are weak and can easily be overpowered by reflections or noise.  Nevertheless, strong and even weak diffractions are easily visible to the human eye. This suggests that a convolutional neural network (CNN), which excels at computer vision problems, may do well at this problem.

Traditional methods for extracting diffractions fall into two broad categories.  On the one hand, there are time domain, coherency-based methods that capture and subtract the reflected energy to reveal the diffracted energy plus noise (e.g. plane wave destruction, Taner et al., 2006).  The quality of the extracted diffractions is dependent on the residual noise level.  On the other hand, there are image domain methods that target the diffractions directly during migration (Schwarz, 2019).  These methods can be computationally demanding and require a good velocity model.  In either case, these traditional approaches suffer from requiring extensive parameter testing.

More recent approaches involve the use of machine learning models, specifically neural networks, to detect diffractions within seismic data (Lowney et al., 2021; Bauer et al., 2025). These models are promising as they can learn the features associated with diffractions directly, resulting in better signal to noise ratio.  Once a good model has been trained these models offer fast, parameter free solutions. However, the burden of work falls back to training, of which the major component is preparing the training data.  In this study, we explore the impact of data labelling approaches to train a DUCK-Net (Dumitru et al. 2023) for extraction of the diffracted wavefield.