Attenuating free‐surface multiples and ghost reflection from seismic data using a trace‐by‐trace convolutional neural network approach

• Published in: Geophysical Prospecting Vol. 72; no. 3; pp. 908 - 937
• Main Authors: Kiraz, Mert S. R., Snieder, Roel, Sheiman, Jon
• Format: Journal Article
• Language: English
• Published: Houten Wiley Subscription Services, Inc 01.03.2024
• Subjects: Air-water interface; Algorithms; Artificial neural networks; Attenuation; Broadband; Correlation coefficient; Correlation coefficients; Data acquisition; Data analysis; Data processing; Ghosts; Graben; Mathematical models; Neural networks; noise; Parameters; Reflection; Seismic data; Seismic data processing; seismics; signal processing; Velocity; Waveforms
• ISSN: 0016-8025; 1365-2478
• DOI: 10.1111/1365-2478.13443

The presence of the air–water interface (or free‐surface) creates two major problems in marine seismic data for conventional seismic processing and imaging: free‐surface multiples and ghost reflections. The attenuation of free‐surface multiples remains one of the most challenging noise attenuation problems in seismic data processing. Current solutions suffer from the removal of the primary events along with the multiple events especially when the primary and multiple events overlap (e.g., adaptive subtraction). The effective attenuation of ghost reflections (or deghosting) requires acquisition‐ and/or processing‐related solutions which generally address the source‐side and receiver‐side ghosts separately. Additionally, an essential requirement for a successful implementation of free‐surface multiple attenuation and seismic dehosting is the requirement of dense seismic data acquisition parameters which is not realistic for two‐dimensional and/or three‐dimensional marine cases. We present a convolutional neural network approach for free‐surface multiple attenuation and seismic deghosting. Unlike the existing solutions, our approach operates on a single trace at a time, and neither relies on the dense acquisition parameters nor requires a subtraction process to eliminate free‐surface multiples, and it removes both the source ghost and receiver ghost simultaneously. We train a network using subsets of the Marmousi and Pluto velocity models and make predictions using subsets of the Sigsbee velocity model. We show that the convolutional neural network predictions give a correlation coefficient of 0.97 on average with the numerically modeled data for the synthetic examples. We illustrate the efficacy of our convolutional neural network–based technique using the Mobil AVO Viking Graben field data set. The application of our algorithm demonstrates that our convolutional neural network–based approach removes different orders of free‐surface multiples (e.g., first and second orders) and recovers the low‐frequency content of the seismic data (which is essential for, for instance, full‐waveform inversion applications and broadband processing) by successfully removing the ghost reflections while preserving and increasing the continuity of the primary reflection.