Ensemble deep learning for improved reconstruction of weak events, conflicting dips, and high frequencies

Seismic data reconstruction is crucial in scenarios where the original seismic data is incomplete, noisy, or corrupted. Using deep learning for the task, a model is trained to learn the characteristics of events from the existing data to predict the missing parts. Observing that a simple U-net fails to predict the poorly represented aspects of the data, we propose a new ensemble model using custom data transformation modules inside the architecture. We target weak events, conflicting dips, and high frequencies that are reported to pose challenges for a conventional deep model. Testing our method for self-supervised reconstruction of consecutive missing traces of two benchmark synthetic data and a real marine dataset shows improvements in reconstruction accuracy.