A machine learning alternative to sparseness

Paloma Lira Fontes, Daniel O. Trad

Radon transform (RT) allows the mapping of different seismic events using different basis functions. By merging RT with machine learning (ML) in the same workflow, we aim to uncover previously unnoticed nonlinearities within these solutions, extending our insights beyond traditional geophysical theory. In this work, we employ the nonlinear capabilities of ML to discern between signal and noise within the model RT space, even when conventional techniques like localization separation and smart mute fall short. This approach becomes particularly valuable in scenarios where achieving complete spatial separation is challenging, for example in the case of multiple or ground roll overlapping primary reflections. Our numerical experiments focus on assessing the U-Net's efficacy in discerning the distinct characteristics of ground roll and multiples, employing various workflows. These include a bridge approach between Hyperbolic and Parabolic RT, aiming to complement and enhance multiple prediction in synthetic data. Additionally, we deploy the Hybrid RT methodology on the Spring Coulee dataset to forecast ground roll attenuation in the frequency domain and examine comparisons, often referred to as crosstalk, between Linear and Parabolic RT spaces. The outcomes illustrate that the U-Net has a certain capability in predicting and attenuating ground roll. However, persisting challenges lie in completely isolating reflections from the ground roll due to shared features within the Linear RT space, compounded by spatial aliasing and the irregular geometry inherent in field data.