Machine Learning alternative to sparseness: a Radon transform application for multiple and ground roll attenuation

Paloma Lira Fontes

Radon transform (RT) allows mapping different seismic events using different basis functions. Integrating RTs with machine learning (ML) presents an innovative approach to addressing non-linear problems in seismic data. By incorporating deep learning (DL) techniques into the framework, the objective is to address challenges encountered in seismic data processing, particularly in separating signal from coherent noise within the model space. A central idea of this work is the utilization of ML as an alternative to traditional sparseness techniques, which can struggle with complex and overlapping seismic events. Through a pixel-by-pixel approach, ML-based approaches leverage DL’s capability to discern non-linear patterns in images, enabling effective segregation of multiples and ground roll from reflections in the RT domain. This approach becomes particularly valuable in scenarios where achieving complete spatial separation is challenging, for example, in multiple or ground roll overlapping primary reflections. This study conducts numerical experiments to assess the U-Net’s effectiveness in discerning ground roll and multiples, employing various workflows to predict RT panels and maximizing RT utility by incorporating multiple channels of information like RGB colour channels in an image. Experiments examined the efficiency of different RT types, such as Hyperbolic RT (HRT) and Parabolic RT (PRT), for training a U-Net model to predict multiples. While U-Net partly succeeded in predicting multiples and highlighting the importance of label selection, it also faced challenges. Transform artifacts linked to input geometry, like truncation and sampling, hampered inference, lowering generalization. Furthermore, tests deploying a Hybrid Linear-Parabolic RT methodology for ground roll suppression in a field dataset helped analyze crosstalk problems between different RT spaces. Different channels provided insights into the leakage of ground roll among the RT used, with the three-channel approach showing promising results to forecast ground roll attenuated RT panels but challenges persisting in fully disentangling ground roll from reflection data. Continued research efforts are crucial to address these challenges and unlock the full potential of ML with seismic.