A Self-Supervised Learning Framework for Multi-Tasks Seismic Processing Via Meta-Learning

Summary Despite the typical commonalities and consistent features found in seismic data, neural networks are frequently trained for singular, specific tasks. Leveraging the acquired features and representations across multiple tasks can reduce the training cost and enhance the network's resilience. Meta learning offers the opportunity for a network to adapt to many tasks efficiently; however, a notable drawback lies in its dependency on labeled data for supervised training, which hinders its application to real data. To address this issue, we employ a self-supervised meta learning paradigm that allows direct training on label-less real data for many seismic processing tasks. In this framework, both the meta-training and meta-testing phases utilize a consistent self-supervised learning (SSL) strategy, namely iterative data refinement. Specifically, the pseudo-labels for the current epoch are derived from the predictions of the network trained in the previous epoch using the original data, while the input data are constructed by processing these predicted pseudo-labels, depending on the specific seismic processing task at hand. Numerical examples demonstrate the effectiveness of the proposed framework in improving quality and reducing cost across various seismic processing tasks under an SSL paradigm.