Literature review on deep learning for the segmentation of seismic images

Abstract

This systematic literature review provides a comprehensive overview of the current state of deep learning (DL) specifically targeted at semantic segmentation in seismic data, with a particular focus on facies segmentation. We begin by comparing the contributions of DL to traditional techniques used in seismic image interpretation. The review then explores the learning paradigms, architectures, loss functions, public datasets, and evaluation metrics employed in seismic data segmentation. While supervised learning remains the dominant approach, recent years have seen a growing interest in semi-supervised and unsupervised methods to address the challenge of limited labeled data. Additionally, we found that the U-Net architecture is the most prevalent backbone for semantic segmentation, appearing in one-third of the articles reviewed. We also present a comprehensive compilation of the results obtained by 24 methods and discuss the challenges and research opportunities in this field. Notably, the lack of standardized protocols for performance comparison, combined with variability in datasets and evaluation metrics across studies, raises questions about what truly constitutes the current state of the art in semantic segmentation of seismic data.