Beyond Labels: A Survey of Label-Efficient Deep Learning Techniques in Seismic Exploration

Abstract

Seismic exploration plays a pivotal role in subsurface characterization and has been greatly promoted by artificial intelligence techniques such as deep learning, yet its reliance on high-quality labeled data poses a significant challenge for conventional supervised learning methods. In recent years, label-efficient deep learning has emerged as a powerful solution to this bottleneck by leveraging unsupervised, self-supervised, semi-supervised, and weakly supervised paradigms. This survey provides a comprehensive overview of these learning paradigms and their growing impact on key stages of the seismic data processing workflow, including preprocessing, imaging and inversion, and geological interpretation. We first outline the theoretical foundations and representative architectures of each paradigm, highlighting their distinct supervision strategies and learning objectives. We then analyze recent advancements in applying label-efficient methods to tasks such as seismic denoising, interpolation, full waveform inversion, facies classification, fault detection, and salt body delineation. By systematically comparing methodologies across application scenarios, we identify their respective advantages, limitations, and domain-specific adaptations. Finally, we discuss the main challenges hindering large-scale deployment, including the lack of standardized benchmarks, difficulty in integrating geophysical constraints, and the interpretability gap, and we also suggest promising directions for future research. This review aims to serve as a comprehensive reference for geoscientists and machine learning practitioners seeking to harness label-efficient deep learning for intelligent and scalable seismic exploration.