Unsupervised and Self-Supervised Deep Learning for Seismic Data Processing: Methods and Applications
Ji Li
Seismic data processing is vital for subsurface imaging but is hindered by incoherent, coherent, and partially coherent noise as well as incomplete or irregular sampling. Classical model-based approaches (e.g., f - x filtering, Radon transform, curvelet thresholding) depend on strong priors and often fail to generalize across noise types and acquisition geometries. Supervised deep learning requires clean labels, which are rarely available in practice. This thesis develops self-supervised and unsupervised learning frameworks that eliminate the need for clean data and exploit the intrinsic structure of seismic signals.
In the self-supervised setting, two strategies are explored. Noise2Noise (N2N) training uses synthetic noisy pairs to learn denoisers without clean targets, effective for random and high-amplitude erratic noise such as blending interference. Blind-spot networks are adopted for unpaired data, and a structured blind-spot (StructBS) model is extended with self-attention and structure-aligned regularization to better handle partially coherent noise like swell noise. In the unsupervised setting, implicit neural representations (INRs) with SIREN are used to fit noisy or undersampled data directly, leveraging the network??s inductive bias for denoising and interpolation. Applications include incoherent/coherent noise attenuation, distributed acoustic sensing (DAS) enhancement, and 5D interpolation, with a Plug-and-Play variant improving robustness on heavily corrupted inputs.
Experiments on synthetic and field data show that these methods achieve high-quality denoising and reconstruction without clean labels, adapt across noise types and survey geometries, and often match or surpass strong model-based baselines. The results demonstrate the potential of scalable, label-free learning priors for modern seismic data processing.