The core problem the researchers set out to solve is deceptively simple to state and brutally hard to solve in practice. Supervised deep learning models, the workhorses of modern image restoration, need pairs of noisy and clean images to learn from. In seismic acquisition, they are not.

Field records are inevitably corrupted by ambient random noise, and obtaining a genuinely noise-free ground-truth record of the subsurface is, as the authors put it, both prohibitively expensive and practically impossible. Annotation compounds the difficulty: labeling seismic data demands meticulous manual work by professional geologists, is time-consuming and labor-intensive, and is prone to human bias. The result is that the large, high-quality datasets on which supervised models thrive simply do not exist for seismic data at the scale needed.

SA-S2S sidesteps this bottleneck entirely by embracing self-supervised learning, a paradigm in which the model extracts meaningful features from the raw data itself, with no manual labels. The method builds on the Self2Self framework and adds a crucial ingredient: a three-dimensional spatial self-attention mechanism. The underlying observation model treats the recorded seismic volume as the sum of the true subsurface reflection response and zero-mean random noise.

Because clean references are unattainable, the network is trained under a blind-spot formulation: a random binary Bernoulli mask stochastically occludes voxels in the input, and the model must infer the hidden amplitudes purely from the surrounding intact context. This masking is deliberately distinct from seismic trace interpolation, which addresses physically missing receiver channels; here it serves strictly as a training constraint that prevents the network from collapsing into a trivial identity mapping that would memorize both reflections and noise alike. The physics behind why this works is elegant.

Seismic reflections exhibit strong structural coherence along subterranean strata, whereas ambient random noise is spatially uncorrelated. When the network is forced to reconstruct masked voxels from their neighbors, it can only succeed by exploiting that coherence, so it learns to rebuild continuous reflection waveforms while the incoherent noise fluctuations average out. Standard local convolutions, however, struggle with large-scale continuity and dipping horizons.

This is where the self-attention module earns its keep: by explicitly computing pairwise structural affinities across the entire three-dimensional volume, it acts as what the authors describe as an adaptive semblance operator, aggregating coherent reflection energy across distant traces while canceling uncorrelated noise. In the encoder, features pass through a PartialConv3d layer with LeakyReLU activation, then branch into an identity residual shortcut and a self-attention path where query, key, and value tensors are generated through one-by-one convolutions and depthwise three-by-three-by-three convolutions, with transposed attention computed through the inner product of queries and keys, normalized by Softmax. The decoder, by contrast, is deliberately simple.

Rather than assuming that deeper and more complex architectures are necessary for effective denoising, the team adopted a streamlined design of cascaded dropout, convolution, and LeakyReLU modules capped with a Sigmoid output layer. Monte Carlo dropout plays a double role: given the scarcity of high-quality seismic data, it enhances generalization, and at test time it provides variance reduction through ensembling. A bridging module between encoder and decoder transfers multi-scale features, allowing high-quality reconstruction even with limited samples.

Training used the AdamW optimizer with a learning rate that started at one ten-thousandth for two thousand epochs and then decayed via cosine annealing to one hundred-thousandth over the remaining one thousand epochs, with a fixed budget of twenty-four thousand steps as the stopping criterion. To evaluate the approach rigorously, the researchers generated two synthetic three-dimensional seismic datasets, each with spatial dimensions of 128 by 128 by 128 voxels, and contaminated them with Gaussian random noise at three intensity levels corresponding to standard deviations of 10, 25, and 50. They benchmarked SA-S2S against a formidable lineup: the structure-oriented filter representing traditional dip-guided smoothing, the deep image prior method, the original Self2Self baseline, Neighbor2Neighbor, Blind2Unblind, and a lightweight three-dimensional adaptation of the Restormer Transformer with roughly a quarter-million parameters.

Across both datasets and all three noise levels, SA-S2S achieved the highest peak signal-to-noise ratio and the best or joint-best structural similarity index in almost every case. On the first dataset it reached 41.1, 37.9, and 33.1 decibels at the three noise levels respectively, improving on the strongest baseline, Self2Self, by 0.6, 0.2, and 0.4 decibels. On the second dataset it posted 32.6, 31.7, and 30.0 decibels, with gains of 0.3, 0.2, and 0.1 decibels over Self2Self.

Visual and spectral analyses reinforced the numbers. Noise separation experiments showed that the structure-oriented filter introduced distortions and lost valid information, the deep image prior method was biased toward high-frequency noise, and Self2Self was biased toward low-frequency noise, sometimes blurring layer boundaries. SA-S2S, in contrast, suppressed noise evenly across the frequency spectrum, producing separation results that closely resembled the ideal noise distribution.

Fast Fourier transform comparisons confirmed that the denoised outputs of SA-S2S tracked the frequency distribution of the original clean data more closely than any competitor, even under the harshest fifty-sigma noise conditions, where all methods suffered some irreversible signal loss yet SA-S2S still reconstructed fault characteristics and stratigraphic layers with striking clarity. The true test of any seismic processing method is whether it generalizes to real field data, and here the team deployed SA-S2S on two geologically contrasting public surveys. The first was the North Sea F3 block in the Dutch offshore sector, featuring gentle structural dips, sigmoid clinoform progradation, and large-throw normal fault systems within Tertiary and Mesozoic strata.

The second was the Kerry3D survey in New Zealand's Taranaki Basin, a structurally complex environment with steep folding, intense faulting, and subtle fault displacements within faulted anticlinal traps. On both surveys, SA-S2S enhanced structural continuity across gentle horizons while sharpening fault edges without distorting geological dips, and its noise residuals contained only minimal signal components, with no visible fault displacements or reflector continuity in the removed noise, evidence that the sharpened boundaries originated from true wavefield energy rather than generative hallucination. An ablation study dissected exactly which components mattered.

Removing skip connections caused the largest damage, dropping the peak signal-to-noise ratio by 3.36 decibels, while disabling dropout triggered a catastrophic 11.47-decibel collapse, confirming that Monte Carlo dropout ensembling is vital for test-time variance reduction. Perhaps surprisingly, eliminating the three-dimensional self-attention module changed the metric by only 0.01 decibels, and multi-seed stability tests showed the attention mechanism improved the mean restoration quality only marginally, from 40.91 to 41.00 decibels, without stabilizing training. The authors interpret attention as a data-driven semblance stacker whose benefit lies in aggregating coherent energy across distant horizons rather than in raw metric gains.

Computational profiling on a single NVIDIA RTX 4070 Ti showed SA-S2S at 3.034 million parameters, 239.81 gigaflops, and a median inference latency of 20.62 milliseconds, comparable to Self2Self and a trade-off the authors consider well justified. The method has honest boundaries. It is designed specifically for random noise attenuation, and strongly coherent noise such as ground roll, guided waves, and acquisition footprints, which shares trace-to-trace continuity with valid reflections, cannot be separated by spatial blind-spot masking alone; under pure coherent noise all tested methods performed poorly, with SA-S2S reaching only 22.91 decibels.

The authors also caution that peak signal-to-noise ratio and structural similarity, being spatial averages over millions of voxels, can dilute localized reconstruction errors, so quantitative scores should be read alongside local zoom-in profiles and fault-plane sharpness. Still, by delivering denoised seismic volumes with well-preserved structural boundaries while requiring no labeled data at all, SA-S2S offers exploration geophysicists a practical, efficient tool that could sharpen the geological maps on which drilling decisions, resource estimates, and high-precision three-dimensional geological models ultimately depend.

Subject of Research: Self-supervised seismic image denoising using a self-attention-based blind-spot network

Article Title: A self-supervised random noise attenuation method for seismic data based on self-attention

Article References: Liu, S., Chen, G., Cao, L., Fan, Y., Zuo, K., & Liu, W. (2026). A self-supervised random noise attenuation method for seismic data based on self-attention. Environmental Earth Sciences, 85(17), Article 444. https://doi.org/10.1007/s12665-026-13165-z

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13165-z

Keywords: seismic data, denoising, self-supervised learning, self-attention, deep learning, 3D seismic imaging, fault detection, signal-to-noise ratio, geological boundaries, oil and gas exploration, blind-spot masking, PSNR