3D U-Net Noise Attenuation
Suppress acquisition noise — surface waves, multiples, random ambient noise — while preserving fine structural detail at every reflector horizon.
The Problem
Marine and land seismic surveys are contaminated by shot-generated noise, surface-wave energy (ground roll), multiple reflections, and ambient low-frequency noise. Conventional bandpass filtering removes noise but also attenuates the signal, blurring horizons and reducing the detectability of thin beds and DHI anomalies.
The Solution
A 3D U-Net learns a data-adaptive noise model directly from examples, separating coherent signal from incoherent noise in the full 3D spatial context. Skip connections preserve high-frequency structural detail. Streaming chunk processing ensures the approach scales to full-field acquisition volumes without GPU memory constraints.
Technical Specifications
Implementation details
| Parameter | Value | Notes |
|---|---|---|
| Architecture | 3D U-Net | Ronneberger, Fischer & Brox (2015). Encoder-decoder with skip connections for full-resolution output. DOI |
| Weight Initialization | Kaiming / He | He et al. (2015). Variance-preserving initialization for ReLU activations, prevents vanishing gradients. DOI |
| Runtime | ONNX + INT8 | Jacob et al. (2018). Post-training INT8 quantization with multi-EP fallback: CUDA → DirectML → CPU. DOI |
| Processing Mode | Streaming 64³ chunks | Tukey (1967). 8-sample cosine taper overlap eliminates block-boundary artefacts during reassembly. |
| Peak Memory | < 8 GB for 50 GB volume | AIMD backpressure controller (Jacobson 1988) dynamically regulates chunk concurrency under memory pressure. |
| Normalization Methods | 7 adaptive methods | none · peak_absolute · zscore · min_max · robust_zscore (MAD, Rousseeuw-Croux 1993) · RMS (Sheriff-Geldart 1995) · percentile clip (Dramsch 2020). |
Quality Gates
Automated acceptance criteria
No result advances to the next pipeline stage unless all gating metrics are satisfied. Failures trigger automatic reprocessing with adjusted hyperparameters.
Peak signal-to-noise ratio on held-out test traces. Failures trigger automatic reprocessing.
Wang et al. (2004). Windowed 11×11 Gaussian σ=1.5, luminance/contrast/structure decomposition.
Wang, Simoncelli & Bovik (2003). Multi-scale SSIM computed at 5 spatial resolutions with power-weighted pooling.
Benchmarks
Synthetic dataset results
Evaluated on publicly available synthetic seismic benchmarks with additive Gaussian noise at multiple SNR levels.
| Dataset | SNR Input | SNR Output | SSIM | Notes |
|---|---|---|---|---|
| SEAM Phase I | 12 dB | 28.4 dB | 0.91 | Salt-body geometry preserved |
| Marmousi2 | 10 dB | 26.8 dB | 0.88 | Complex fold structure intact |
| Overthrust | 14 dB | 29.1 dB | 0.93 | Thrust fault reflectors retained |
References
Peer-reviewed foundations
- 1Ronneberger, O., Fischer, P. & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI. DOI
- 2He, K., Zhang, X., Ren, S. & Sun, J. (2015). Delving Deep into Rectifiers. ICCV. DOI
- 3Jacob, B. et al. (2018). Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. CVPR. DOI
- 4Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. (2004). Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE TIP 13(4). DOI
- 5Wang, Z., Simoncelli, E. P. & Bovik, A. C. (2003). Multi-scale Structural Similarity for Image Quality Assessment. Asilomar. DOI
- 6Tukey, J. W. (1967). An Introduction to the Calculations of Numerical Spectrum Analysis. Spectral Analysis of Time Series.
- 7Rousseeuw, P. J. & Croux, C. (1993). Alternatives to the Median Absolute Deviation. JASA 88(424). DOI
- 8Dramsch, J. S. (2020). 70 Years of Machine Learning in Geoscience in Review. Advances in Geophysics 61. DOI
See denoising on your data
Upload a SEG-Y file and watch the 3D U-Net suppress noise while preserving structural detail in under a minute.