---
tags:
- seismic
- ground-roll
- denoising
- unet
- resunet
- dncnn
- attention-unet
- pytorch
library_name: pytorch
---

# Ground-Roll Attenuation Benchmark

Deep-learning-based coherent noise (ground roll) suppression on pre-stack seismic shot gathers, using the SEG C3 synthetic dataset.

## Task

Given a noisy shot gather contaminated by dispersive ground-roll noise, the model predicts the additive noise component. The denoised signal is obtained by:

```
denoised = noisy_input - predicted_noise
```

This is a **paired regression** task trained with a noise-label objective (the ground-truth noise component). The supervision target is the residual between the noisy input and the clean reference.

## Dataset

- **Source**: SEG C3 pre-stack synthetic data, 9 regular shot gathers
- **Geometry**: 201 traces × 625 time samples per shot, dt = 2 ms
- **Noise modeling**: Reflection signals modeled with the acoustic wave equation; ground roll modeled with the elastic wave equation to capture its dispersive, low-velocity character
- **Split**: Shot-level (FFID) sequential 7:1:1 — 7 training shots, 1 validation, 1 held-out test

### Noise Intensity Levels

Five ground-roll intensity levels produce paired noisy / noise-label records:

| Level | SNR (dB) | PSNR (dB) | SSIM | MAE | MSE | RMSE |
|-------|----------|-----------|------|-----|-----|------|
| 1.0   | 2.71     | 15.81     | 0.9480 | 0.030624 | 0.026232 | 0.161964 |
| 3.0   | -6.83    | 6.27      | 0.9418 | 0.091871 | 0.236091 | 0.485892 |
| 5.0   | -11.27   | 1.83      | 0.9402 | 0.153118 | 0.655809 | 0.809820 |
| 7.0   | -14.19   | -1.09     | 0.9395 | 0.214366 | 1.285386 | 1.133749 |
| 9.0   | -16.37   | -3.27     | 0.9390 | 0.275613 | 2.124822 | 1.457677 |

*Metrics computed on the full dataset in the original amplitude domain before normalization.*

## Model Architectures

- **ResUNet** (`res_unet`) — U-Net with residual blocks replacing plain double-conv layers (He et al., 2016; Zhang et al., 2018). Base channels: 32, depth: 4.
- **UNet** (`unet`) — Classic encoder-decoder with skip connections (Ronneberger et al., 2015). Base channels: 32, depth: 4.

### Preprocessing

- **Normalization**: `max_abs`, global scope — the entire dataset scaled to [-1, 1]
- **Patching**: Overlapping 2D patches (128 × 256) with 50% overlap, channel-last format (1, H, W)

## Repository Structure

```
models/
├── unet/
│   ├── level1.0_seed42/
│   │   ├── best.pt          # Best checkpoint (minimum validation loss)
│   │   └── config.yaml      # Full training configuration
│   ├── level1.0_seed43/
│   ├── level1.0_seed44/
│   ├── level3.0_seed42/
│   └── ...
└── res_unet/
    └── ...
```

Each subdirectory corresponds to one experiment: a model architecture trained at a specific noise level with a specific random seed.

## Training Details

| Hyperparameter | Value |
|----------------|-------|
| Loss | MSE (predicted noise vs. label noise) |
| Optimizer | AdamW (lr=1e-3, weight_decay=1e-4) |
| Scheduler | Cosine annealing (min_lr=1e-6) |
| Epochs | 200 |
| Gradient clipping | 1.0 (max norm) |
| Batch size | 196 |
| Distributed training | 2 × NVIDIA RTX 4090, DDP |
| Seeds | 42, 43, 44 per experiment |

## Usage

```python
import torch
from huggingface_hub import hf_hub_download

# Download a checkpoint
repo = "GeoBrain/coherent-noise-attenuation"
model_key = "res_unet"
level = "3.0"
seed = "42"

ckpt_path = hf_hub_download(
    repo_id=repo,
    filename=f"models/{model_key}/level{level}_seed{seed}/best.pt",
)

# Load state dict
state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)

# For full model loading, instantiate the corresponding architecture
# and load the state dict (see config.yaml for exact architecture params).
```

See the companion benchmark documentation for detailed experimental setup and full evaluation results.

## Results (SNR dB)

| Model | Params (M) | Level 1.0 | Level 3.0 | Level 5.0 | Level 7.0 | Level 9.0 |
|-------|:----------:|:---------:|:---------:|:---------:|:---------:|:---------:|
| Raw (noisy) | — | 2.71 | −6.83 | −11.27 | −14.19 | −16.37 |
| UNet | 7.76 | 28.39±1.09 | 23.07±0.10 | 20.22±0.32 | 17.90±0.28 | 16.60±0.46 |
| ResUNet | 8.11 | 31.62±0.56 | 22.65±0.93 | 18.11±2.12 | 16.60±1.44 | 14.61±0.01 |
| DnCNN | 0.56 | 31.67±0.11 | 30.02±0.43 | 22.91±3.74 | — | — |
| Attention UNet | 7.85 | 28.79±0.46 | 23.10±1.00 | 19.66±0.22 | 17.57±0.11 | 16.61±0.19 |

Mean ± std over 3 seeds. All models achieve 15–31 dB SNR improvement. DnCNN delivers the best performance at low-to-mid noise levels with the smallest footprint (0.56 M parameters). See the benchmark documentation for per-level detailed metrics (PSNR, SSIM, MAE, MSE, RMSE).

## References

- Ronneberger et al., U-Net: Convolutional Networks for Biomedical Image Segmentation, MICCAI 2015
- He et al., Deep Residual Learning for Image Recognition, CVPR 2016
- Zhang et al., Image Denoising via Deep CNN (DnCNN), IEEE TIP 2017
- Oktay et al., Attention U-Net: Learning Where to Look for the Pancreas, MIDL 2018
- SEG C3 Velocity Model: https://wiki.seg.org/wiki/C3