SEG C3 Ground-Roll Dataset

Paired noisy-input / noise-label SEG-Y volumes for supervised ground-roll attenuation, derived from the SEG C3 synthetic velocity model.

Task

Noise-label regression: given a noisy pre-stack shot gather, predict the additive ground-roll noise component. The clean signal is recovered as:

denoised = noisy_input - predicted_noise

The noise labels serve as regression targets. Both input and label are 3D SEG-Y volumes with identical geometry.

Example 1

Noisy Data

Clean Data

Ground-Roll Noise Label

representative ground-roll attenuation sample at noise level 3.0.

Example 2

Noisy Data

Clean Data

Ground-Roll Noise Label

representative ground-roll attenuation sample at noise level 3.0.

Example 3

Noisy Data

Clean Data

Ground-Roll Noise Label

representative ground-roll attenuation sample at noise level 3.0.

File Structure

Each noise level has a matched pair of SEG-Y files:

Level	Noisy Input	Noise Label	Size (approx)
1.0	SEGC3_shots1_9_noisy_1.0.sgy	SEGC3_shots1_9_noise_1.0.sgy	~951 MB × 2
1.5	SEGC3_shots1_9_noisy_1.5.sgy	SEGC3_shots1_9_noise_1.5.sgy	~951 MB × 2
2.0	SEGC3_shots1_9_noisy_2.0.sgy	SEGC3_shots1_9_noise_2.0.sgy	~951 MB × 2
2.5	SEGC3_shots1_9_noisy_2.5.sgy	SEGC3_shots1_9_noise_2.5.sgy	~951 MB × 2
3.0	SEGC3_shots1_9_noisy_3.0.sgy	SEGC3_shots1_9_noise_3.0.sgy	~951 MB × 2
4.5	SEGC3_shots1_9_noisy_4.5.sgy	SEGC3_shots1_9_noise_4.5.sgy	~951 MB × 2
5.0	SEGC3_shots1_9_noisy_5.0.sgy	SEGC3_shots1_9_noise_5.0.sgy	~951 MB × 2
7.0	SEGC3_shots1_9_noisy_7.0.sgy	SEGC3_shots1_9_noise_7.0.sgy	~951 MB × 2
9.0	SEGC3_shots1_9_noisy_9.0.sgy	SEGC3_shots1_9_noise_9.0.sgy	~951 MB × 2

Total: 9 noisy + 9 noise SEG-Y files

Loading Data

import segyio
import numpy as np

def read_shot_gather(path, traces_per_shot=201):
    '''Read a regular SEG-Y file into (n_shots, n_traces, n_time).'''
    with segyio.open(path, "r", strict=False) as src:
        n_traces_total = src.tracecount
        n_shots = n_traces_total // traces_per_shot
        n_time = src.samples.size
        data = np.zeros((n_shots, traces_per_shot, n_time), dtype=np.float32)
        for i in range(n_shots):
            for j in range(traces_per_shot):
                data[i, j, :] = src.trace[i * traces_per_shot + j]
    return data

# Load a level-3.0 pair
noisy  = read_shot_gather("noisy/SEGC3_shots1_9_noisy_3.0.sgy")
noise  = read_shot_gather("noise/SEGC3_shots1_9_noise_3.0.sgy")
signal = noisy - noise   # clean reference

With huggingface_hub:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="GeoBrain/ground-roll",
    filename="noisy/SEGC3_shots1_9_noisy_3.0.sgy",
    repo_type="dataset",
)

Train / Val / Test Split

Shot-level sequential split by FFID (field file ID), avoiding trace leakage:

Split	Shots	Fraction
Train	7	77.8%
Val	1	11.1%
Test	1	11.1%

The split is done at loading time (not pre-saved as separate files) so users can adjust the ratios.

Preprocessing Recipe

The companion benchmark applies:

1. Normalization: max_abs, global scope — the entire noisy volume scaled to [-1, 1]; same stats applied to the noise label
2. Patching: Overlapping 2D patches (128 traces × 256 time samples), 50% overlap, yielding (1, H, W) tensors

No spherical-divergence correction is applied (raw amplitudes are used).

Benchmark Results

See the companion model repository for full benchmark results across UNet, ResUNet, DnCNN, and Attention UNet architectures at each noise level.

Citation

If you use this dataset, please cite the SEG C3 model and the companion benchmark:

@misc{seg_c3_ground_roll,
  title={SEG C3 Ground-Roll Attenuation Benchmark},
  howpublished={https://huggingface.co/datasets/GeoBrain/ground-roll},
}

References

• SEG C3 Velocity Model: https://wiki.seg.org/wiki/C3
• segyio library: https://github.com/equinor/segyio