Marine Multiples Attenuation Dataset

Paired noisy-input / multiples-noise-label SEG-Y volumes for supervised marine multiples attenuation.

Task

Noise-label regression: given a noisy pre-stack shot gather, predict the additive multiples component. The denoised signal is recovered as:

denoised = noisy_input - predicted_multiples

The uploaded noise label is the supervised target. The clean reference used by the benchmark is computed as:

clean_reference = noisy_input - multiples_label

Example 1

Noisy Data

Clean Data

Multiples Noise Label

representative multiples attenuation sample.

Example 2

Noisy Data

Clean Data

Multiples Noise Label

representative multiples attenuation sample.

Example 3

Noisy Data

Clean Data

Multiples Noise Label

representative multiples attenuation sample.

File Structure

Kind	Path	Size
noise	noise/multiples.sgy	3161.4 MB
noisy	noisy/total_nodw.sgy	3161.4 MB

Total: 1 noisy + 1 noise SEG-Y files

Loading Data

import segyio
import numpy as np

def read_shot_gather(path, traces_per_shot=638):
    '''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

noisy = read_shot_gather("noisy/total_nodw.sgy", traces_per_shot=638)
multiples = read_shot_gather("noise/multiples.sgy", traces_per_shot=638)
clean_reference = noisy - multiples

With huggingface_hub:

from huggingface_hub import hf_hub_download

noisy_path = hf_hub_download(
    repo_id="GeoBrain/multiples",
    filename="noisy/total_nodw.sgy",
    repo_type="dataset",
)

Benchmark Split

The companion benchmark uses shot-level FFID splitting to avoid trace leakage. The current multiples configs use:

Split	Shots
Train	510
Val	64
Test	64

The split is done at loading time, so users can adjust it in their own configs.

Preprocessing Recipe

The companion benchmark applies:

1. Normalization: max_abs, global scope on the noisy input; the same scale is applied to the multiples label.
2. Patching: overlapping 2D patches on the trace-time plane.
3. Metric handling: SNR can skip near-zero clean-reference patches with min_signal_energy.

No spherical-divergence correction is applied in the denoising training script.

Citation

If you use this dataset, cite the dataset repository:

@misc{marine_multiples_attenuation,
  title={Marine Multiples Attenuation Dataset},
  howpublished={https://huggingface.co/datasets/GeoBrain/multiples},
}

References

• segyio library: https://github.com/equinor/segyio