GeoBrain/deblending-common-receiver

Deblending Common-Receiver Dataset

Paired common-receiver SEG-Y volumes for supervised seismic deblending.

Task

This dataset is used for direct paired regression:

denoised = model(input)

Where:

• input is the pseudo-deblended common-receiver gather volume
• target is the cleaner forwarding-acoustic common-receiver gather volume

This is not a noise-label regression dataset.

File Structure

Kind	Path	Size
input	input/T02_pseudo_deblended_common_receiver_mod.sgy	819.0 MB
target	target/T02_forwarding_acoustic_common_receiver.sgy	819.0 MB

Loading Data

import numpy as np
import segyio


def read_common_receiver_volume(path, traces_per_gather=270):
    with segyio.open(path, "r", ignore_geometry=True) as src:
        n_traces_total = src.tracecount
        n_time = src.samples.size
        n_gathers = n_traces_total // traces_per_gather
        data = segyio.tools.collect(src.trace[:]).astype(np.float32, copy=False)
        data = data.reshape(n_gathers, traces_per_gather, n_time)
    return data


input_volume = read_common_receiver_volume(
    "input/T02_pseudo_deblended_common_receiver_mod.sgy",
    traces_per_gather=270,
)
target_volume = read_common_receiver_volume(
    "target/T02_forwarding_acoustic_common_receiver.sgy",
    traces_per_gather=270,
)

With huggingface_hub:

from huggingface_hub import hf_hub_download

input_path = hf_hub_download(
    repo_id="GeoBrain/deblending-common-receiver",
    filename="input/T02_pseudo_deblended_common_receiver_mod.sgy",
    repo_type="dataset",
)

Benchmark Split

The current benchmark uses a sequential gather split:

Split	Gathers
Train	300
Val	43
Test	43

The split is applied on common-receiver gather index.

Preprocessing Recipe

The companion benchmark currently applies:

1. shared max_abs normalization between input and target
2. per-gather normalization through the shared shot normalization scope
3. overlapping 2D trace-time patches

Current patch settings:

• patch_trace: 128
• patch_time: 256
• patch_overlap: 0.5

Notes

• The benchmark uses common-receiver gathers, not common-shot gathers.
• Input and target must keep identical ordering and identical shape.
• The current benchmark reads these files by fixed sequential grouping with 270 traces per gather.

Citation

If you use this dataset, cite the dataset repository:

@misc{deblending_common_receiver_dataset,
  title={Deblending Common-Receiver Dataset},
  howpublished={https://huggingface.co/datasets/GeoBrain/deblending-common-receiver},
}

References

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