GeoBrain/random-noise-seg-c3-clean

SEG C3 Clean Seismic Dataset

Clean pre-stack SEG-Y shot gathers used as the source volume for the random-noise-suppression benchmark. Synthetic Gaussian and Poisson noise is added during training and inference by the benchmark scripts; this repository stores the clean reference volume only.

Source

The clean seismic data is derived from the public SEG C3 synthetic benchmark dataset.

Preview Images

clean

Dataset File

• File: data/SEG_45Shot_shots1-9.sgy
• Size: 950.1 MB
• Shots: 9
• Traces per shot: 201
• Samples per trace: 625
• Sample interval: 0.002 s
• Total traces: 363609

Benchmark Usage

The random-noise-suppression configs expect this file at:

/root/Desktop/data/SEG_45Shot_shots1-9.sgy

The benchmark reads the SEG-Y as a regular shot-gather volume:

from tools.segy_read import read_regular_shots

shots = read_regular_shots(
    "data/SEG_45Shot_shots1-9.sgy",
    traces_per_shot=201,
    time_downsample=1,
)
print(shots.shape)  # (9, 201, 625)

Training uses shot-level splitting:

train: 7 shots
val:   1 shot
test:  1 shot

Preprocessing uses per-shot max_abs normalization, dt=0.002, patch_trace=128, and patch_time=256.

Citation

If you use this dataset, cite the dataset repository:

@misc{seg_c3_clean_seismic_dataset,
  title={SEG C3 Clean Seismic Dataset},
  howpublished={https://huggingface.co/datasets/GeoBrain/random-noise-seg-c3-clean},
}