GeoBrain/seismic-interpolation-benchmark

Seismic Trace Interpolation Benchmark

Deep-learning-based missing-trace reconstruction on pre-stack seismic shot gathers. Given a shot gather with masked (missing) traces, the model reconstructs the full gather. Each experiment is one architecture trained on one missing-trace scenario with one random seed.

Datasets

Model directories are suffixed with the dataset the model was trained on:

• *_mobil — Mobil field dataset (pre-stack seismic shot gathers)
• *_seg_c3 — SEG C3 synthetic dataset (wiki.seg.org/wiki/C3): 9 regular shots, 201 traces x 625 time samples, dt = 2 ms

Per-dataset benchmark results: results_mobil.md | results_seg_c3.md

Models

• Chai2020 UNet (chai2020_unet) — 2D U-Net (Chai et al., 2020, IEEE TGRS, DOI 10.1109/TGRS.2019.2961015): 50 layers = 19 convolutions (5x5, same padding) + 18 ReLU + 4 max-pool + 4 upsample + 4 concat, 1x1 output convolution, for regularly missing data reconstruction.
• CA-Unet (li2022_caunet) — CA-Unet (Li et al., 2022, IEEE LGRS, DOI 10.1109/LGRS.2021.3128511): U-Net encoder-decoder with Coordinate Attention blocks (directional H/W pooling, shared 1x1 conv, H-Swish, split, Sigmoid gating). Trained with hybrid SSIM + L1 loss.
• WRDL (liu2022_wrdl) — WRDL (Liu et al., 2022, IEEE TGRS, DOI 10.1109/TGRS.2022.3152984): wavelet-based residual deep learning for seismic data reconstruction.
• PConv U-Net (pan2020_pconv_unet) — PConv U-Net (Pan et al., 2020, Computers & Geosciences, DOI 10.1016/j.cageo.2020.104609): partial-convolution U-Net for seismic data regularization.
• CFunet (park2022_cfunet) — CFunet (Park et al., 2022, IEEE TGRS, DOI 10.1109/TGRS.2022.3190292): coarse-refine network with upsampling techniques and Fourier loss for missing-trace reconstruction.
• Gated Transformer v9 (gated_transformer_v9) — In-house gated transformer (v9) baseline; evaluated on the same scenarios as the paper baselines.
• ANet (yu2022_anet) — ANet (Yu and Wu, 2022, IEEE TGRS, DOI 10.1109/TGRS.2021.3068279): two stride-2 downsampling convolutions, six residual blocks, one non-local attention module, two upsampling + convolution groups. Trained with hybrid SSIM + L1 loss.

Model availability differs per dataset: the Mobil benchmark covers ANet and CA-Unet; the SEG C3 benchmark covers all architectures above except PConv U-Net. The park2022_cfunet_continuous family (SEG C3 only) is a legacy continuous-only training variant of CFunet kept in its own subtree.

Missing-Trace Scenarios

• Uniform masking (30% / 50% / 70% missing, every k-th trace kept)
• Random trace dropout (30% / 50%)
• Consecutive missing-trace blocks (20 / 30 / 40 traces)
• Scenario availability follows each paper's setting; see each config.yaml for exact mask parameters

Preprocessing

Spherical divergence correction (power = 1.2, except where a paper setting skips it), global max_abs normalization to [-1, 1], overlapping (trace x time) patches. See each config.yaml for the exact objective (hybrid SSIM + L1 for the attention models).

Repository Structure

models/
├── yu2022_anet_mobil/            # Mobil field-dataset benchmark
│   └── uniform_miss50/seed42/
│       ├── best.pt               # Best checkpoint (minimum validation loss)
│       └── config.yaml           # Full training configuration
├── yu2022_anet_seg_c3/           # SEG C3 benchmark
│   └── ...
└── ...

Usage

import torch
from huggingface_hub import hf_hub_download

repo = "GeoBrain/seismic-interpolation-benchmark"
ckpt_path = hf_hub_download(
    repo_id=repo,
    filename="models/yu2022_anet_seg_c3/uniform_miss50/seed42/best.pt",
)
state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)

# Instantiate the architecture with the params from the sibling config.yaml
# and load the state dict.

References

• Chai et al., "Deep Learning for Regularly Missing Data Reconstruction", IEEE TGRS, 2020. DOI: 10.1109/TGRS.2019.2961015
• Li et al., "CA-Unet: Coordinate Attention U-Net for Seismic Data Reconstruction", IEEE LGRS, 2022. DOI: 10.1109/LGRS.2021.3128511
• Liu et al., "Seismic Data Reconstruction via Wavelet-Based Residual Deep Learning", IEEE TGRS, 2022. DOI: 10.1109/TGRS.2022.3152984
• Pan et al., "A Partial Convolution-Based Deep-Learning Network for Seismic Data Regularization", Computers & Geosciences, 2020. DOI: 10.1016/j.cageo.2020.104609
• Park et al., "Coarse-Refine Network With Upsampling Techniques and Fourier Loss for the Reconstruction of Missing Seismic Data", IEEE TGRS, 2022. DOI: 10.1109/TGRS.2022.3190292
• Yu and Wu, "Attention and Hybrid Loss Guided Deep Learning for Consecutively Missing Seismic Data Reconstruction", IEEE TGRS, 2022. DOI: 10.1109/TGRS.2021.3068279