Marine Multiples Attenuation Benchmark
Deep-learning-based free-surface multiple attenuation on marine pre-stack seismic shot gathers.
Task
Given a shot gather containing primaries and free-surface multiples, the model predicts the additive multiple component. The attenuated primary estimate is obtained by:
denoised = noisy_input - predicted_noise
This is a paired regression task trained with a noise-label objective. The supervision target is the multiple wavefield stored in the paired noise-label volume.
Dataset
- Input volume:
/data/shared/benchmark/multiples/noisy/total_nodw.sgy - Multiple-label volume:
/data/shared/benchmark/multiples/noise/multiples.sgy - Geometry: regular shot gathers with 638 traces per shot
- Split: shot-level sequential split from the training configs, typically 510 train shots, 64 validation shots, and 64 held-out test shots
The uploaded checkpoints are trained on the fixed paired marine multiples benchmark used by scripts/multiples_attenuation.
Metrics are computed on the held-out test shots in the normalized domain. If a batch-evaluation workbook is supplied, its values are embedded below.
Model Architectures
- DNNDAT (
dnndat) — DNNDAT-style convolutional encoder-decoder for marine multiple suppression (Wang et al., 2022). U-Net-like encoder-decoder with 28 convolutional layers and dropout. - SAGAN (
sagan) — Self-attention GAN generator for seismic surface-related multiple suppression (Tao et al., 2022). U-Net generator with a bottleneck self-attention block.
Uploaded Checkpoints
- DNNDAT: 4 checkpoints
- SAGAN: 3 checkpoints
Preprocessing
- Normalization:
max_abs, global scope — the entire dataset scaled to [-1, 1] - Patching: overlapping 2D patches, usually 256 traces × 512 time samples with 50% overlap
- Tensor format: PyTorch NCHW patches
(batch, 1, trace, time)
Repository Structure
models/
├── unet/
│ ├── seed42/
│ │ ├── best.pt # Best checkpoint (minimum validation loss)
│ │ └── config.yaml # Full training configuration
│ ├── seed43/
│ ├── seed44/
│ └── ...
└── res_unet/
└── ...
Each subdirectory corresponds to one experiment: a model architecture trained with a specific random seed.
Training Details
| Hyperparameter | Value |
|---|---|
| Loss | MSE on the predicted multiple/noise component |
| Optimizer | Adam / AdamW / SGD, depending on model config |
| Scheduler | Cosine annealing (min_lr=1e-6) |
| Epochs | 100-200, depending on model config |
| Gradient clipping | 1.0 (max norm) |
| Seeds | 42, 43, 44 per experiment |
Usage
import torch
from huggingface_hub import hf_hub_download
# Download a checkpoint
repo = "GeoBrain/multiples-attenuation"
model_key = "res_unet"
seed = "42"
ckpt_path = hf_hub_download(
repo_id=repo,
filename=f"models/{model_key}/seed{seed}/best.pt",
)
# Load state dict
state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)
# For full model loading, instantiate the corresponding architecture
# and load the state dict (see config.yaml for exact architecture params).
See the companion benchmark documentation for detailed experimental setup and full evaluation results.
Results
Results pending — run batch_evaluate.py to populate.
References
- Ronneberger et al., U-Net: Convolutional Networks for Biomedical Image Segmentation, MICCAI 2015
- He et al., Deep Residual Learning for Image Recognition, CVPR 2016
- Zhang et al., Image Denoising via Deep CNN (DnCNN), IEEE TIP 2017
- Oktay et al., Attention U-Net: Learning Where to Look for the Pancreas, MIDL 2018
- Kiraz et al., Attenuating free-surface multiples and ghost reflection from seismic data using a trace-by-trace convolutional neural network approach, Geophysical Prospecting 2024
- Tao et al., Seismic Surface-Related Multiples Suppression Based on SAGAN, IEEE Geoscience and Remote Sensing Letters 2022
- Wang et al., Seismic multiple suppression based on a deep neural network method for marine data, Geophysics 2022