SWST / README.md
AlumniBBQ's picture
Upload README.md with huggingface_hub
7b2d1f8 verified
|
Raw History Blame Contribute Delete
3.06 kB
metadata
license: cc-by-nc-4.0
library_name: pytorch
tags:
  - seismic-denoising
  - surface-wave-suppression
  - ground-roll
  - self-supervised-pretraining
  - transformer

SWST — Surface-Wave Suppression Transformer

Trained checkpoint for SWST, a two-stage Transformer for suppressing surface waves (ground roll) in land seismic data:

  1. Self-supervised pretraining: a deterministic low-frequency spatiotemporal-energy detector locates candidate surface-wave patches in unlabeled field records; a sparse masked autoencoder with SwinV2-style window attention reconstructs masked patches and learns surface-wave morphology and spatial continuity without paired clean/noisy data.
  2. Supervised fine-tuning: the sparse encoder is converted into a parameter-compatible dense encoder that predicts the complete surface-wave component; a convolutional residual head refines patch boundaries and local waveforms. The restored record is clip(input − predicted noise).

Model description

Architecture mae_swinv2_sparse_refine_base_patch16 (sparse SwinV2 encoder + MAE decoder + refinement head)
Parameters 111,554,353
Checkpoint format PyTorch-Lightning final.ckpt
File size 881,285,912 bytes
SHA-256 3cf478d283dd5601aa34924d256863e447f8b4eb2df5ed2d70f5b674465b584a
Input single-channel seismic section, normalized to [-1, 1]
Output predicted surface-wave (noise) component, same shape
License CC BY-NC 4.0

Quick start

A self-contained reproducible example (input data, inference script with the manuscript's sliding-window protocol, and expected outputs for verification) is available in the repro_example/ folder of the accompanying code repository referenced in the manuscript Data Statement.

Minimal usage with the source package on disk:

import torch
from MAE.models_restoration import get_mae_restoration_model

model = get_mae_restoration_model(
    "mae_swinv2_sparse_refine_base_patch16", img_size=256, in_chans=1
).eval()
ckpt = torch.load("final.ckpt", map_location="cpu")
state = {k[len("model."):]: v for k, v in ckpt["state_dict"].items()
         if k.startswith("model.")}
model.load_state_dict(state, strict=True)

section = torch.randn(1, 1, 256, 256)        # your normalized section patch
predicted_noise = model(section)             # [1, 1, 256, 256]
restored = torch.clamp(section - predicted_noise, -1.0, 1.0)

For full 2000 x 256 records use the sliding-window inference from the repro_example/ scripts (256-pixel windows, stride 128, separable Hann blending).

Training data

Pretraining used unlabeled field records (500 Hz sampling) from multiple survey areas; fine-tuning used synthetic pairs built from surface-wave-free references and field-derived surface-wave components. No confidential field data is required to use or evaluate the checkpoint.

Intended use

Research on seismic noise attenuation and reproducible evaluation of the manuscript's results. Not validated for production processing.