SeisDispFusion-NCF

Dataset creators: Xin Liu, Ziye Yu This dataset was developed by Xin Liu and Ziye Yu for research on seismic dispersion analysis, integrating noise cross-correlation (NCF) waveforms, dispersion curves, and 2D dispersion images for machine learning applications.

Overview

SeisDispFusion-NCF is a multimodal seismic dataset organized by station pairs.
It integrates dispersion curves, noise cross-correlation (NCF) waveforms, and 2D dispersion images into a unified HDF5 structure.

The dataset is designed to support data-driven seismic inversion, uncertainty estimation, and multimodal learning tasks.

---

Key Features

• Station-pair-based organization (A–B and B–A unified)
• Multimodal data fusion
  • Dispersion curves (period–velocity)
  • Noise cross-correlation waveforms (NCF)
  • 2D dispersion images (frequency/period–velocity)
• Uncertainty-aware representation
  • Dispersion masks indicating valid observations
• Multiple observations per path
  • Multiple NCF waveforms per station pair
  • Multiple dispersion-image representations (both A_B and B_A preserved)
• Train/Test split included

---

Dataset Variants

Two versions of the dataset are provided to support different use cases:

1. ncf_disp_dataset.h5

• Contains:
  • Dispersion curves (disp_periods, disp_velocity, disp_mask)
  • Noise cross-correlation waveforms (NCF)
• Suitable for:
  • Classical inversion workflows
  • Lightweight machine learning models
  • Curve-based regression tasks

---

2. ncf_disp_dataset_with_disp_image.h5

• Contains:
  • All contents of the base dataset
  • 2D dispersion images (disp_image)
• Suitable for:
  • CNN / Transformer-based models
  • Representation learning
  • Uncertainty-aware modeling using dispersion energy maps

---

Summary

Dataset	Dispersion	NCF	2D Image
ncf_disp_dataset.h5	✓	✓	✗
ncf_disp_dataset_with_disp_image.h5	✓	✓	✓

Dataset Structure

The dataset is stored in HDF5 format:

root
├── train_keys
├── test_keys
└── paths
└── <STA1-STA2>
├── disp_periods        (float32)  [N]
├── disp_velocity       (float32)  [N]
├── disp_mask           (uint8)    [N]
├── disp_sta1_coord     (float32)  [2]
├── disp_sta2_coord     (float32)  [2]

├── disp_image
│   ├── pair_order_list
│   ├── item_0000
│   │   ├── image_2d    (float32)  [F, V]
│   │   ├── period      (float32)
│   │   ├── freq        (float32)
│   │   └── velocity    (float32)
│   └── item_0001
│       └── ...

└── ncf
    ├── item_0000
    │   ├── time        (float32)
    │   ├── waveform    (float32)
    │   ├── sta1_coord  (float32)
    │   └── sta2_coord  (float32)
    └── item_0001

---

Data Description

1. Dispersion Curves

• disp_periods: standardized period axis (2–50 s)
• disp_velocity: phase velocity values
• disp_mask: indicates valid measurements (1 = valid, 0 = missing)

This representation enables:

• regression tasks
• uncertainty-aware training
• masked learning

---

2. Noise Cross-Correlation (NCF)

Each station pair may contain multiple NCF observations:

• time: time axis
• waveform: cross-correlation amplitude

These data capture wave propagation information and can be used for:

• travel-time learning
• waveform-based inversion
• representation learning

---

3. 2D Dispersion Images

Each station pair may include multiple dispersion-image representations:

• image_2d: dispersion energy map
• period / freq: horizontal axis (depending on representation)
• velocity: vertical axis

Important:

• Both A_B and B_A orientations are preserved
• Each is stored as an independent item_xxxx

This allows:

• CNN/Transformer-based feature learning
• uncertainty modeling from dispersion energy
• comparison between symmetric paths

---

Data Splits

• train_keys: list of training station pairs
• test_keys: list of testing station pairs

Splitting is performed at the station-pair level, ensuring independence.

---

Intended Use

This dataset is suitable for:

1. Seismic Inversion

• Dispersion-based inversion
• Waveform-based inversion
• Hybrid inversion (curve + image + waveform)

2. Machine Learning

• Multimodal learning
• Representation learning
• Transformer/CNN models on dispersion images
• Variational models (VAE) for uncertainty estimation

3. Uncertainty Quantification

• Masked regression
• Heteroscedastic modeling
• Bayesian inference

---

Design Philosophy

This dataset is built on the idea that:

A larger dataset does not necessarily imply more useful information.

Instead, it emphasizes:

• data quality representation (mask)
• multiple observational views (multimodal)
• structural interpretability

---

Limitations

• No absolute ground truth for real seismic structure
• Dispersion curves may contain missing or noisy values
• 2D dispersion images depend on preprocessing methods
• NCF quality varies across station pairs

---

Data Loaders

Two PyTorch data loaders are provided in the utils/ directory:

• utils/dispdataset.py
• utils/dispdataset2d.py

These loaders support two related but different learning settings built on top of the HDF5 datasets.

---

1. utils/dispdataset.py

Purpose

dispdataset.py provides a PyTorch Dataset and DataLoader for learning from:

• noise cross-correlation (NCF) waveforms
• 1D dispersion curves

It is designed for waveform-to-dispersion tasks or other models that use waveform input and dispersion-vector targets. :contentReference[oaicite:2]{index=2}

Main Class

• NCFDispersionDataset
• helper function: build_dataloader(...) :contentReference[oaicite:3]{index=3} :contentReference[oaicite:4]{index=4}

Input File

This loader is intended for the base HDF5 dataset:

• data/ncf_disp_dataset.h5

It can also be used on the image-enhanced HDF5 file if the required waveform/dispersion fields are present. :contentReference[oaicite:5]{index=5}

Features

• Supports split="train" and split="test"
• Reads one waveform sample per station pair
• If multiple NCF items are available for the same path, one can be selected randomly
• Pads or truncates waveforms to a fixed length
• Returns dispersion velocity, period axis, and mask
• Returns station coordinates
• Optionally returns the NCF time axis :contentReference[oaicite:6]{index=6} :contentReference[oaicite:7]{index=7} :contentReference[oaicite:8]{index=8}

Returned Fields

Each sample contains:

• key: station-pair key
• waveform: tensor of shape [waveform_length]
• disp: tensor of shape [49]
• mask: tensor of shape [49]
• periods: tensor of shape [49]
• sta1_coord: tensor of shape [2]
• sta2_coord: tensor of shape [2]
• ncf_time: tensor of shape [waveform_length] if return_time=True :contentReference[oaicite:9]{index=9}

Example

from utils.dispdataset import build_dataloader

loader = build_dataloader(
    h5_path="data/ncf_disp_dataset.h5",
    split="train",
    batch_size=16,
    num_workers=0,
    waveform_length=1536,
    random_ncf=True,
)

batch = next(iter(loader))

print(batch["waveform"].shape)   # [B, 1536]
print(batch["disp"].shape)       # [B, 49]
print(batch["mask"].shape)       # [B, 49]
print(batch["periods"].shape)    # [B, 49]

Notes

• Waveforms are padded or truncated to a fixed length.
• Dispersion labels are stored as 1D vectors over the standardized period axis.
• If a path contains multiple NCF observations, the loader can randomly select one during training.

---

2. utils/dispdataset2d.py

Purpose

dispdataset2d.py provides a PyTorch Dataset and DataLoader for learning from:

• 2D dispersion images
• Gaussian soft labels constructed from 1D dispersion curves
• dispersion curves and masks

It is designed for image-based or multimodal learning on FTAN / dispersion-image representations.

Main Class

• DispImageDataset
• helper function: build_dataloader(...)
• visualization helper: plot_sample(...)

Input File

This loader is intended for the image-enhanced HDF5 dataset:

• data/ncf_disp_dataset_with_disp_image.h5

Features

• Supports split="train" and split="test"
• Requires valid disp_image entries
• Requires valid dispersion labels when require_valid_label=True
• Randomly selects one FTAN / dispersion-image item if multiple items are available for the same station pair
• Interpolates the original 2D image onto a unified [V, T] grid
• Builds Gaussian soft labels centered at the dispersion velocity for each valid period
• Can optionally return the image with channel dimension [1, V, T]

Label Construction

For each period index t:

• if disp_mask[t] == 0 or disp_velocity[t] <= 0, the entire label column is set to zero
• otherwise, a Gaussian distribution is constructed on the velocity axis, centered at disp_velocity[t]

This produces a 2D target map of shape [V, T].

Returned Fields

Each sample contains:

• image: tensor of shape [1, V, T] or [V, T]
• label: tensor of shape [V, T]
• mask_1d: tensor of shape [T]
• disp_periods: tensor of shape [T]
• disp_velocity: tensor of shape [T]
• velocity_axis: tensor of shape [V]
• key: station-pair key
• ftan_index: selected FTAN item index

Example

from utils.dispdataset2d import build_dataloader

loader = build_dataloader(
    h5_path="data/ncf_disp_dataset_with_disp_image.h5",
    split="train",
    batch_size=4,
    num_workers=0,
    velocity_min=1.0,
    velocity_max=5.0,
    num_velocity_bins=256,
    label_sigma=0.08,
    random_ftan=True,
    require_valid_label=True,
)

batch = next(iter(loader))

print(batch["image"].shape)         # [B, 1, V, T]
print(batch["label"].shape)         # [B, V, T]
print(batch["disp_periods"].shape)  # [B, T]
print(batch["disp_velocity"].shape) # [B, T]

Visualization Example

dispdataset2d.py also provides a plotting utility:

from utils.dispdataset2d import build_dataloader, plot_sample

loader = build_dataloader(
    h5_path="data/ncf_disp_dataset_with_disp_image.h5",
    split="train",
    batch_size=4,
    num_workers=0,
)

batch = next(iter(loader))

plot_sample(
    image=batch["image"][0],
    label=batch["label"][0],
    disp_periods=batch["disp_periods"][0],
    disp_velocity=batch["disp_velocity"][0],
    velocity_axis=batch["velocity_axis"][0],
    disp_mask=batch["mask_1d"][0],
    key=batch["key"][0],
    ftan_index=int(batch["ftan_index"][0]),
)

This function plots:

1. the input dispersion image
2. the Gaussian target label
3. the dispersion curve

---

Which Loader Should I Use?

Use utils/dispdataset.py when:

• your model takes NCF waveform as input
• your target is a 1D dispersion vector

Use utils/dispdataset2d.py when:

• your model takes a 2D dispersion image as input
• your target is a 2D soft label map
• you want to use FTAN / image-based learning workflows

---

Practical Notes

• On macOS and Python 3.12, it is recommended to test with num_workers=0 first.
• dispdataset2d.py filters out samples without valid image-based inputs and valid dispersion labels.
• dispdataset.py is the lightweight loader for waveform-based experiments, while dispdataset2d.py is intended for image-based multimodal learning.

---

Dataset Attribution

This dataset is derived from the following original dataset:

Feng, Jikun (2021),
“Ambient noise cross-correlations in the North China Craton”,
Mendeley Data, V1.
https://doi.org/10.17632/m9ry8nbfwj.1

We acknowledge the original author for providing the ambient noise cross-correlation data.
The current dataset has been further processed and extended, including:

• Reorganization into station-pair based HDF5 structure
• Extraction and standardization of dispersion curves
• Integration of noise cross-correlation (NCF) waveforms
• Construction of 2D dispersion images (FTAN-like representations)
• Generation of machine learning-ready labels and metadata

Users of this dataset should cite both the original dataset and this processed version where appropriate.

---

Citation

If you use this dataset, please cite:

Yu, Z., & Liu, X. (2025).
A Framework for Uncertainty Estimation in Seismology Data Processing with Application to Extract Rayleigh Wave Dispersion Curves from Noise Cross-correlation Functions.
arXiv:2503.20460
https://doi.org/10.48550/arXiv.2503.20460

This dataset is developed as part of the above work, which introduces an uncertainty-aware framework for extracting Rayleigh wave dispersion curves directly from noise cross-correlation functions (NCFs), along with posterior uncertainty estimation.

Original data source:

Feng, J. (2021).
Ambient noise cross-correlations in the North China Craton.
Mendeley Data, V1.
https://doi.org/10.17632/m9ry8nbfwj.1

---

License

This dataset is released under the apache-2.0 License.

---

Contact

For questions or collaborations:

• Author: Ziye Yu, yuziye@cea-igp.ac.cn; Xin Liu