GeoBrain/first-break-picking-segy-with-masks

First-Break Picking SEG-Y Dataset with Masks

This dataset contains prestack seismic SEG-Y records and their corresponding first-break picking masks. It is designed for supervised first-break picking and first-arrival detection in seismic data.

Task

First-break picking with mask supervision: given a prestack seismic record, predict a mask that indicates the first-break / first-arrival position.

The dataset is organized as paired SEG-Y files:

• data/: input seismic SEG-Y files
• label/: corresponding first-break mask SEG-Y files

Each input file has a corresponding mask file with matched geometry.

The masks can be used as labels for supervised learning models. Depending on the downstream implementation, the mask can be interpreted as a binary / soft label map over trace-time samples.

In the companion benchmark, the mask is used as a binary step-mask target: samples before the first-break position are labeled as 0, and samples from the first-break position onward are labeled as 1.

File Structure

.
├── data/
│   ├── Brunswick_valid.sgy
│   ├── Halfmile_valid.sgy
│   └── Lalor_valid.sgy
└── label/
    ├── Brunswick_valid_mask.sgy
    ├── Halfmile_valid_mask.sgy
    └── Lalor_valid_mask.sgy

File Summary

Dataset	Input SEG-Y	Label SEG-Y	Traces	Samples/Trace	Sample Interval	Input Size	Label Size
Brunswick	data/Brunswick_valid.sgy	label/Brunswick_valid_mask.sgy	4,490,714	751	2 ms	13.57 GiB	13.57 GiB
Halfmile	data/Halfmile_valid.sgy	label/Halfmile_valid_mask.sgy	1,093,842	751	2 ms	3.30 GiB	3.30 GiB
Lalor	data/Lalor_valid.sgy	label/Lalor_valid_mask.sgy	2,027,587	1,501	1 ms	11.79 GiB	11.79 GiB

Total listed dataset size is approximately 57.3 GiB.

Benchmark Usage

The companion benchmark treats first-break picking as a segmentation-style prediction problem over trace-time samples.

Common preprocessing and training settings used by the released models include:

data:
  data_dir: data
  label_dir: label
  label_threshold: 0.5
  prediction_threshold: 0.5
  split:
    train: 0.8
    val: 0.1
    test: 0.1
    shuffle_ffids: true
  patch:
    trace: 128
    time: 512
    trace_stride: 64
    time_stride: 256

preprocess:
  normalize_mode: max_abs
  normalize_scope: gather
  clip_percentile: 99.5

For the combined benchmark setting, the listed SEG-Y pairs are used together. For single-dataset experiments, one input SEG-Y file and its corresponding mask are used at a time.

Loading Example

from pathlib import Path

import numpy as np
import segyio


def read_segy(path):
    path = Path(path)
    with segyio.open(str(path), "r", ignore_geometry=True) as src:
        traces = np.stack([np.asarray(trace, dtype=np.float32) for trace in src.trace])
        samples = np.asarray(src.samples)
    return traces, samples


amplitude, time_samples = read_segy("data/Brunswick_valid.sgy")
mask, _ = read_segy("label/Brunswick_valid_mask.sgy")

print(amplitude.shape)
print(mask.shape)

With huggingface_hub:

from huggingface_hub import hf_hub_download

repo_id = "GeoBrain/first-break-picking-dataset"

input_path = hf_hub_download(
    repo_id=repo_id,
    filename="data/Brunswick_valid.sgy",
    repo_type="dataset",
)

mask_path = hf_hub_download(
    repo_id=repo_id,
    filename="label/Brunswick_valid_mask.sgy",
    repo_type="dataset",
)

Related Model Release

The corresponding trained model checkpoints are published separately:

GeoBrain/first-break-picking

---

初至拾取 SEG-Y 数据集及掩码标签

本数据集包含叠前地震 SEG-Y 记录及其对应的初至拾取掩码标签，面向监督式初至拾取和地震初至到达检测任务。

任务

基于掩码监督的初至拾取：给定一份叠前地震记录，模型需要预测一个表示初至 / 首波到达位置的掩码。

数据集以成对 SEG-Y 文件组织：

• data/：输入地震 SEG-Y 文件
• label/：对应的初至拾取 mask SEG-Y 文件

每个输入文件都有一个几何匹配的 mask 文件。

数据集说明

• 领域：勘探地球物理 / 地震资料处理
• 任务：初至拾取 / 首波到达检测
• 输入格式：SEG-Y 地震记录
• 标签格式：SEG-Y 初至拾取掩码
• 数据类型：叠前地震道集
• 监督方式：成对的输入 SEG-Y 与 mask SEG-Y 文件

这些 mask 可作为监督学习模型的标签。根据下游实现方式，mask 可以被解释为 trace-time 采样点上的二值或软标签图。

在配套 benchmark 中，mask 被用作二值 step-mask 标签：初至位置之前的采样点标为 0，从初至位置开始及其之后的采样点标为 1。

文件结构

.
├── data/
│   ├── Brunswick_valid.sgy
│   ├── Halfmile_valid.sgy
│   └── Lalor_valid.sgy
└── label/
    ├── Brunswick_valid_mask.sgy
    ├── Halfmile_valid_mask.sgy
    └── Lalor_valid_mask.sgy

文件概览

数据集	输入 SEG-Y	标签 SEG-Y	道数	每道采样点	采样间隔	输入大小	标签大小
Brunswick	data/Brunswick_valid.sgy	label/Brunswick_valid_mask.sgy	4,490,714	751	2 ms	13.57 GiB	13.57 GiB
Halfmile	data/Halfmile_valid.sgy	label/Halfmile_valid_mask.sgy	1,093,842	751	2 ms	3.30 GiB	3.30 GiB
Lalor	data/Lalor_valid.sgy	label/Lalor_valid_mask.sgy	2,027,587	1,501	1 ms	11.79 GiB	11.79 GiB

当前列出的数据总大小约为 57.3 GiB。

Benchmark 使用方式

配套 benchmark 将初至拾取建模为 trace-time 采样点上的分割式预测问题。

已发布模型使用的常见预处理和训练设置如下：

data:
  data_dir: data
  label_dir: label
  label_threshold: 0.5
  prediction_threshold: 0.5
  split:
    train: 0.8
    val: 0.1
    test: 0.1
    shuffle_ffids: true
  patch:
    trace: 128
    time: 512
    trace_stride: 64
    time_stride: 256

preprocess:
  normalize_mode: max_abs
  normalize_scope: gather
  clip_percentile: 99.5

综合 benchmark 设置会同时使用当前列出的成对 SEG-Y 文件。单数据集实验则每次只使用一个输入 SEG-Y 文件及其对应 mask。

读取示例

from pathlib import Path

import numpy as np
import segyio


def read_segy(path):
    path = Path(path)
    with segyio.open(str(path), "r", ignore_geometry=True) as src:
        traces = np.stack([np.asarray(trace, dtype=np.float32) for trace in src.trace])
        samples = np.asarray(src.samples)
    return traces, samples


amplitude, time_samples = read_segy("data/Brunswick_valid.sgy")
mask, _ = read_segy("label/Brunswick_valid_mask.sgy")

print(amplitude.shape)
print(mask.shape)

使用 huggingface_hub 下载：

from huggingface_hub import hf_hub_download

repo_id = "GeoBrain/first-break-picking-dataset"

input_path = hf_hub_download(
    repo_id=repo_id,
    filename="data/Brunswick_valid.sgy",
    repo_type="dataset",
)

mask_path = hf_hub_download(
    repo_id=repo_id,
    filename="label/Brunswick_valid_mask.sgy",
    repo_type="dataset",
)

相关模型发布

对应训练模型已单独发布：

GeoBrain/first-break-picking