cangyeone/SeismicX-Cont-mini

SeismicX-Cont Mini Two-Hour Subset

This folder is the compact, Zenodo-archived two-hour mini release for SeismicX-Cont. It is designed for quick download, tutorial use, software smoke tests, and checking that the HDF5, annotation, SQLite, dataloader, picker, and validation workflow all fit together before using the full 14-day data product.

Zenodo record: https://zenodo.org/records/20047415 DOI: https://doi.org/10.5281/zenodo.20047415

Dataset preview

The Hugging Face Dataset Viewer displays a lightweight file index rather than the full HDF5 waveform contents. The waveform data are stored as structured HDF5 files and should be accessed using the provided documentation, notebook, and scripts. The preview table is intended only to summarize available files, time periods, networks, and data types.

Relation to the Full Dataset

This mini release is a two-hour subset of the complete SeismicX-Cont data product. It is intended for quick trials and examples, not for full-release analyses. The complete data product is hosted on both Hugging Face and ModelScope, with a DOI for the Hugging Face record:

https://huggingface.co/datasets/cangyeone/SeismicX-Cont
https://doi.org/10.57967/hf/8993
https://www.modelscope.cn/datasets/cangyeone/SeismicX-Cont

The mini release contains only two one-hour windows:

• Dense Ridgecrest hour: 2019-07-06 04:00-05:00 UTC, 6948 reference picks.
• Quiet-period hour: 2021-11-14 16:00-17:00 UTC, 7 reference picks.

The mini package is about 3.0 GiB. The two waveform files are about 1.7 GiB and 1.2 GiB, and the response JSON is about 119 MiB, so all examples below are intended to run on a laptop. Full-dataset download and rebuild notes are also stored under scripts/README_download_data.md so that the mini archive carries the reproducibility instructions with the executable scripts.

1. Folder Layout

Run commands from this folder unless noted otherwise:

cd publish_mini

Main files:

data/
  hdf5/
    continuous_waveform_usa_20190706_04.h5
    continuous_waveform_usa_20211114_16.h5
  index/
    waveform_index.sqlite
  label/
    annotations_mini_two_hours.json
    stations.csv
  response/
    instrument_responses_mini.json
  picks/
    # generated picker JSONL files are written here
notebooks/
  seismicx_cont_mini_quickstart.ipynb
pickers/
  phasenet.jit
  pnsn.v3.diff.jit
  skynet.jit
scripts/
  README_download_data.md
  README_event_comparison.md
  README_real_association.md
  build_consensus_picks_json.py
  build_waveform_index.sh
  compare_associated_events.py
  download_instrument_responses.py
  evaluate_picks.py
  example_dataloader.py
  example_query_waveform.py
  getwaveform_cha_names.py
  makeh5.py
  makeh5_flex_seg.py
  plot_dataset_overview.py
  plot_essd_qc_figures.py
  readh5.py
  response_json_to_stationxml.py
  run_consensus_picks.sh
  run_picker_to_jsonl.py
  run_pickers.sh
  run_real_association.py
  upload_files_to_hf.py
  upload_files_to_ms.py
  upload_restfiles_hf.py
  upload_restfiles_ms.py
  verify_full_dataset.py
  write_manifest.sh
essd_scripts/
  reproduce_manuscript_outputs.sh
  audit_manuscript_numbers.py
  plot_dataset_overview.py
  plot_monitoring_regime_characterization.py
  plot_essd_qc_figures.py
utils/
  hdf5_waveform_dataset.py
  hdf5_waveform_index.py
  waveform_index_api.py

2. Environment

Use the same Python environment as the complete SeismicX-Cont release. The basic inspection examples need h5py, numpy, and torch; picker and evaluation examples also use the dependencies imported by the scripts.

Quick check:

python - <<'PY'
import h5py, numpy, torch
print("h5py", h5py.__version__)
print("numpy", numpy.__version__)
print("torch", torch.__version__)
PY

If shell scripts are not executable after download, run:

chmod +x scripts/*.sh scripts/*.py

The scripts also respect a PYTHON environment variable. For example:

PYTHON=/Users/anaconda3/bin/python3 ./scripts/verify_full_dataset.py

3. First Five-Minute Check

These commands verify that the package is complete and that the SQLite index can find waveform segments.

python scripts/verify_full_dataset.py --scan-hdf5
python scripts/example_query_waveform.py \
  --db data/index/waveform_index.sqlite \
  --starttime 2019-07-06T04:00:00 \
  --endtime 2019-07-06T04:05:00 \
  --limit 5

Expected high-level counts from verify_full_dataset.py --scan-hdf5:

HDF5 files: 2
waveform_segments: 22908
stations: 971
n_picks: 6955
2019_dense: 6948
2021_quiet: 7

If these counts are close to the above values, the mini release is ready for the examples below.

4. Query Waveforms From SQLite

The SQLite database is a waveform coverage index. It tells you which HDF5 file and dataset path contain a requested station-channel-time interval.

Query and merge one channel into an in-memory waveform array:

python scripts/example_query_waveform.py \
  --db data/index/waveform_index.sqlite \
  --network BK \
  --station BDM \
  --channel BHZ \
  --starttime 2019-07-06T04:00:00 \
  --endtime 2019-07-06T04:01:00 \
  --limit 3

This is the simplest way to confirm that data/index/waveform_index.sqlite and data/hdf5/*.h5 are mutually consistent.

The same SQLite query path can remove native instrument response or simulate a target response, matching the dataloader options:

python scripts/example_query_waveform.py \
  --db data/index/waveform_index.sqlite \
  --network BK \
  --station BDM \
  --channel BHZ \
  --starttime 2019-07-06T04:00:00 \
  --endtime 2019-07-06T04:00:10 \
  --limit 1 \
  --remove_response \
  --response_pre_filt 0.2 0.5 20 45 \
  --response_water_level none

When response processing is enabled, the printed instrument_processing record shows the matched native response epoch and any simulated target-response id.

5. Inspect Annotations

The two-hour annotation subset is:

data/label/annotations_mini_two_hours.json

It preserves the original event, station, phase, status, and provenance fields for the selected one-hour windows. A quick summary:

python - <<'PY'
import json
from pathlib import Path

path = Path("data/label/annotations_mini_two_hours.json")
obj = json.loads(path.read_text())
print(json.dumps(obj["summary"], indent=2, ensure_ascii=False))
PY

Use the 2019 hour to test dense aftershock behavior and the 2021 hour to test low-event-rate monitoring behavior. The 2021 hour is not pure noise; it is the least labeled nonzero hour selected so that the quiet subset still has reference annotations.

6. Use the PyTorch HDF5 Dataloader

Run the minimal dataloader example:

python scripts/example_dataloader.py --n_samples 2

The loader reads the HDF5 files, groups components by station-time window, resamples to the requested sampling rate, and returns tensors in the form:

waveform: torch.Tensor [time, 3]
station_id: network.station.location
channels: selected component channels
starttime: waveform start time
sampling_rate: output sampling rate

To point it at a single HDF5 file:

python scripts/example_dataloader.py \
  --h5_input data/hdf5/continuous_waveform_usa_20190706_04.h5 \
  --n_samples 2

The mini package also includes an instrument-response JSON derived from the complete response inventory:

data/response/instrument_responses_mini.json

Use it when you want the dataloader to remove the native instrument response:

from utils.hdf5_waveform_dataset import HDF5WaveformDataset

dataset = HDF5WaveformDataset(
    h5_file="data/hdf5/continuous_waveform_usa_20190706_04.h5",
    mode="three",
    instrument_response_json="data/response/instrument_responses_mini.json",
    remove_instrument_response=True,
    response_output="VEL",
    response_pre_filt=(0.2, 0.5, 20.0, 45.0),
    response_water_level=None,
)

item = dataset[0]
print(item["waveform"].shape)
print(item["instrument_processing"])

To remove the native response and then simulate a target response, select the target response by response_id or by station-channel metadata:

dataset = HDF5WaveformDataset(
    h5_file="data/hdf5/continuous_waveform_usa_20190706_04.h5",
    mode="single",
    instrument_response_json="data/response/instrument_responses_mini.json",
    remove_instrument_response=True,
    response_output="VEL",
    response_pre_filt=(0.2, 0.5, 20.0, 45.0),
    response_water_level=None,
    simulate_instrument_response=True,
    simulation_response_selector={
        "network": "BK",
        "station": "BDM",
        "location": "00",
        "channel": "BHZ",
        "time": "2019-07-06T04:00:00Z",
    },
    simulation_output="VEL",
)

The response correction and simulation are computed with ObsPy. The mini dataloader and SQLite query API both use the same JSON lookup, station-channel-time matching, and instrument_processing metadata convention.

7. Run One Picker

The quickest picker test uses the bundled PhaseNet TorchScript model:

python scripts/run_picker_to_jsonl.py \
  --h5_input "data/hdf5/continuous_waveform_usa_*.h5" \
  --picker_model pickers/phasenet.jit \
  --output_jsonl data/picks/phasenet.mini.phase.jsonl \
  --canonical_input_length 360000 \
  --max_picks_per_sample 0 \
  --no_auto_restart

Default output:

data/picks/phasenet.mini.phase.jsonl

Run a different bundled model by passing the model path and output path:

python scripts/run_picker_to_jsonl.py \
  --h5_input "data/hdf5/continuous_waveform_usa_*.h5" \
  --picker_model pickers/pnsn.v3.diff.jit \
  --output_jsonl data/picks/pnsn_v3_diff.mini.phase.jsonl \
  --canonical_input_length 360000 \
  --max_picks_per_sample 0 \
  --no_auto_restart

For GPU or Apple Silicon MPS, set DEVICE when using the full three-model script below, or pass --device directly to scripts/run_picker_to_jsonl.py.

8. Run the Three Bundled Pickers

This runs PhaseNet, PNSN v3 diff, and SkyNet on the two HDF5 files:

DEVICE=cpu ./scripts/run_pickers.sh

Use DEVICE=mps on Apple Silicon or DEVICE=cuda on CUDA machines:

DEVICE=mps ./scripts/run_pickers.sh

Generated JSONL files:

data/picks/phasenet.mini.phase.jsonl
data/picks/pnsn_v3_diff.mini.phase.jsonl
data/picks/skynet.mini.phase.jsonl

Each JSONL line is one station-time inference record with picker outputs. These files are generated products and are intentionally not included as required input files.

9. Evaluate Picker Output

After generating a picker JSONL file, evaluate it against the mini annotation subset:

python scripts/evaluate_picks.py \
  --auto-jsonl data/picks/pnsn_v3_diff.mini.phase.jsonl \
  --label-json data/label/annotations_mini_two_hours.json \
  --index-db "${LOCAL_PICK_DB:-/tmp/seismicx_cont_picks.sqlite}" \
  --outdir eval_picks/example \
  --waveform-db data/index/waveform_index.sqlite \
  --tp-tol 1.5 \
  --err-window 5.0 \
  --build-index \
  --drop-existing \
  --plot

Outputs are written under:

eval_picks/example/

The evaluation script uses:

• data/label/annotations_mini_two_hours.json as the reference annotation set.
• data/index/waveform_index.sqlite as the waveform coverage index.
• A 1.5 s true-positive tolerance.
• A 5.0 s residual fitting window.

The mini evaluation is only a smoke test. It should not be interpreted as a relative model-assessment result because the subset has just two one-hour windows.

10. Build Optional Consensus Candidates

If the three picker JSONL files exist, build consensus-supported candidate arrivals:

./scripts/run_consensus_picks.sh

Default output:

data/label/consensus_picks.json

The default rule requires at least three models to agree within 1.5 s on the same station and phase. This layer is optional. It is useful for auditing candidate detections and testing workflow extensions, but it is not the primary reference label set.

11. Run REAL Association

The mini package also includes a bridge from SeismicX picker JSONL output to the REAL association algorithm. It compiles scripts/real/real.c, converts phase-pick JSONL records to REAL input files, runs association-based event detection, and writes a workflow JSONL that contains both time-window phase candidates and associated earthquake events. Downstream earthquake location can be attached by users through the associated-arrival list in each event record.

Run from the publish_mini/ folder:

python scripts/run_real_association.py \
  --picks-jsonl data/picks/pnsn_v3_diff.mini.phase.jsonl \
  --output-jsonl data/associated/real_events.jsonl \
  --starttime 2019-07-06T04:00:00Z \
  --endtime 2019-07-06T05:00:00Z

Default output:

data/associated/real_events.jsonl
data/associated/real_events.summary.json

The JSONL contains one real_association_run metadata record, one real_input_pick record for each phase pick that entered REAL after filtering and NMS, and one real_event record for each associated event. Each event record contains an origin-time estimate, optional preliminary latitude/longitude/depth when provided by the associator, association counts, the parameters used, and the associated arrivals. Each arrival keeps a source_pick.real_input_pick_id that points back to the corresponding time-window phase record, as well as the original picker JSONL line, station id, phase name, probability, channel family, and source HDF5 file.

The default settings are conservative smoke-test settings: --min-prob 0.85, 0.5 s station-phase NMS, and a small REAL trigger requirement. Lowering --min-prob can recover more candidate arrivals but will increase runtime quickly during dense aftershock periods. Use --keep-work-dir if you want to inspect the generated REAL station.dat, pick files, catalog_sel.txt, and phase_sel.txt. Use --no-include-input-picks for an event-only JSONL in a large production run.

For details, see:

scripts/README_real_association.md

Compare the associated events with the mini catalog:

python scripts/compare_associated_events.py \
  --pred-jsonl data/associated/real_events.jsonl \
  --catalog data/label/annotations_mini_two_hours.json \
  --outdir eval_events/real_vs_catalog_mini

An event is counted as a true positive when the epicentral distance error is less than 20 km and the origin-time error is less than 3 s. The generic event JSONL format accepted by the comparison script is documented in:

scripts/README_event_comparison.md

12. Notebook Walkthrough

For an interactive version of the workflow, open:

notebooks/seismicx_cont_mini_quickstart.ipynb

The notebook covers:

• checking package files and sizes;
• reading the annotation summary;
• querying the SQLite waveform index;
• plotting one HDF5 waveform segment;
• instantiating the PyTorch dataloader;
• preparing picker, evaluation, and consensus commands.

13. Rebuild the Mini Package

Most users do not need to rebuild the mini package. Use this section only if you have the complete SeismicX-Cont annotation file and the source MiniSEED archive.

Rebuild the two hourly HDF5 files from the source waveform archive with the common flexible HDF5 builder:

SEISMICX_RAW_ROOT=/path/to/continous_usa/data

python scripts/makeh5_flex_seg.py \
  --input_dir "$SEISMICX_RAW_ROOT/07/06" \
  --loc_file data/label/stations.csv \
  --output data/hdf5/continuous_waveform_usa.h5 \
  --split_interval hour \
  --include_split_file_id 20190706_04 \
  --num_workers "${NUM_WORKERS:-8}" \
  --max_pending "${MAX_PENDING:-16}" \
  --compression gzip \
  --compression_opts 4

python scripts/makeh5_flex_seg.py \
  --input_dir "$SEISMICX_RAW_ROOT/11/14" \
  --loc_file data/label/stations.csv \
  --output data/hdf5/continuous_waveform_usa.h5 \
  --split_interval hour \
  --include_split_file_id 20211114_16 \
  --num_workers "${NUM_WORKERS:-8}" \
  --max_pending "${MAX_PENDING:-16}" \
  --compression gzip \
  --compression_opts 4

The script expects source waveform folders such as:

$SEISMICX_RAW_ROOT/07/06
$SEISMICX_RAW_ROOT/11/14

Rebuild the SQLite waveform index:

./scripts/build_waveform_index.sh

This script writes the SQLite database to local temporary storage first and then copies it into data/index/, because some external drives are unreliable for SQLite database creation. You can choose the temporary path explicitly:

LOCAL_DB=/tmp/seismicx_cont_waveform_index.sqlite \
  ./scripts/build_waveform_index.sh

Update the SHA256 and file-size manifests:

./scripts/write_manifest.sh

The script writes manifest_sha256.txt and manifest_filesizes.tsv, which list the released paths, SHA256 checksums, and byte counts for package inspection.

14. Troubleshooting

ModuleNotFoundError: No module named 'utils' : Run commands from the publish_mini folder, or use the provided scripts rather than moving files into another directory.

Permission denied for shell scripts : Run chmod +x scripts/*.sh scripts/*.py.

SQLite errors on an external drive : Build the database on local storage with LOCAL_DB=/tmp/... and let build_waveform_index.sh copy the finished database back into data/index/.

Picker is slow on CPU : Start with one model using run_picker_to_jsonl.py. Use DEVICE=mps or DEVICE=cuda with run_pickers.sh when hardware is available.

Missing source waveform archive during rebuild : The downloaded mini package already includes HDF5 files. Rebuilding HDF5 is only possible if you also have the original MiniSEED source archive.

REAL association is slow : Start with the default one-hour command and --min-prob 0.85. Lower probability thresholds can greatly increase the number of input picks and the REAL search time.

15. What to Cite

When using this mini subset in examples or tutorials, cite the complete SeismicX-Cont release and the original waveform/data sources described in the paper. The complete dataset is available at:

https://www.modelscope.cn/datasets/cangyeone/SeismicX-Cont

The mini subset is a convenience package derived from the complete SeismicX-Cont data-product release.