Ensemble Deep Learning to Mine Cascadia Offshore Seismicity

This project builds a high-resolution earthquake catalog for coastal and offshore Cascadia (2010–2015) from the Cascadia Initiative ocean-bottom seismometers (OBS) and regional land networks, using an ensemble deep-learning pipeline:

1. Detection & phase picking with an ensemble picker (ELEP; Yuan et al., 2023),
2. Phase association with the GENIE graph neural network (PyOcto was also tested),
3. Relocation with GraphDD double-difference relocation, refined by waveform cross-correlation differential times (HypoDD was also tested),
4. Quality control, magnitude estimation, and comparison against established catalogs — USGS ComCat and Morton et al., 2023.

The manuscript describing this work is authored in paper/ (Quarto → LaTeX, synced to Overleaf; see paper/README.md).

Authors

Repository / software — code and analysis in this repository:

Author	Affiliation	Contribution
Marine Denolle (mdenolle@uw.edu)	UW Earth & Space Sciences	project lead, magnitudes, QC/post-processing
Hiroto Bito (hbito@uw.edu)	UW Earth & Space Sciences	picking, amplitude, QC & post-processing pipeline
Qibin Shi (qibins@uw.edu)	UW Earth & Space Sciences	code
Yiyu Ni (niyiyu@uw.edu)	UW Earth & Space Sciences	code
Nathan T. Stevens (ntstevens@uw.edu)	Pacific Northwest Seismic Network	code
Ian W. McBrearty	Stanford Geophysics	graph networks — GENIE association & GraphDD relocation
Yifan Yu	Stanford Geophysics	graph networks — GENIE association & GraphDD relocation
Zoe Krauss (zkrauss@uw.edu)	UW Oceanography	results cross-checking

Manuscript — the working paper draft in paper/ has a longer author list: Marine Denolle, Hiroto Bito, Qibin Shi, Yiyu Ni, Ian W. McBrearty, Zoe Krauss, Nathan T. Stevens, Yifan Yu, and Gregory C. Beroza (Stanford Geophysics).

Repository structure

📜 README.md · INSTALL.md · LICENSE · CITATIONS.cff
📜 pixi.toml          # primary environment (pixi); environment.yml / requirements.txt are fallbacks
📜 Makefile           # `make paper` (manuscript), `make figs` (paper figures)
📜 verify_environment.py   # `pixi run verify` — dependency smoke test
📜 download_data.sh   # rsync figure/pipeline input catalogs from the lab server
📦 0_data_availability
📦 1_picking          # ensemble ELEP picking: parallel_pick_20{10-15}*.py, picking_utils*.py
📦 2_association      # (association is run with GENIE; see the graph-network tooling)
📦 3_post_processing  # amplitudes, catalog merge, ANSS concat, QC metrics, cross_correlation/, quality_control/
📦 4_relocation       # relocation post-processing + QC + magnitudes
 ┣ 📜 calculate_amplitudes.py            # per-pick peak amplitudes
 ┣ 📜 qc_metrics_all_regions_*.ipynb     # QC thresholds (P>=4, S>=4, RMS<2.5)
 ┣ 📦 cross_correlation                  # CC differential-time dataset builders
 ┣ 📦 quality_control                    # QC/examine notebooks
 ┗ 📦 magnitude                          # ML magnitude pipeline (phase1..6) + methods docs
📦 data               # catalogs + config (see "Data"); large CSVs live in data/split_files/
📦 utils              # data_client.py, plot_utils.py, qc_utils.py, split_large_csvs.py, reconstruct_split_csvs.py
📦 figures            # fig1..fig6 notebooks for the manuscript
📦 paper              # Quarto → Seismica manuscript, auto-synced to Overleaf
📦 .claude/skills     # pre-submission-reviewer + plain-voice (paper-iteration agents)

Installation

Recommended: pixi — one command, reproducible.

# 1. Install pixi (once)
curl -fsSL https://pixi.sh/install.sh | bash

# 2. Clone
git clone https://github.com/Denolle-Lab/cascadia_obs_ensemble.git
cd cascadia_obs_ensemble

# 3a. Public / EarthScope FDSN (default — works anywhere)
pixi install
# 3b. UW internal — also installs pnwstore waveform-archive access
pixi install --environment internal
# 3c. Manuscript toolchain (Quarto + LaTeX) — only if building the paper
pixi install --environment paper

# 4. Verify the install (imports the core deps + notebook stack)
pixi run verify

# 5. Use it
pixi run notebook        # Jupyter (notebook workflow)
pixi run pick            # example CLI entry point (1_picking)

Conda fallback: conda env create -f environment.yml && conda activate seismo_cobs pip fallback: pip install -r requirements.txt (note: basemap and pygmt need conda-forge / a system GMT; obspy is easiest via conda-forge).

Data

The pipeline produces three nested datasets (each a smaller, higher-quality subset: ~39.6 M ELEP picks → 1.09 M associated picks / 116,591 events → 1.00 M relocated picks / 63,887 events → 31,020 events after QC). See data/LINEAGE.md for the full creation-time lineage, row counts, and which files are final vs alternative.

Three distinct data products span the pipeline; they are not the same artifact:

1. Raw ELEP picks (per-station P/S picks, 2010–2015) — the picker output, upstream of association. Currently on Google Drive (private until archived on Zenodo): Picks folder.
2. Relocated catalog + picks (shipped in this repo, data/) — the GraphDD-relocated event catalog (Cascadia_relocated_catalog_ver_3.csv) and its phase picks (Cascadia_relocated_catalog_picks_ver_3*, with optional Wood-Anderson amplitudes). These exceed GitHub's file-size limit, so they are committed as ≤50 MB chunks in data/split_files/ (via utils/split_large_csvs.py). Rebuild the monoliths with:

   pixi run python utils/reconstruct_split_csvs.py        # --list to preview

3. Final cross-correlated + QC-filtered catalog used in the paper figures (origin_2010_2015_reloc_cog_ver3_cc_p_4_s_4_rms_2_5.csv, plus arrival_*, assoc_*, all_stations_*, comparison catalogs) — these live on the lab server, not in the repo. Fetch them with:

   ./download_data.sh mdenolle@<host>            # small catalogs (fig1/4/5/6)
   ./download_data.sh mdenolle@<host> --with-picks   # + arrival/assoc (fig3, large)

Filename conventions: reloc = GraphDD-relocated; cog = center-of-gravity cluster step; cc = cross-correlation-refined; p_4_s_4_rms_2_5 = QC filter (≥4 P picks, ≥4 S picks, RMS < 2.5 s). data/ds01.csv is Morton et al. (2023); nodes_*/vel_*.csv are GraphDD velocity/region config; jgrb52524-* are external published supplements.

Zenodo archive (planned)

For publication, archive the analysis-ready monolithic CSVs (not the GitHub chunks) as a single Zenodo record — see data/ZENODO.md for the proposed package layout and the download recipe. Superseded *_old amplitude files are kept for provenance but excluded from the archive.

Building the manuscript

make paper       # paper/main.qmd -> paper/main.tex + main.pdf (Quarto -> Seismica -> tectonic)
make figs        # collect figure PNGs into paper/figures/ (needs the data above)

See paper/README.md for the authoring + Overleaf-sync workflow.