pnw_seismic_event_detection

QuakeXNet · Mt. Rainier Surface Event Catalog (2010–2025)

End-to-end pipeline for building a 15-year catalog of surface seismic events near Mt. Rainier. The workflow runs the QuakeXNet deep-learning detector on continuous waveforms from stations within 50 km of the volcano, aggregates detections across the network, locates them with ENVELOC, and validates the results against the PNSN and ESEC catalogs.

   
Surface events (SU) 114,775
Explosions (PX) 13,766
Total located 128,541
Period Jan 2010 – Dec 2025

Catalog Overview

Catalog overview – event locations, monthly counts, and distance distribution

Interactive dashboard – explore the full catalog at akashkharita.github.io/pnw_seismic_event_detection/data/enveloc_dashboard.html: heatmap / point-cloud toggle, time-series, distance histogram, and per-class / per-station-count filters.

Repository Structure

src/          Detection and location scripts
notebooks/    Analysis, validation, and exploration notebooks
data/         Catalog outputs and dashboard
utils/        Station metadata

Notebooks

Detection & Pipeline

Notebook What it shows
plotting_commonly_detected_events.ipynb Gathers per-station daily detections, finds events seen at ≥ 4 stations, and plots waveforms + probability curves for network-level events
walking_through_generate_common_events.ipynb Step-by-step walkthrough of the common-event aggregation logic (time-alignment, grouping, station-count filtering)
creating_station_json_file_for_the_detector.ipynb Builds the stations.json configuration used by the detector

Location

Notebook What it shows
enveloc_location_for_detected_events.ipynb Runs ENVELOC on network-level detections; benchmarks sequential vs. parallel waveform downloading
combining_all_the_locations_into_single_file.ipynb Merges per-day ENVELOC outputs into the master catalog CSV
testing_location_using_enveloc.ipynb Tests and tunes ENVELOC hyperparameters on a subset of events

Catalog Validation & Diagnostics

Notebook What it shows
quakexnet_diagnostic.ipynb Comprehensive diagnostic: PNSN recall analysis, missed-event classification (confused / ambiguous / missed), single-event trace-through, co-association check, and V1 vs V3 catalog comparison
comparing_with_pnsn_catalog.ipynb Matches QuakeXNet detections to the PNSN catalog by distance and time tolerance; computes precision/recall
validating_enveloc_locations_for_esec_events.ipynb Cross-checks located events against the ESEC documented surface-event catalog (quality score 1–5)
analyzing_quakexnet_15_years_detection_results.ipynb Bulk analysis of the full 15-year detection run: class distributions, temporal trends, and PNSN comparison

Surface Event Characterization

Notebook What it shows
su_clustering_new_new.ipynb Clusters surface events by source type using 128-dim QuakeXNet embeddings (best-window extraction → UMAP → HDBSCAN), with spectrogram visualization per cluster
single_event_analysis.ipynb Deep-dive on a single icefall event: compares STA/LTA, deep-learning, and kurtosis pickers, motivating the need for ENVELOC
single_event_infrasound_analysis.ipynb Infrasound analysis for individual surface events
visualizing_infrasound_data_for_mt_rainier_surface_events.ipynb Surveys infrasound signatures across the catalog to identify event types with strong infrasound expression
plotting_the_infrasound_location_with_original_one.ipynb Compares ENVELOC seismic locations against infrasound-derived locations

Model Diagnostics

Notebook What it shows
quakexnet_diagnostic.ipynb (see Validation above)
testing_surface_event_pickers.ipynb Benchmarks different onset-picking strategies on surface event waveforms

Using the trained QuakeXNet model with SeisBench

1) Install SeisBench

pip install seisbench

2) Copy the model definition into SeisBench

SeisBench discovers models from its internal seisbench/models/ package:

cp src/quakexnet.py <PATH_TO_SITE_PACKAGES>/seisbench/models/quakexnet.py

To find the correct models/ directory:

python -c "import seisbench, os; print(os.path.join(os.path.dirname(seisbench.__file__), 'models'))"

3) Register QuakeXNet inside seisbench.models

Open <PATH_TO_SITE_PACKAGES>/seisbench/models/__init__.py and add:

from .quakexnet import QuakeXNet

4) Add trained weights to the SeisBench cache

mkdir -p ~/.seisbench/models/v3/quakexnet
cp src/base.pt.v3 ~/.seisbench/models/v3/quakexnet/base.pt.v3
echo '{}' > ~/.seisbench/models/v3/quakexnet/base.json.v3

Final layout:

~/.seisbench/models/v3/quakexnet/
├── base.pt.v3
└── base.json.v3

5) Import and use the model

import seisbench.models as sbm

model = sbm.QuakeXNet()

Detection Pipeline

src/custom_daily_detection.py

Runs QuakeXNet on continuous waveform data for each station and logs per-station event detections.

Workflow:

  1. Loads the pre-trained QuakeXNet model and station list from stations.json
  2. Downloads waveform data from IRIS via ObsPy for each station
  3. Runs model inference with a 100 s window and 10 s stride, producing per-window probabilities for eq, px, and su
  4. Smooths probability curves with a 5-sample moving average (~50 s)
  5. Detects events: start ≥ 0.15, end < 0.15, kept only if max probability ≥ 0.5

Output — one CSV per station:

station network class auc mean_prob max_prob start_time end_time
PARA CC eq 3.37 0.35 0.54 2025-12-13T14:44:22Z 2025-12-13T14:45:52Z

src/custom_generate_common_events.py

Aggregates per-station detections into network-level common events.

Workflow:

  1. Merges all per-station CSVs for the chosen date range
  2. Rounds start times to the nearest 10 s to align slightly offset detections
  3. Groups by rounded start time; computes num_stations, most_common_class, mean_auc/max/prob
  4. Keeps only events detected at ≥ 4 stations (default)

Output — one CSV per day:

rounded_start num_stations stations most_common_class mean_auc
2025-08-03 20:03:30+00:00 4 [‘RCM’, ‘RER’, ‘STAR’, ‘OBSR’] su 4.72