TremorCNN

Low-Frequency Tremor Detection with Deep Learning

A convolutional neural network that detects tectonic tremor in real seismic waveform data fetched live from the EarthScope/IRIS archive. Built as a proof-of-concept for automated tremor detection on the Cascadia subduction zone.

---

What is Tectonic Tremor?

Tectonic tremor is a weak, sustained seismic signal (1-8 Hz) generated by slow slip along fault boundaries. In Cascadia, tremor episodes occur every ~12-14 months and are closely linked to slow slip events that load stress onto the locked fault zone capable of producing a magnitude 9+ megathrust earthquake. Detecting tremor automatically across years of continuous data is a key open problem in seismology.

---

Approach

Real waveform (IRIS/EarthScope)
        ↓
Bandpass filter (1-8 Hz)
        ↓
Short-Time Fourier Transform → spectrogram patch (45 x 10)
        ↓
2D CNN binary classifier (tremor / no tremor)
        ↓
Predicted label + confidence

Labels are derived from the PNSN/Wech tremor catalog for confirmed Cascadia ETS episodes (Aug 2010, Sep 2011, Oct 2013). Quiet windows are drawn from periods with no known ETS activity across multiple times of day and seasons.

---

Results

Metric	Value
Test Accuracy	87.0%
Precision	0.186
Recall	0.741
F1	0.297
ROC-AUC	0.869
True Positives (tremor correctly detected)	109
False Positives (quiet incorrectly flagged)	477
True Negatives (quiet correctly identified)	3326
False Negatives (tremors missed)	38

Trained on 15,796 spectrogram windows, tested on 3,950 held-out windows using chronological split: train/test boundary at 2016-01-05. Decision threshold: 0.80.

Training loss curve:

Inference on unseen station UW.MRBL (Marblemount, WA) - Sep 15 2011 ETS(Episodic Tremor and Slip):

---

Data

All waveform data fetched live via ObsPy from the EarthScope/IRIS FDSN (Federation of Digital Seismograph Networks) web service. No local waveform files required are data streams directly from the archive at runtime.

• Network: UW (Pacific Northwest Regional Seismic Network)

• Stations: GNW (Green Mountain WA), FORK (Forks WA), LEBA (Lebam WA)

• Channel: BHZ/HHZ/EHZ (broadband vertical), resampled to 40 Hz

• Tremor labels: PNSN catalog CSVs (tremor_events-YYYY.csv), 80,479 individual tremor event timestamps

• ETS episodes: 2010, 2011, 2012, 2013, 2014, 2015-16, 2017, 2018

• Quiet periods: Mar 2010, Apr 2011, Jun 2012, Jun 2014, Jun 2015, Jun 2016, Jun 2019

• Model weights: MorphKnight1440/TremorCNN

• Tremor catalog dataset: MorphKnight1440/CascadiaTremorCatalog

---

Model Architecture

Input: (1, 45, 10) spectrogram patch
→ Conv2d(1→16, 3x3) + ReLU + MaxPool2d(2)   # → (16, 22, 5)
→ Conv2d(16→32, 3x3) + ReLU + MaxPool2d(2)  # → (32, 11, 2)
→ Flatten → Linear(704→64) + ReLU
→ Linear(64→1)

Binary Cross-Entropy loss, Adam optimizer (lr=1e-3), 20 epochs.

---

Reproducing

git clone https://github.com/Eshwar1440/TremorCNN
cd TremorCNN
pip install obspy torch numpy scipy matplotlib

Train from scratch

python train.py

Downloads are needed to train, Please look at Data section for Hugging Face Dataset I used.

Run inference on unseen data (pretrained weights included)

python predict.py

---

Limitations

• Cascadia subduction zone only, not tested on other tectonic settings
• 18,007 training windows vs ~95,000 in the reference paper
• Catalog labels have ~2.5-minute resolution, some window boundaries are approximate
• Low precision (0.134) at default 0.7 threshold, threshold tuning needed for deployment
• 2020/2021 ETS episodes excluded due to waveform archival gaps on GNW

---

Future Work

• Multi-station generalization (test on UW.LRIV, CN network stations)
• Multi-station voting (classify it as tremor if more than 2 stations agree.)
• Replace CNN with transformer architecture for long-range temporal patterns
• Retrain on noise windows the model mistakes for tremor to cut false alarms
• ROC curve analysis and threshold tuning for operational deployment
• Integration with live EARTHSCOPE stream for real-time detection

---

Related Work

This project independently reproduces the core pipeline of:

Rouet-Leduc, B., Hulbert, C., McBrearty, I. W., & Johnson, P. A. (2020). Probing Slow Earthquakes With Deep Learning. Geophysical Research Letters, 47, e2019GL085870. DOI: 10.1029/2019GL085870

That paper trained a CNN on ~165,000 Cascadia tremor examples from the PNSN catalog and showed the model generalizes across subduction zones and to the San Andreas Fault at Parkfield.

---

Author

Eshwar Singh Rajaputana.