REAVER: Real-time Earthquake Prediction with Attention-based Sliding-Window Spectrograms

Lotfy H. Abdel Khaliq, Sabine Janzen, Wolfgang MaassIJCAI 2024

Abstract

Earthquake early warning systems (EEWS) are critical for mitigating seismic disasters, yet they often face a trade-off between detection speed and accuracy. This paper presents REAVER, a novel EEWS that utilizes attention-based sliding-window spectrograms to detect seismic P- and S-waves. REAVER achieves a detection latency of just 0.08 seconds with 98.8% accuracy. By processing seismic data as Mel-Spectrograms, the system identifies subtle frequency shifts at wave onset that traditional models overlook.

Introduction

In the context of earthquake early warning, every millisecond of lead time can save lives. Traditional methods like STA/LTA (Short-Term Average/Long-Term Average) are computationally efficient but prone to false alarms and lack regional generalizability. Conversely, modern deep learning models such as PhaseNet or EQTransformer offer high accuracy but typically require long input windows (30–60 seconds), introducing significant lag. REAVER addresses this gap by providing an ultra-fast, highly precise system capable of triggering warnings the moment a seismic wave reaches a sensor.

Methodology

REAVER employs a four-layer architecture designed for real-time performance:

  1. Data Acquisition Layer: Real-time 3-channel seismic data (Vertical, North-South, East-West) is streamed via SeedLink servers.
  2. Pre-processing Layer: To capture transient frequency changes, REAVER uses a 4-second sliding window with a 99% overlap. These windows are transformed into Mel-Spectrograms using Short-Time Fourier Transforms (STFT), treating seismic signals similarly to audio data.
  3. Model Layer: The core is a multi-stage encoder-decoder. The encoder uses four stages of convolutional layers to extract features. Two specialized decoders for P-waves and S-waves utilize spatial self-attention mechanisms to focus on wave-specific signatures.
  4. Warning Layer: Probabilities ($P_p$ and $P_s$) are continuously evaluated; an alert is triggered if values exceed a 0.7 threshold.
Overview of the REAVER four-layer architecture, from data acquisition to warning generation.

The model was trained on the Stanford Earthquake Dataset (STEAD), utilizing over 51,000 earthquake waveforms and 11,000 noise samples to ensure robustness against non-seismic signals like traffic.

Results and Key Insights

REAVER demonstrates superior performance across key metrics compared to both classical and deep learning benchmarks.

  • Detection Speed: REAVER's mean detection time difference ($\Delta t_{mean}$) is 0.08 seconds, representing a 70% improvement over STA/LTA and a 95% improvement over PhaseNet (1.67s).
  • Accuracy and Robustness: The system achieved a classification accuracy of 98.81%, with high precision (0.99) and recall (0.98).

The success of REAVER stems from treating seismic waves as audio. Mel-Spectrograms effectively highlight the "instant tiniest changes" in frequency that raw waveforms obscure. Furthermore, the 99% window overlap ensures near-continuous monitoring, catching wave onsets almost immediately.

Discussion and Future Work

REAVER represents a significant advancement in phase-picking speed. However, several avenues for future research remain. While the system excels at detection, integrating magnitude and location estimation would enhance its utility. Additionally, while trained on global data, real-world performance is subject to local geological variations.

Real-time monitoring dashboard showing seismic waveforms and station location.

Finally, while algorithmic latency is minimized, total warning time in deployment will still be influenced by network hardware and transmission delays. Future iterations will focus on edge deployment to further reduce end-to-end latency.

Summary updated

BibTeX
@inproceedings{khaliq2024reaver,
  author = {Lotfy H. Abdel Khaliq and Sabine Janzen and Wolfgang Maass},
  title = {{REAVER: Real-time Earthquake Prediction with Attention-based Sliding-Window Spectrograms}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure