An Interpretable Physics Informed Multi-Stream Deep Learning Architecture for the Discrimination between Earthquake, Quarry Blast and Noise

TL;DR

The paper tackles the problem of distinguishing earthquakes, quarry blasts, and noise from single-station seismic data. It introduces PI-CRNN, a three-branch architecture with Time-Domain SincNet, Multi-Resolution Spectrogram, and Physics Block streams, fused and followed by a Bi-LSTM with attention; physics is incorporated via $E_{tot}(t) = Z(t)^2 + N(t)^2 + E(t)^2$ and $R_{ZH}(t) = Z(t)^2 /(N(t)^2 + E(t)^2 + epsilon)$. On the Curated Pacific Northwest dataset, the method achieves 97.56% test accuracy with perfect noise recall and interpretable saliency maps that reveal distinct seismic signatures for earthquakes and blasts. The results show a clear advantage over a non-physics CRNN, a classical STA/LTA–P/S approach, and a PINN-based loss approach, highlighting the value of architectural inductive bias for reliable, interpretable seismology. The work provides a scalable framework for AI-driven seismic monitoring with improved generalization and transparency, and outlines future directions toward multi-station data, transfer learning, and self-supervised learning.

Abstract

The reliable discrimination of tectonic earthquakes from anthropogenic quarry blasts and transient noise remains a critical challenge in single station seismic monitoring. In this study, we introduce a novel Physics Informed Convolutional Recurrent Neural Network (PI CRNN) that embeds seismological domain knowledge directly into the feature extraction and learning process. We adapt a multistream architecture with three parallel encoders: (i) Time Domain: SincNet Encoder, (ii) MultiResolution Spectrogram branch, and, (iii) Physics Branch, followed by a fusion and a bidirectionalLSTM module. Evaluated on the Curated Pacific Northwest AI ready Seismic Dataset, the PI CRNN achieves an overall classification accuracy of 97.56 percent on an independent test set. It outperforms a standard CRNN baseline, a classical P to S amplitude ratio method, and a Physics Informed Neural Network (PINN) that enforces physical constraints via the loss function. Furthermore, the model demonstrates perfect precision in noise rejection (100 percent Recall). Interpretability analysis using saliency maps confirms that the architecture successfully learns distinct physical signatures, identifying bimodal P- and S-wave arrivals for earthquakes versus singular impulsive onsets for blasts. This work establishes a scalable, transparent framework for AI-driven seismology, proving that architectural inductive bias provides an alternative reliable approach compared to purely data-driven approaches.

Figures (9)

• Figure 1: Statistical distribution of attributes for the earthquake class within the dataset. (Top Left) The magnitude distribution shows majority of events classified as micro-seismicity (Magnitude 1–2). (Top Right) Source depth distribution indicates that most events are shallow crustal earthquakes occurring between 0 and 20 km depth. (Bottom) The distribution of the average Signal-to-Noise Ratio (SNR) shows that the dataset is dominated by lower SNR events (10–30 dB), reflecting the challenging conditions typical of local monitoring environments
• Figure 2: Model Architecture of PI-CRNN
• Figure 3: Confusion Matrix for the proposed PI-CRNN model
• Figure 4: Tectonic Earthquake: The saliency map exhibits a clear bimodal structure. The first peak aligns with the P-wave onset, followed by a dominant peak corresponding to the S-wave arrival. Consistently, the feature map from the time-domain branch shows a distinct double-activation pattern, indicating that the network has successfully learned to temporally separate the P and S phases.
• Figure 5: Blast: The saliency map is characterized by a single dominant peak. The corresponding frequency-domain features display vertically broad bands rather than the two distinct bands observed in the earthquake case, suggesting that the model captures an instantaneous, high-frequency energy release. Likewise, the time-domain feature map exhibits a single, sharp vertical activation.