[Paper Review] Combining Deep Learning with Physics Based Features in Explosion-Earthquake Discrimination

Qingkai Kong, Ruijia Wang+4 more|arXiv (Cornell University)|Mar 11, 2022

Seismology and Earthquake Studies35 references72 citations

TL;DR

This paper proposes a hybrid deep learning model that combines physics-based seismic features—such as high-frequency P/S amplitude ratios and ML–MC magnitude differences—with waveform-derived features from convolutional neural networks for improved discrimination between earthquakes and explosions. The fusion enhances generalization across new regions, with Grad-CAM visualization revealing that the model focuses on distinct frequency bands in P and S waves to make decisions.

ABSTRACT

This paper combines the power of deep-learning with the generalizability of physics-based features, to present an advanced method for seismic discrimination between earthquakes and explosions. The proposed method contains two branches: a deep learning branch operating directly on seismic waveforms or spectrograms, and a second branch operating on physics-based parametric features. These features are high-frequency P/S amplitude ratios and the difference between local magnitude (ML) and coda duration magnitude (MC). The combination achieves better generalization performance when applied to new regions than models that are developed solely with deep learning. We also examined which parts of the waveform data dominate deep learning decisions (i.e., via Grad-CAM). Such visualization provides a window into the black-box nature of the machine-learning models and offers new insight into how the deep learning derived models use data to make the decisions.

Motivation & Objective

• To improve the generalization of deep learning models for explosion-earthquake discrimination across new geographic regions.
• To integrate physically interpretable seismic features (e.g., P/S amplitude ratios, ML–MC differences) into a deep learning framework to enhance model robustness.
• To investigate how deep learning models rely on specific waveform components using interpretability techniques like Grad-CAM.
• To demonstrate that combining data-driven and physics-informed features leads to more reliable and transportable seismic source classification.

Proposed method

• A dual-branch neural network architecture is used: one branch processes raw seismic waveforms or spectrograms via convolutional and recurrent layers, while the other processes engineered physics-based features.
• Physics-based features include high-frequency P/S amplitude ratios and the difference between local magnitude (ML) and coda duration magnitude (MC), known to distinguish explosion and tectonic sources.
• Features from both branches are concatenated and passed through fully connected layers for final classification.
• Model interpretability is achieved using Grad-CAM to visualize attention maps on time-series and spectrogram inputs, identifying which waveform regions influence predictions.
• Data augmentation is applied by randomly sampling start times for each event, increasing training data diversity and balancing class distributions.
• The model is trained and evaluated on four regional datasets (SPE, BASE, MSH, SSIP) with consistent preprocessing: detrending, Hanning tapering, 1–20 Hz bandpass filtering, and resampling to 40 Hz.

Experimental results

Research questions

• RQ1 Can combining physics-based features with deep learning improve model generalization in explosion-earthquake discrimination across new regions?
• RQ2 Which parts of the seismic waveform does the deep learning model rely on most for classification decisions?
• RQ3 How do physics-based features contribute to performance when the model is applied to data outside the training region?
• RQ4 To what extent do the deep learning and physics-based branches complement each other in the decision-making process?
• RQ5 Can visualization techniques like Grad-CAM reveal meaningful patterns consistent with seismological understanding?

Key findings

• The hybrid model achieves better generalization performance on unseen regional data compared to a deep learning model using waveforms alone.
• The addition of physics-based features improves model performance on new regions, indicating that these features provide information not fully captured by the deep learning branch.
• Grad-CAM visualization shows that the deep learning model focuses on different frequency bands in P and S waves, with peak attention on high-frequency components.
• The model’s attention on P and S wave phases aligns with seismological knowledge, supporting the physical interpretability of the learned features.
• Coda waves, particularly at longer times, also contribute to discrimination, especially at farther distances.
• The model’s reliance on specific frequency bands suggests potential for dynamic or adaptive P/S ratio calculations in future work.