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.
Accurate discrimination between earthquakes, anthropogenic explosions, and transient background noise remains a fundamental challenge in observational seismology. This capability is critical for seismological monitoring agencies to ensure the integrity of seismic catalogs (used for hazard assessment) and for ensuring compliance with international treaties such as the Comprehensive Nuclear-Test-Ban Treaty (CTBT) [1]. As seismic networks are expanding, with time the volume of data generated has far exceeded the capacity for manual analysis. This necessitates the development of reliable, robust, automated classification systems.
Classically, the discrimination of earthquakes and blasts is performed using physicsbased methods, such as the amplitude ratio [2] and spectral energy distribution [3]. These methods are rooted in the fundamental source physics: tectonic earthquakes are shear-source mechanisms that generate strong S-waves, whereas explosions are isotropic, compressional sources that produce dominant P-waves with relatively weak shear energy. While these methods are transparent and provide physical interpretability, they often struggle with low-magnitude events, low signal-to-noise ratio (SNR) and/or events recorded at local and regional distances.
As machine learning advanced, it provided an alternative approach to addressing this problem. For instance, [4] employed a Support Vector Machine (SVM) to discriminate between earthquakes and explosions, using the amplitudes of the P-wave and S-wave components of seismic signals to construct feature vectors. [5] investigated the applicability of Random Forest algorithms to classify events recorded during operational seismic monitoring in the Upper Rhine Graben area. [6] evaluated multiple machine learning techniques including Decision Trees, K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), and Logistic Regression to distinguish between natural earthquakes and nuclear explosions.
While classical machine learning significantly advanced automation, deep learning (DL) fundamentally transformed this landscape by enabling neural networks to autonomously learn and extract hierarchical features directly from raw input data. In earthquake monitoring, these approaches have demonstrated remarkable success across multiple tasks, including event detection and seismic phase picking (e.g., [7][8][9][10][11]), magnitude estimation (e.g., [9,[12][13][14]), and first-motion polarity estimation (e.g., [9,15]). The rapid progress in deep learning-based seismic discrimination has been strongly supported by the development and curation of large, high-quality benchmark datasets, such as STEAD [16] and INSTANCE [17], which have enabled standardized training, validation, and comparative evaluation of models.
The earliest application of neural networks to the problem of distinguishing earthquakes from blasts dates back to 1990, when [18] proposed a simple feed forward network trained on regional spectral data which achieved accuracy above 93% on their test set. In the late 2010s, researchers revisited this problem using deep learning, for instance [19] explored the use of convolutional and recurrent neural networks to discriminate explosive and tectonic sources for local distances. [20] used a deep learning model based on convolutional neural networks (CNNs) to classify natural earthquakes and blasts, achieving promising results on a dataset recorded in Hainan. [21] combined a Compact Convolutional Transformer with a Capsule Neural Network to distinguish earthquakes from quarry blasts using data recorded by the Egyptian Seismic Network (ENSN) from January to December 2021. [22] developed EQTypeNet, a tri-branch CNN that integrates waveform, spectrogram, and event P/S ratio features for seismic event classification and detection across China. Similarly, [23] addressed the classification of three categories of seismic events: quarry blasts, earthquakes, and noise, by developing CNN models trained on labeled waveform data recorded at the SUR station of the Gujarat State Seismic Network between 2007 and 2022.
Despite these successes, purely data-driven deep learning models are often viewed as “black boxes” that lack physical transparency and struggle with generalization when applied to regions outside their training data. To address this Physics-Informed Neural Networks (PINNs) have emerged as a promising solution that embeds physical laws directly into neural network training. Early efforts by [24] demonstrated that combining deep learning with physics-based features, such as the P/S wave amplitude ratio, significantly enhanced the generalization performance of models when applied to unseen regions. Further advancements have utilized PINNs to solve more complex geophysical problems related to source characterization. [25] introduced a Bayesian framework that integrates seismic waveform features with geospatial context to jointly classify event types and quantify their
This content is AI-processed based on open access ArXiv data.