Automated Classification of Seismic Signals for Real-Time Hazard Monitoring

Landslides and earthquakes are among the most significant natural hazards, often generating seismic signals with partially overlapping characteristics that complicate reliable discrimination, particularly in mountainous and tectonically active regions. Accurate identification of landslide-induced seismic signals is essential for developing reliable landslide catalogs, improving hazard assessment, and enabling real-time monitoring systems. In this study, we present an advanced machine learning and deep learning–based framework for the classification of seismic signals associated with landslides and earthquakes, using real observational data and dimensionality-reduction techniques. Seismic waveform data were collected from permanent seismic stations operated by the Seismological Observatory and the Earthquake Engineering Department at the Indian Institute of Technology Roorkee. Earthquake events were identified using established regional and global earthquake catalogues, while the landslide catalogue was independently developed by our research group through systematic analysis of seismic records, field evidence, and event validation. This self-developed landslide catalogue provides a high-confidence dataset for supervised learning and represents a significant contribution to regional mass-movement monitoring efforts. The seismic signals were initially characterized using a comprehensive set of signal descriptors derived from previous studies on landslide and earthquake seismology. Approximately 97 time-domain, frequency-domain, and statistical parameters were extracted for each event, capturing waveform amplitude, energy distribution, spectral content, and temporal evolution. While these features effectively describe seismic signal behavior, their high dimensionality introduces redundancy and may degrade classification performance. To address this challenge, multiple Principal Component Analysis (PCA) approaches, including conventional and kernel-based PCA, were employed to reduce dimensionality while preserving the most informative components relevant for class discrimination.

Following dimensionality reduction, advanced machine learning classifiers were applied to distinguish between landslide- and earthquake-generated seismic signals. The classification framework was trained using combinations of real data and synthetically augmented samples generated through CTGAN (Conditional Tabular Generative Adversarial Network) to improve class balance and model robustness. Model performance was evaluated using independent test datasets derived from raw, unseen seismic signals, ensuring a realistic assessment of generalization capability. Across different PCA–classifier combinations, the proposed framework achieved high classification accuracy, consistently exceeding 95% and reaching values close to 97% for optimal model configurations. Precision, recall, F1-score, and ROC–AUC metrics further demonstrate the reliability and stability of the classification results. Importantly, the trained models were validated directly on raw seismic data, highlighting their ability to generalize beyond feature-engineered training sets. This result indicates strong potential for operational deployment. The proposed methodology provides a scalable and automated approach for discriminating landslide-induced seismicity from earthquakes and can be integrated into continuous seismic monitoring systems.

Overall, this study demonstrates the effectiveness of combining seismic signal processing, dimensionality reduction, and advanced machine learning for landslide detection. The developed framework has significant implications for the real-time development of accurate landslide catalogs and offers a promising pathway toward improving early warning capabilities using continuous data streams from regional seismometer networks.