Development of an AI-based Seismic Signal Classification System for Earthquake Detection

Thesis Abstract

The increasing frequency and complexity of seismic activities necessitate the development of advanced, accurate, and timely earthquake detection systems to mitigate associated risks and enhance early warning capabilities. Traditional seismic monitoring methods, primarily reliant on manual interpretation of seismic signals, often face limitations related to latency, scalability, and sensitivity to noise, which hinder prompt and reliable earthquake detection—particularly in regions with dense seismic networks or substantial anthropogenic noise. This study aims to develop a robust artificial intelligence (AI)-based seismic signal classification system that automatically distinguishes earthquake signals from non-seismic noise, thereby improving detection accuracy and response times. The specific objectives of the research include (1) examining the characteristics of seismic signals associated with earthquake events and non-earthquake activities; (2) designing a machine learning model, utilizing deep learning techniques such as convolutional neural networks (CNNs), trained on a comprehensive dataset of labeled seismic recordings; (3) evaluating the performance of the model against existing classification methods in terms of accuracy, precision, recall, and F1-score; (4) implementing real-time data processing and visualization interfaces for operational deployment; and (5) assessing the system's robustness across different seismic noise environments and regional settings. The research adopts a quantitative, experimental design, integrating signal processing, machine learning, and statistical validation methodologies. The population comprises seismic recordings sourced from a global network of 150 seismic stations, with a focus on data collected over a span of five years (2017-2022). A stratified random sampling technique is employed to select a representative subset of approximately 10,000 labeled seismic signals encompassing diverse earthquake magnitudes, depths, and non-seismic noise types such as cultural, atmospheric, and industrial disturbances. Data collection instruments include digital seismic sensors providing continuous waveform data, supplemented by existing curated datasets from national and international seismic agencies. Data preprocessing involves noise filtering, feature extraction via spectrogram analysis, and normalization tailored for input into deep learning models. The core of the methodology involves training and validating a CNN architecture optimized through hyperparameter tuning, with the training process guided by supervised learning principles. Model performance is evaluated using confusion matrices, ROC curves, and metrics including accuracy, precision, recall, and F1-score, with cross-validation to prevent overfitting. Additionally, the study applies statistical significance testing, such as paired t-tests, to compare the proposed model against baseline classifiers like Support Vector Machines (SVM) and Random Forests. Expected findings include a significant increase in seismic event detection accuracy compared to traditional automated classifiers, with the CNN model achieving over 95% accuracy and high sensitivity in detecting low-magnitude earthquakes. The system is anticipated to demonstrate resilience in noisy environments and adaptability to different regional seismic characteristics, confirming its operational viability for real-time earthquake early warning systems. This research contributes novel insights into the application of deep learning techniques for seismic signal processing, filling a notable gap in existing literature concerning the integration of AI for real-time seismic monitoring, especially in resource-limited settings. It advances the theoretical framework by aligning signal characteristics with modern computational models, supporting the development of scalable, automated seismic analysis tools. The findings have the potential to influence policy and practice by informing the design of more responsive early warning systems, ultimately improving public safety, disaster preparedness, and emergency response strategies. In conclusion, the study underscores the efficacy of AI-driven approaches in seismic signal classification and advocates for their wider adoption in operational seismic networks. Recommendations include further research into multisensor data fusion, expanded validation across diverse geographic regions, and the integration of the system with existing seismic infrastructure for comprehensive earthquake risk mitigation.