Thesis
Seismic signal classification and detection based on deep learning
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17355
- Person Identifier (Local)
- 202088427
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Passive seismic monitoring is important for understanding subsurface processes such as landslides, mining activities, and geothermal systems, enabling the prediction and mitigation of their effects. However, continuous seismic monitoring produces vast datasets containing various sources of seismicity that require accurate classification. Manual detection and labeling of these events is both time-consuming and prone to inconsistency, even when performed by the same expert. To address these challenges, this thesis first proposes an automated joint detection and classification method for characterizing seismic events using Convolutional Neural Networks (CNNs). Despite their effectiveness and high accuracy, deep learning models, such as CNNs, face two significant limitations: the lack of interpretability due to their “black-box” nature and the large amount of manually labeled data required for training. Interpretability is particularly important in seismic applications where reliable detection and classification of earthquakes are essential for infrastructure safety. To ensure both accuracy and explainability, the second contribution of this thesis is a novel methodology for data labeling, verification, and re-labeling through CNNs enhanced by Layer-wise Relevance Propagation (LRP), a popular explainable AI tool. This approach aims to provide transparency in seismic event detection, improving trustworthiness in AI-driven decisions. Manual labeling of seismic events is often inefficient and contradictory to the goal of automated detection. To overcome the time and resource inefficiency of manual labeling, the third contribution of the thesis is a self-supervised learning (SSL) model that reduces the dependency on large amounts of annotated data while maintaining high detection accuracy. This model significantly reduces the manual effort involved in labeling seismic events, thereby improving the efficiency and scalability of seismic monitoring systems. The proposed CNN-based models achieve approximately 90% accuracy, effectively distinguishing between seismic sources such as rockfalls and earthquakes. Furthermore, the LRP-based method enhances interpretability of seismic classification by generating explainable relevance maps, which visually highlight the most influential parts of the seismic signal that contributed to the model’s decision. These maps help experts understand model reasoning, validate whether the model is focusing on geophysically meaningful features, and identify potential mislabeling or overlooked patterns in the data.The SSL model achieves accuracy comparable to state-of-the-art fully supervised methods while requiring only 5% to 30% of the labeled data. Specifically, the lower end (around 5%) is sufficient for distinguishing well-separated classes like rockfalls, while more complex scenarios with overlapping signal characteristics, such as those between quakes and a subset of earthquakes, may require up to 20% labeled data. This flexibility significantly reduces the manual labeling burden without sacrificing detection precision. Together, these contributions offer an accurate, reliable, efficient, and explainable deep learning-based framework for seismic event detection and classification, advancing the state of seismic signal monitoring and analysing.
- Advisor / supervisor
- Pytharouli, Stella
- Stankovic, Vladimir
- Stankovic, Lina
- Resource Type
- DOI
Relations
Items
Thumbnail | Title | Date Uploaded | Visibility | Actions |
---|---|---|---|---|
|
PDF of thesis T17355 | 2025-06-03 | Public | Download |