검색 상세

Analysis of Seismicity using Deep Learning Methods in the Southern Korean Peninsula

심층학습을 활용한 한반도 남부 지역의 지진 활동 분석

  • Jongwon Han (한종원, 대학원 지구환경과학과 지구물리학전공)

초록/요약

This thesis presents a deep learning-based seismic analysis framework and a long-term earthquake catalog by utilizing a local earthquake detection model optimized for the southern Korean Peninsula. Furthermore, it aims to enhance automated earthquake analysis techniques and systematically generate foundational datasets. As a case study, the deep learning-based framework is applied to the southeastern and southwestern regions of the Korean Peninsula to validate its effectiveness, thereby highlighting the significance of compiling a comprehensive earthquake catalog for further intraplate seismic studies. At first, using the aftershock monitoring data of the 2016 ML 5.8 Gyeongju earthquake sequence, the study demonstrated that deep learning-based earthquake detection models can detect more microearthquakes with faster computation speed and higher accuracy than conventional methods, by recognizing previously undetected small-magnitude seismic signals. However, limitations such as the inability to discriminate blast signals and omission of large earthquakes highlighted the need for optimized training and further model development tailored to the study region. Based on this need, a retrained deep learning model using the seismic dataset recorded in the Korean Peninsula was developed, and it created a new catalog of earthquakes for the southern Korean Peninsula from 2012 to 2021. This catalog includes an additional 21,475 earthquakes not reported in the previous routine catalog, enabling a more detailed understanding of natural seismicity within Korea. Throughout this process, I also achieved updates in the 1D velocity model, development of station correction terms, adjustments in local magnitude, blast signal discrimination, and aftershock de-clustering, establishing a comprehensive analytical framework and presenting various seismological characteristics. My methods also identified northeast-and northwest-linear alignments of seismicity, suggesting the influence of pre-existing faults and boundaries between crustal provinces. This implies reactivation of some pre-existing faults in intraplate regions. My interpretation is supported by analyses with Mohr circle for stress which shows the current stress regime favoring the reactivation of the nearly orthogonal pre-existing faults rather than reactivation of optimal conjugate faults. To perform more refined focal mechanism analyses on the extensive seismic catalogs, the RPNet model was developed for automatic P-wave first-motion polarity determination. Incorporating advanced deep learning techniques such as inception modules and attention mechanisms, the RPNet offers robust performance under uncertain P-wave onset conditions. By employing Monte Carlo dropout, the model also provides quantifiable prediction uncertainties, allowing for a more reliable assessment of results. This enhances the accuracy of focal mechanism analysis, supporting a deep understanding of the faulting mechanisms and stress distribution through the reliable analysis of large-scale seismic data. This thesis includes analyzing seismogenic faults and earthquake clusters in southeastern Korea (near the 2016 ML 5.8 Gyeongju earthquake region) and southwestern Korea (near the 2024 ML 4.8 Buan earthquake region). In the southeastern region, data from the Gyeongju Hi-density Broadband Seismic Network (GHBSN) were used to analyze seismogenic faults and focal mechanisms, revealing reactivation of faults associated with Miocene sedimentary basins under the current compressive stress regime. In the southwestern region, analysis of the 2024 ML 4.8 Buan earthquake cluster highlighted seismogenic fault segments revealed by aftershock distribution, with evidence of linearly distributed earthquake clusters near the Okcheon Belt. This thesis demonstrates the utility of deep learning in efficiently processing large-scale seismic data and systematically analyzing intraplate earthquake patterns, contributing to essential foundational technologies and data for advancing the understanding of seismogenic fault behavior mechanisms in intraplate regions.

more

목차

ABSTRACT i
국문 초록 iv
PREFACE vii
TABLE OF CONTENTS viii
LIST OF FIGURES xii
LIST OF TABLES xxiii
CHAPTER 1. Introduction 1
1.1. General introduction 1
1.2. Research overview 5
CHAPTER 2. Methodological background 7
2.1 Deep learning applications in seismology 7
2.2 Deep learning phase picker 11
2.3 Deep learning techniques 12
2.3.1. Convolutional Neural Network 12
2.3.2. Recurrent Neural Network 14
2.3.3. Attention mechanism 16
2.3.4. Inception module 17
2.3.5. Monte Carlo dropout 18
2.3.6. Seismic phase association using the Gaussian Mixture model 19
2.3.7. Aftershock de-clustering using the random forest model 20
CHAPTER 3. Seismic event and phase detection using deep learning for the 2016 Gyeongju earthquake sequence 22
3.1. Abstract 22
3.2. Introduction 23
3.3. Data and Method 26
3.3.1. Data, Processing, and Previous Event Catalogs 26
3.3.2. DL Model for Earthquake Detection and Phase Picking 29
3.3.3. Post-processing of the DL Model Results Based on Waveform Similarity 29
3.4. Result and Discussion 31
3.5. Conclusion 42
CHAPTER 4. Research catalog of inland seismicity in the southern Korean Peninsula from 2012 to 2021 using deep learning techniques 44
4.1. Abstract 44
4.2. Introduction 46
4.3. Data processing and workflows 52
4.4. Result and discussion 55
4.4.1. Retraining of a deep learning picker using hybrid datasets and subsequent event detection 55
4.4.2. Event discrimination 61
4.4.3. Adjustments for the event parameter determination 64
4.4.4. Comparison of the proposed catalog with the KMA catalog 68
4.4.5. De-clustering and b-value estimation of natural seismicity 74
4.4.6. Pattern of natural seismicity 75
4.5. Conclusion 85
4.6. Supplementary 87
4.6.1. Magnitude equation 87
4.6.2. 1D velocity model update 88
4.6.3. Event declustering based on the random-forest algorithm (Aden Antoniów et al., 2022) 89
CHAPTER 5. RPNet: Robust P-wave first-motion polarity determination using deep learning 91
5.1. Abstract 91
5.2. Introduction 92
5.3. Data for training and verification 96
5.4. Benchmark test for previous deep-learning models 99
5.5. RPNet model structures and training 104
5.6. Model performance and verification 107
5.6.1. RPNet model structures and training 107
5.6.2. Verification using Hi-net data 111
5.6.3. Model application to the 2016 MW 7.0 Kumamoto earthquake sequence 113
5.7. Conclusion 117
5.8. Supplementary 119
5.8.1. Quality assessment of unknown (K) labels for 1,000 randomly selected data 119
5.8.2. mRPNet: A minor version of RPNet predicting the unknown (K) class 122
5.8.3. Model performance on the Hi-net verification dataset using the receiver operating characteristic (ROC) curves 124
CHAPTER 6. Seismogenic faults and linearly distributed intraplate earthquake clusters in local areas of the Korean Peninsula 125
6.1. Introduction 125
6.1.1. Gyeongju Hi-density Broadband Seismic Network in the southeast Korean Peninsula (chapter 6.2) 126
6.1.2. 2024 ML 4.8 Buan earthquake in the southwest Korean Peninsula (chapter 6.3) 126
6.2. Seismogenic faults in the southeastern Korean Peninsula 128
6.2.1. GHBSN and training dataset 128
6.2.2. Method 130
6.2.3. Deep learning model optimization 132
6.2.4. Seismicity of the southeastern Korean Peninsula 136
6.2.5. Focal mechanisms 142
6.2.6. Interpretations on seismogenic fault segments 145
6.3. Linearly distributed earthquake clusters in the southwest Korean Peninsula 149
6.3.1. Seismic continuous waveform and the earthquake catalog 149
6.3.2. Analysis process of ML 4.8 Buan earthquake sequence 151
6.3.3. Analysis process of earthquake clusters in adjacent region 154
6.3.4. 2024 ML 4.8 Buan earthquake sequence 158
6.3.5. Linearly distributed intraplate earthquake clusters 160
6.4. Conclusions 163
CHAPTER 7. Summary and conclusion 164
7.1. Summary 164
7.2. Conclusion 167
REFERENCES 170
APPENDICES 189

more
반출 Meta View 목록