순환신경망 기반 천해역 탄성파 탐사자료의 해수면 기인 다중반사파 제거 연구

다국어 초록 (Multilingual Abstract)

Shallow-water marine seismic data are characterized by strong surface-related multiples with short-period features caused by the shallow water depth. These surface-related multiples generate spurious reflection events that do not correspond to actual subsurface interfaces, thereby hindering accurate geological interpretation. Consequently, effective suppression of surface-related multiples is essential for precise subsurface imaging and reliable geological analysis. To this end, conventional multiple-suppression techniques based on physical principles have been widely employed, including predictive deconvolution, the Radon transform, and surface-related multiple elimination (SRME), which are generally applied to pre-stack seismic data. However, when these conventional methods are applied to shallow-water seismic data, their performance deteriorates significantly due to the strong similarity in traveltime and phase characteristics between primary reflections and surface-related multiples. As a result, residual multiples that are not fully removed at the pre-stack stage often remain in the stacked section. To overcome the limitations of conventional approaches, recent studies have increasingly focused on deep-learning-based multiple suppression techniques. In this study, a deep-learning-based approach is proposed to directly suppress surface-related multiples in seismic stack sections using a recurrent neural network–based bidirectional long short-term memory (BiLSTM) architecture. To generate synthetic training data that realistically represent shallow-water seismic conditions, the water depth and acquisition geometry of the target survey area were taken into account, and synthetic seismic time-series data were generated using a source wavelet estimated from field data. In addition, to enable the network to learn not only the temporal characteristics but also the frequency-dependent behavior of surface-related multiples, the continuous wavelet transform (CWT) was applied to construct time–frequency domain input data for network training. Surface-related multiple suppression was performed by predicting the multiple components using the trained network and subtracting the predicted multiples from the raw seismic data. The application to synthetic data demonstrated that the proposed BiLSTM network accurately predicted the traveltime and phase characteristics of surface-related multiples, showing strong agreement with the reference data. Quantitative evaluation using the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) further confirmed the superior prediction performance of the proposed method. Furthermore, application to field data acquired in the Arctic shallow-water environment showed effective attenuation of surface-related multiples and improved continuity of primary reflection events, thereby validating the practical applicability of the proposed approach. This study presents an alternative framework that complements the limitations of conventional multiple-suppression methods in shallow-water and coastal seismic data processing and is expected to contribute to the automation and advancement of marine geophysical data processing. Keywords : Shallow water, Seismic exploration, Surface-related multiple, Deep-learning, Bidirectional long short-term memory (BiLSTM)

목차 (Table of Contents)

• List of Tables·ⅲ
• List of Figuresⅲ
• ABSTRACT·ⅵ
• 1. 서론 1
• 1.1 연구의 필요성· 1
• 1.2 연구의 구성 4
• 2. 이론 5
• 2.1 해양 탄성파 탐사 5
• 2.1.1 자료취득 5
• 2.1.2 자료처리 8
• 2.1.2.1 전처리· 9
• 2.1.2.2 중합 전 처리 9
• 2.1.2.3 중합 후 처리 10
• 2.1.3 자료해석 11
• 2.2 다중반사파 12
• 2.2.1 해수면 기인 다중반사파 12
• 2.2.2 내부 다중반사파 12
• 2.3.3 페그레그 다중반사파 12
• 2.3 전통적 다중반사파 제거 기법· 15
• 2.3.1 예측 디콘볼루션 15
• 2.3.2 라돈 변환 18
• 2.3.3 자유면 기인 다중반사파 제거 20
• 2.4 딥러닝 네트워크· 23
• 2.4.1 순환신경망 23
• 2.4.2 장단기 기억 신경망 25
• 2.4.3 양방향 장단기 기억 신경망 27
• 3. 연구 방법 29
• 3.1 천해역 탐사 환경을 반영한 합성자료 생성· 30
• 3.1.1 1차 반사파 신호 생성· 31
• 3.1.2 해수면 기인 다중반사파 모사 33
• 3.1.3 반사파 합성 및 잡음 추가 35
• 3.2 입력자료 변환기법을 통한 학습자료 생성 37
• 3.3 다중반사파 예측 네트워크 설계 40
• 3.4 다중반사파 예측 네트워크 학습 41
• 3.5 해수면 기인 다중반사파 예측 성능 분석 44
• 4. 수치 예제 47
• 4.1 현장자료 취득 탐사해역 특성· 47
• 4.2 현장 유사 합성자료를 통한 예측 성능 분석 49
• 4.3 천해역 탄성파 탐사자료 해수면 기인 다중반사파 제거 52
• 4.4 입력자료 변환 기법에 따른 다중반사파 제거 성능 비교· 58
• 5. 논의 60
• 6. 결론 61
• 참고문헌· 63
• 국문초록· 71