17:15 〜 19:15
[SSS06-P06] Machine Learning-based Surface Wave Analysis for High-Resolution Seismic Imaging in North China
キーワード：Machine learning, North China, Dispersion curve extraction, Eikonal tomography
Surface wave dispersion curves can be extracted from ambient noise cross-correlation functions between seismic stations for subsurface imaging. In dense seismic arrays, the large number of stations results in an overwhelming volume of dispersion curves, making manual extraction impractical. To address this, we develop a neural network-based method for automated extraction. Using data from the ChinArray in eastern North China, we apply an image analysis approach to extract 19,163 dispersion curves for training. The trained model achieved a root mean square error of less than 0.01 km/s in the 5–50 s period range compared to automatically extracted curves.
To enhance the generalization capability of the neural network, we perform transfer learning using a new small dataset including 2000 pairs of dispersion curves and spectrograms from western North China. This fine-tuning expands the model's effective period range and significantly improves performance with a small dataset.
The quality of amplitude spectrograms is a key factor in dispersion curve prediction. While we filter dispersion curves based on the signal-to-noise ratio of cross-correlation functions, quality control of amplitude spectrograms remained limited. To address this, we design a simplified VGG Net-D architecture for spectrogram classification, reducing layers to balance performance and computational efficiency. The difference in phase velocity error between predicted and extracted dispersion curves remained below 0.1 km/s, with higher errors at short periods and study area boundaries due to sparse data coverage.
Finally, we derive Rayleigh wave phase velocity distributions via eikonal tomography. Results indicate that short-period waves (8–12 s) show low phase velocity in sediment-rich basins such as Hetao Graben, Shanxi Rift, and Bohai Bay Basin, and high phase velocity in mountainous and uplift regions. At 20 s period, low phase velocity anomalies from the Datong volcanic area extend westward beneath Hetao Graben and southward toward Shanxi Rift, suggesting deeper crustal structures. For periods longer than 24 s, the Bohai Bay Basin maintains its low phase velocity, while the Datong volcanic area shows more pronounced anomalies, potentially indicating a magmatic upwelling channel.
To enhance the generalization capability of the neural network, we perform transfer learning using a new small dataset including 2000 pairs of dispersion curves and spectrograms from western North China. This fine-tuning expands the model's effective period range and significantly improves performance with a small dataset.
The quality of amplitude spectrograms is a key factor in dispersion curve prediction. While we filter dispersion curves based on the signal-to-noise ratio of cross-correlation functions, quality control of amplitude spectrograms remained limited. To address this, we design a simplified VGG Net-D architecture for spectrogram classification, reducing layers to balance performance and computational efficiency. The difference in phase velocity error between predicted and extracted dispersion curves remained below 0.1 km/s, with higher errors at short periods and study area boundaries due to sparse data coverage.
Finally, we derive Rayleigh wave phase velocity distributions via eikonal tomography. Results indicate that short-period waves (8–12 s) show low phase velocity in sediment-rich basins such as Hetao Graben, Shanxi Rift, and Bohai Bay Basin, and high phase velocity in mountainous and uplift regions. At 20 s period, low phase velocity anomalies from the Datong volcanic area extend westward beneath Hetao Graben and southward toward Shanxi Rift, suggesting deeper crustal structures. For periods longer than 24 s, the Bohai Bay Basin maintains its low phase velocity, while the Datong volcanic area shows more pronounced anomalies, potentially indicating a magmatic upwelling channel.