Contemporary deep learning models in seismology rely on ground motion records for various tasks. The study evaluates the influence of auxiliary information on model effectiveness, highlighting a potential gap in methodologies for deep learning from single-station ground motion recordings. Experimental results show a strong reliance on P and S phase arrival information, indicating the need for further research to enhance seismic data analysis with deep learning techniques.
The study uses two fundamental CNN architectures, ResNet and TCN, to assess their efficacy in learning from ground motion records. By conducting hyperparameter searches and ablation studies, the authors investigate how deeply models can learn from only ground motion records without auxiliary information like P/S phase arrival times. The experiments reveal a significant improvement in model performance when incorporating P/S phase information as an input channel.
The findings suggest a strong correlation between epicentral distances and P/S phase information, emphasizing the importance of this auxiliary data for accurate predictions. The study concludes that TCN models outperform ResNet models and that including P/S phase information enhances training capabilities while reducing overfitting behavior. Future research is needed to explore alternative architectures tailored to localized seismic data for improved seismic data analysis.
To Another Language
from source content
arxiv.org
Deeper Inquiries