DEEP LEARNING-BASED SPECTRAL ENHANCEMENT OF AMBIENT SEISMIC NOISE DATA FOR IMAGING UPPER MANTLE STRUCTURES (Invited Presentation)

Enhancing the vertical resolution of seismic ambient noise is important to understand the Earth's interior and its behavior. Particularly, ambient seismic noise data are utilized to invert subsurface seismic velocity models and interpret the upper mantle structures. However, there is a limit to directly conduct quantitative seismic inversion and interpretation methods to these noise data due to their low signal-to-noise ratio (SNR) and narrow frequency bandwidth. Although we can effectively improve the SNR by using current seismic processing techniques to overcome the limitations, it is challenging to broaden the frequency bandwidth of the original seismic data. This study presents a deep learning application to extend the spectral bandwidth of recorded noise data and image the lithospheric structures such as Moho and Lithosphere-Asthenosphere-Boundary (LAB) horizons. First of all, we use the seismic noise data recorded in the Korean Peninsula, 2014. Before the spectral broadening, we improve the SNR of these datasets by using seismic interferometry and reflection processing methods. Then, we apply a 1-D convolutional U-Net deep learning method to these datasets to generate synthetic seismic traces. Especially, we train the U-net model by using the prior information, for example, the estimated wavelet and reflectivity series. Second, we model the spectrally broadened traces corresponding with the input data based on the successfully trained U-net model. As we compare the amplitude spectra before and after our deep learning application, the applied data have broader frequency spectra (0-6.5Hz), while the original data have 0-3.0 Hz. Third, we perform seismic attribute analysis to emphasize the subsurface information in the improved data, like sparse-spike deconvolution and amplitude envelope attributes. From the attribute sections, we effectively identify the Moho and LAB, better than the raw seismic sections. In conclusion, we broaden the frequency spectrum of seismic data and improve its vertical (temporal) resolution based on deep learning. It is also evident that this enhancement allows us to invert the reliable seismic velocity models and interpret more intense upper mantle structures.