부산대학교 도서관

Ambient seismic noise tomography has been applied to probe the Earth's structure. To accurately map geological structures, a considerable amount of time is required to pick fundamental and higher modes of dispersion curves. We used frequency‐Bessel (F‐J) transform to calculate phase velocity‐frequency diagrams to obtain more higher modes of dispersion curves from the cross‐correlation function. Several studies have recently focused on picking dispersions automatically using deep learning to reduce time consumption. However, the generalization of neural networks has a degradation for untrained diagrams to some degree. Here, based on domain adaptation in computer vision, we used gamma transform to change the image contrast of the phase velocity‐frequency diagrams, rendering the test data closer to the training data. Introducing domain adaptation into the dispersion region extraction effectively improves the generalization of the neural network. Here, the dispersion regions are the regions in the phase velocity‐frequency diagram located around the dispersion curves where the energy is above a given threshold. We validated our method by using phase velocity‐frequency diagrams from different areas. We used one synthetic phase velocity‐frequency diagram and three phase velocity‐frequency diagrams of different areas to test domain adaptation. In particular, we tested one phase velocity‐frequency diagram that belongs to the same area for the training diagram, except for the processing steps. The results showed that our domain adaptation method successfully enhanced the generalization of the neural network. After domain adaptation, our trained network could effectively extract more higher modes of dispersion regions than before. Our dispersion curves picking method combined with domain adaptation can pick sufficient dispersion information for numerous phase velocity‐frequency diagrams. Furthermore, our method can facilitate the study of ambient noise tomography and illumination of the Earth's interiors. Our research provides a strategy for enhancing the generalization of neural networks for other deep learning‐based geophysical image segmentation tasks. Plain Language Summary: Seismic surface wave has been used to probe the structure of our planet at regional and global scales. To pick dispersion curves efficiently and quickly, several researchers have attempted to utilize neural networks for the automatic picking of dispersion curves. However, in the automatic dispersion picking method, an inadaptation phenomenon occurs for different types of data that originate from different geological areas or have been processed by different manual processing methods. We used gamma transform to conduct image transformation in an attempt to alleviate inadaptation for different types of data. We verified the results of our domain adaptation with several samples. Changing the image contrast (i.e., amplitude of image) via gamma transform, we improved the performance of the trained neural network and used it to artificially control the quantity to which regions are extracted. Our method enabled solving the inadaptation problem. Key Points: The extraction ability of neural network using domain adaptation was enhanced significantly and verified by untrained phase velocity‐frequency diagramsThe modified gamma transform effectively enhanced the extraction of dispersion regions and simultaneously suppressed the noiseOur results validate that the extent to which dispersion regions are extracted could be controlled by a change in gray‐level distribution [ABSTRACT FROM AUTHOR]