Extracting volcanic rock velocity from reflection seismic data using deep learning

The accuracy of the velocity field stands as the foremost determinant impacting the quality of migration imaging, thus underscoring the significance of establishing a robust velocity model in the context of complex geological body imaging. However, within the realm of volcanic rock development, the intricate lithology and lithofacies of volcanic rock, alongside spatial overlap and scale variability, have presented formidable challenges for conventional manual methodologies in explicating the spatial distribution of volcanic rock masses. Consequently, these methods have struggled to furnish precise structural interpretation models suitable for velocity modeling techniques like grid tomography. In this study, we have leveraged deep learning tools to effectively and precisely delineate the spatial distribution range of volcanic rock masses from seismic data within the imaging domain, thereby reshaping the technical process of grid tomography velocity modeling. Addressing the problem of volcanic rock mass detection as a semantic segmentation task, we trained a network to execute pixel-by-pixel prediction aimed at identifying pixels corresponding to a high likelihood of volcanic rock mass presence. To enhance the network’s recognition accuracy, this study introduced a cavity convolution and two functional modules, augmenting the performance of a conventional U-Net. The proposed methodology utilizes three-dimensional reflection seismic data for network training and validation. Ultimately, a practical dataset is employed to substantiate the reliability and efficacy of the method.