Machine learning based resolution improvement of velocity structures with prior model

It is important to derive a high-resolution velocity model from seismic data because it provides information for the accurate subsurface imaging as well as the additional information of subsurface material which cannot be obtained from the subsurface structure image. Conventionally, algorithms such as Full Waveform Inversion (FWI) or stochastic inversion, which use low-resolution velocity structures as initial models, were mainly used to obtain high-resolution velocity structures. However, these algorithms require intensive computational resources for high-resolution velocity inversion. Recently, due to the improvement of computing technology and the advancement of machine learning algorithms, studies using machine learning have been widely conducted in various fields. Studies using machine learning are being actively conducted in the field of seismic data processing. In this study, machine learning was used to efficiently derive a high-resolution velocity model with low computational cost. As input data, a low-resolution velocity model and seismic migration data were used. Additionally, for higher accuracy, a prior model constructed based on well data and a weight model containing uncertainty information of the prior model were also used. To sum up, four types of data, which were low-resolution velocity model, migration data, prior model and weight model, were used to construct the training dataset for the supervised machine learning. The trained model was applied to the test dataset which were not used during training stage and the result showed that the proposed method could recover the accurate high-resolution velocity model from the low-resolution velocity information.