LAUSR

Fault interpretation is very important for reservoir characterization in seismic petroleum exploration. Recently, different machine learning methods have been widely performed in fault detection of seismic data. Faults have multiscale characteristics and present abrupt changes in seismic data. Based on this characteristic, we introduce a three-dimensional scattering wavelet transform (3-D SCWT) convolutional neural network (CNN) method, called 3-D SCWTnet, which contains the 3-D SCWT and the 3-D Unet. Three-dimensional SCWT has multiscale and multidirection characteristics to delineate the spatial characteristics at different scales and angles of the faults in seismic data. Three-dimensional Unet is the 3-D CNN for the faults classification task. The 3-D SCWTnet makes full use of the multiscale properties of 3-D SCWT and deep learning network. The synthetic seismic data is used as the training data for the 3-D SCWTnet, the predicted results of synthetic data and field seismic cube obtain a higher accuracy than the traditional 3-D Unet.