This study presents a novel neural network model to explore its application in automatically interpreting subsurface faults from seismic images. A Wavelet Convolutional Neural Network (CNN) model that incorporates discrete wavelet decomposition is presented, and its capability in segmenting subsurface faults is analyzed.
In this study, different neural network models are developed to compare their performance in segmenting subsurface faults. Sliced 2D seismic images are used as the input of the models. Pre-interpreted images with fault locations are used as the output of the models. Different CNN models are created using different pooling methods, including a traditional U-Net model with average pooling method, and an advanced Wavelet CNN model using wavelet pooling method. The results show that the Wavelet CNN model, which incorporates discrete wavelet transformation as the pooling layer, has the best performance comparing to traditional models in segmenting subsurface faults from input seismic images. It is more effective in saving edge features during pooling operations and outperforms the traditional U-Net model in segmenting subsurface faults from seismic images.