Enhanced Automatic Segmentation of Salt Bodies from Seismic Images Using Wavelet Convolutional Neural Networks

Abstract

Summary With the improvement of computational power and big data analytics methods, automatic seismic data processing and interpretation methods are being developed to extract information from seismic data and make interpretations more efficiently. This study explores the methods in automatically identifying salt bodies from seismic images using a novel convolution neural networks (CNN) combined with wavelet decomposition analyses. Traditional CNN models use max pooling or mean pooling as the pooling layers. However, there exists limitations for these two pooling methods, which may result in loss of details from the original input. This study combines wavelet decomposition with CNN and applies it for the task of identify salt bodies from seismic images. Since the input seismic images are synthesized from reflected sound waves, adding wavelet analyses to the CNN model is expected to increase the model performance by improving the ability to capture wave features. The results show that the wavelet CNN model has a better performance compared to traditional models by incorporating wavelet decomposition. The model has a higher testing accuracy in identifying salt bodies from seismic images.