Interpretability-Guided Convolutional Neural Networks For Seismic Fault Segmentation

Delineating the seismic fault, which is an important type of geologic structures in seismic images, is a key step for seismic interpretation. Comparing with conventional methods that design a number of hand-crafted features based on the observed characteristics of the seismic fault, convolutional neural networks (CNNs) have proven to be more powerful for automatically learning effective representations. However, the CNN usually serves as a black box in the process of training and inference, which would lead to trust issues. The inability of humans to understand the CNN would be more problematic, especially in critical areas like seismic exploration, medicine and financial markets. To include domain knowledge to improve the interpretability of the CNN, we propose to jointly optimize the prediction accuracy and consistency between explanations of the neural network and domain knowledge. Taking the seismic fault segmentation as an example, we show that the proposed method not only gives reasonable explanations for its predictions, but also more accurately predicts faults than the baseline model.