The Fault Detection Method of the Seismic Image Based on Semantic Segmentation

Seismic image fault detection is of practical significance in solid mineral resources investigation, petroleum geology investigation, and other multidisciplinary fields. Traditional fault detection requires manual interpretation, limited by inefficiency and subjectivity. Nowadays, deep learning methods have significantly improved segmentation tasks, but most of these methods were supported by multitudinous seismic data. However, the semantic segmentation method achieves enhanced performance in insufficient data. This paper proposed Channel-UNet, a novel U-Net architecture that integrates the channel attention module to recalibrate feature weight adaptively. Simultaneously, to obtain adequate seismic images, we synthesize a dataset of 2000 images containing apparent fault features and corresponding labels. The experimental results indicate that our network has the highest average IOU and Dice coefficient on the synthetic seismic datasets compared with other semantic segmentation networks. In addition, we further confirm that when applying the Hough Transform as a post-process, the results of Channel-UNet on the actual seismic images perfectly match the predicted results and labels given by the interpreter. The experimental results reveal that the Channel-UNet delineates the fault features in seismic sections more accurately.