Deep Learning for Detecting Thrust Faults in Subduction Zones

Subduction plate boundary faults and splay faults in accretionary wedges are capable of generating some of the largest earthquakes and tsunamis on Earth. Owing to the complexity of geological structures and the inherent ambiguity in geophysical data, comprehensively characterizing potential thrust fault systems presents a considerable challenge. Current automated fault detection methods, primarily targeting normal faults, show limited efficacy in complex fault systems of subduction zones. Treating the task of fault detection as a binary image segmentation issue, we propose a supervised end-to-end fully convolutional neural network (U-Net) to automatically and accurately delineate thrust faults from seismic data. To circumvent the labour-intensive and potentially subjective manual labelling process required for model training, we have designed a workflow to efficiently auto-generate more than 10000 training pairs comprising both 2D synthetic seismic images and their corresponding labelled images of the thrust faults simulated in the seismic images. Each synthetic seismic image includes randomly undulating stratigraphic strata and faults with dip angles between 5 and 40 degrees, aiming to simulate realistic and varied geological structures and thrust fault features in subduction zone, which equipped the U-Net model to achieve a 91% accuracy rate in fault detection within the test dataset. The example from the Hikurangi subduction zone, New Zealand demonstrates that the U-Net trained by only synthetic data is superior to conventional automatic methods, such as unsupervised methods or supervised methods trained by normal faults, in delineating more than 70% thrust faults from seismic images. To enhance the U-Net model's adaptation to specific regional fault characteristics and reduce the interference from noise, we incorporated a select set of real 2D seismic images and manually interpreted fault labels into the transfer learning process, which significantly improved its prediction accuracy and make the results clearer. From the comprehensive 2D characterizations based on the U-Net model, we can further extract 3D thrust fault systems and quantitatively measure their geometric parameters.