DETECTING LOW-ANGLE FAULTS ON SEISMIC MIGRATION IMAGES USING NESTED RESIDUAL U-NET

Detecting faults is crucial for site characterization and monitoring in geological carbon storage projects because faults can damage the integrity of natural seals that might cause CO2 leakage and induce seismicity. The robust automation for detection and segmentation of faults on seismic migration images is a rapidly developing technology and yet still poses many challenges. In particular, the lack of expertly interpreted fault volumes to use for training data requires the use of synthetic seismic images and fault labels. While the detection of high angle faults has been highly successful, low-angle normal and reverse faults are challenging to detect because of their large throws, variable angles, and rotational features. We generate synthetic seismic images as training data that incorporates high-, low-, and variable-angle fault features. We use these data to train a nested residual U-Net for fault detection and enhance its ability to detect variable dipping faults on seismic migration images. Additionally, we assess the ability of the network for varying noise levels, geological formation thickness, and seismic wavelet frequency. We use seismic migration images from Iron Mountain in Utah to verify the ability of the network to detect low-angle faults in addition to high-angle faults.