Automated CO2 Plume Detection from Time-Lapse Seismic Using Local Orthogonalization and Deep Learning

Reliable interpretation of time-lapse (4D) seismic data is essential for monitoring injected CO and ensuring safe long-term storage, yet observed differences between vintages are often dominated by acquisition non-repeatability, near-surface variability, and imaging artifacts. We present an automated workflow that combines Local Orthogonalization Weighting (LOW), a physics-guided method that isolates dynamic fluid-induced changes, with a compact 3D deep learning model (TLNet) that predicts CO plume probability from paired baseline–monitor volumes. LOW suppresses repeatable geology and stabilizes the dynamic anomaly, enabling simple value and texture gates to produce volumetric plume labels without manual interpretation. TLNet learns from these seismic-driven labels, generating high-resolution plume probability volumes that capture thin-layer confinement, lateral variability, and intra-tier fingering. The workflow is demonstrated on the Sleipner CO storage project, and the LOW component is further evaluated on the Cranfield and Duri fields to assess robustness across marine and land surveys, shallow and deep targets, and CO and thermal-EOR settings.