Automated multi-class classification of amplitude anomaly due to carbon capture and storage using transformer models for time-lapse seismic data: A case study using Sleipner data

Accurate monitoring of CO 2 plume evolution is critical for the safety and efficiency of Carbon Capture and Storage (CCS) projects. However, conventional seismic interpretation and standard Convolutional Neural Networks (CNNs) often struggle to resolve complex, overlapping plume layers, particularly when amplitude contrast is low. This study establishes a novel Transformer-based AI workflow using U-SegFormer, a Transformer-based architecture with U-Net-like skip connections, to perform multi-class segmentation of CO 2 plumes in the Sleipner 4D seismic dataset. Unlike binary (plume/non-plume) approaches, our multi-class strategy simultaneously delineates nine distinct horizons, enabling layer-by-layer analysis of subsurface dynamics. Comparative benchmarks demonstrated that U-SegFormer, combined with Focal Loss, significantly outperformed standard CNN and SegFormer models. While the highest accuracy was achieved in the high-amplitude shallow layers, the proposed model also demonstrated improved sensitivity to deep, low-amplitude horizons compared to conventional methods. The model also successfully identified vertical migration paths (chimneys), offering enhanced structural interpretation. Temporal generalization tests revealed that the model trained on 2006 data maintained high prediction accuracy for the 2008 survey (a two-year interval). However, an extended four-year prediction (2010) showed performance degradation in deeper layers due to progressive amplitude attenuation, while shallow layers remained robust. These findings suggest that deep learning-based prediction accuracy can serve as a data-driven indicator for optimizing the frequency of 4D seismic acquisition. This study demonstrates the high potential of Transformer-based models to enhance quantitative monitoring and adaptive management of CCS sites in geological storage sites.