Enhancing Early CO2 Leakage Detection using Pre-Trained Deep Learning Models

This study presents a deep learning framework for early CO2 leakage detection from time-lapse seismic data, leveraging a pre-trained ResNet152 model originally trained for natural images classification. The framework integrates forward seismic modeling, deep feature extraction, and leakage mass prediction. ResNet152 is adapted to extract embeddings from time-lapse seismic shot gathers, which are compared using a Siamese network setup. Time-lapse embeddings are projected into a 2D latent space using UMAP, enabling anomaly detection and temporal pattern analysis. A multilayer perceptron (MLP) regressor is then trained to estimate leakage mass from the extracted embeddings. Tests on synthetic datasets from the Kimberlina CO2 storage site simulations show that the model effectively captures the temporal evolution of seismic signal related to CO2 leakage, particularly during early stages. UMAP projections reveal clear separability between baseline, no-leakage, and leakage scenarios. Despite being trained on non-seismic data, the model produces informative embeddings that support leakage detection. The MLP regressor achieves satisfactory performance under realistic noise levels. This framework demonstrates the potential of using pre-trained CNNs for seismic monitoring tasks and is extendable to other architectures. It offers a computationally efficient, generalizable approach that can serve as an early-warning to flag anomalies and guide targeted surveys.