<p>The characterization of CO₂ plume migration during geological sequestration is a challenging inverse problem due to inherent non-uniqueness and uncertainties. Reliable assessment of plume behavior is critical for storage security but is complicated by reservoir heterogeneity and the limited availability of monitoring data. This study establishes a cost-effective and robust protocol for CO₂ plume characterization by analyzing pressure signals from injection and monitoring wells. The proposed workflow couples deep neural network models with the Ensemble Kalman Filter (EnKF). An inverse model adapted from an autoencoder-based architecture is trained to predict the spatial heterogeneities in petrophysical properties using field pressure data. Two forward-looking models are employed to simulate pressure responses and the spatiotemporal evolution of free CO₂ Plumes. The forward-looking models and EnKF models are integrated to structure an i-EnKF approach and iteratively refine predictions from the inverse model. The i-EnKF approach aligns the prediction from the inverse model with the field observation data and reduces uncertainty via the plume characterization. The methodology was evaluated using both an idealized case and a real aquifer field case. With over 6,000 training samples for the Sleipner dataset and 2,000 for the synthetic case, the trained expert system demonstrated strong predictive capability in image-to-image regression tasks. Results indicate that the approach significantly reduces inversion uncertainty and reliably forecasts the spatiotemporal dynamics of CO₂ species. After injection ceases, plume evolution becomes less sensitive to injection-well pressures, which highlights the importance of monitoring-well placement for long-term storage assessment. Compared with conventional inversion techniques, the workflow requires fewer and more affordable data inputs, providing a cost-effective and accurate solution for CO₂ plume characterization in saline aquifers.</p>