Post-stack seismic impedance inversion based on sparse-coded Mamba seismic model

Abstract

Post-stack seismic impedance inversion plays a vital role in reservoir characterization and seismic attribute analysis, enabling the interpretation of lithological properties and the prediction of shale distribution in coal seam roofs and floors—critical for mine safety and working face design. However, conventional inversion methods often suffer from low resolution, strong non-uniqueness, and a heavy reliance on low-frequency initial models, limiting their generalization and practical reliability. To overcome these limitations, this study introduces an enhanced seismic inversion network that integrates a Mamba-based encoder with a masked autoencoder (MAE) to effectively capture global dependencies and structural features in seismic data. Furthermore, a physics-guided learning framework is proposed by incorporating a forward operator and well-log constraints, enhancing the physical consistency and stability of the inversion process. Validation on the Marmousi2 model and field datasets demonstrates that the proposed method significantly outperforms conventional approaches, increasing the Pearson correlation coefficient(PCC) from 0.90 to 0.95 at blind well locations and reducing the root mean square error (RMSE) by 0.11 compared to Constrain Sparse Spike Inversion (CSSI). The results confirm that our model achieves higher spatial continuity and geological plausibility with improved computational efficiency. This work provides a robust deep learning solution for high-resolution impedance inversion, offering substantial practical value for enhanced reservoir characterization and reduced exploration risk.