Porosity prediction from prestack seismic data via deep learning: incorporating a low-frequency porosity model

Abstract Porosity prediction from seismic data is of considerable importance in reservoir quality assessment, geological model building, and flow unit delineation. Deep learning approaches have demonstrated great potential in reservoir characterization due to their strong feature extraction and nonlinear relationship mapping abilities. However, the reliability of porosity prediction is often compromised by the lack of low-frequency information in bandlimited seismic data. To address this issue, we propose incorporating a low-frequency porosity model based on geostatistical methodology, into the supervised convolutional neural network to predict porosity from prestack seismic angle gather and seismic inversion results. Our study demonstrates that the inclusion of the low-frequency porosity model significantly improves the reliability of porosity predictions in a heterogeneous carbonate reservoir. The low-frequency information can be compensated to enhance the network's capabilities of capturing the background porosity trend. Additionally, the blind well tests validate that considering the low-frequency constraint leads to stronger model prediction and generalization abilities, with the root mean square error of the two blind wells reduced by up to 34%. The incorporation of the low-frequency reservoir model in network training also remarkably enhances the geological continuity of seismic porosity prediction, providing more geologically reasonable results for reservoir characterization.