BDCNet: An effective method for extracting the borehole dispersion curves and obtaining formation P- and S-wave velocities

This study introduces Borehole Dispersion Curves Network (BDCNet), a deep learning framework built on the DeepLabV3+ architecture, designed to automate the extraction and classification of borehole dispersion curves from acoustic logging data. Addressing three critical challenges—automated dispersion curve extraction, aliased mode interference, and direct S-wave velocity determination—BDCNet integrates image segmentation with Amplitude and Phase Estimation (APES) and Prony dispersion inversion techniques to compute formation P-wave and S-wave velocities. By embedding geological prior knowledge into training through suppressing aliased modes as background labels and leveraging physical relationships such as the low-frequency cutoff velocity of pseudo-Rayleigh waves for S-wave velocity estimation, the network translates spatial dispersion features into physical parameters. Experimental results demonstrate over 99% accuracy in velocity inversion under challenging conditions including large receiver spacings and severe aliasing. Notably, BDCNet predicts P-wave velocities with 96% accuracy even when P-wave signals are weak by utilizing implicit dispersion-to-velocity mapping. This method eliminates manual steps in traditional approaches, enhancing precision and offering a robust tool for reservoir parameter analysis.