<p>Accurate identification of stratigraphic interfaces and embedded cavities is critical for geological modeling, engineering design, and infrastructure maintenance. However, integrating geophysical and geological data remains challenging due to discrepancies in spatial resolution, signal-to-noise ratio, and geological representation. This study presents a stratigraphic U-Net model that combines geophysical and geological information to improve subsurface interpretation. An initial U-Net model is trained using a preliminary P-wave velocity model derived from conventional seismic inversion. The predicted interfaces, however, show noticeable discrepancies with borehole observations because of manual velocity picking and limited borehole data. To address these limitations, the stratigraphic U-Net model is retrained using the initial predictions together with both drilled and synthetic boreholes. This iterative strategy significantly improves the identification of stratigraphic interfaces, particularly at greater depths. The stratigraphic U-Net model is further integrated with an opening detection model to identify embedded cavities within geological layers. Model evaluation shows that geological diversity captured by borehole data contributes more to prediction accuracy than simply increasing the borehole number. A field case study demonstrates that the proposed framework accurately identifies stratigraphic interfaces and subsurface cavities, providing a robust workflow for reducing geological uncertainty and improving the spatial coverage and reliability of subsurface characterization.</p>