This study applies explainable deep learning (XDL) to uncover dynamically significant coherent structures in the turbulent boundary layer on the suction side of a NACA 4412 wing section subjected to strong adverse pressure gradients (APG). A short-horizon U-Net surrogate is trained using high-resolution large-eddy simulation (LES) data, and gradient-SHAP feature attribution is utilized to quantify the local contribution of flow features to the short-time evolution of velocity fluctuations. Unlike classical kinematic detectors, this approach isolates regions based strictly on their importance for reconstructing future events. Comparisons with Q events, streaks and vortex clusters indicate that SHAP-identified structures provide a more complete description of the turbulent flow than traditional definitions: they correspond primarily to ejection events near the wall but evolve into broad, complex outer-layer structures under strong APG, encompassing multiple classical features. In contrast, vortex clusters are shown to possess marginal dynamical importance. These findings demonstrate that gradient SHAP provides a physically interpretable, data-driven perspective on non-equilibrium turbulence, offering a robust alternative to standard structure-identification methods for aerodynamic applications.