<p>The dike of the Wenzaobang River is subjected to long-term wave erosion, leading to hidden safety hazards such as surface subsidence, voids, tilting of flood control walls, and localized cracking, which pose a threat to navigation safety. To accurately investigate the distribution of these hazards, this study, in conjunction with regional geological conditions, constructed a geological-geophysical forward model for soil loosening and void hazards within the dike body, clarifying their typical response characteristics in Ground Penetrating Radar (GPR) images: soil loosening manifests as disordered waveforms and disrupted in-phase axes, while voids present typical hyperbolic reflections. Building on this, the YOLOv11 deep learning network was introduced to automatically identify and locate hazards in the measured radar data, accurately delineating areas of soil loosening and void hazards within the dike project. Testing showed that the YOLOv11 model achieved an identification precision of 98% and a recall of 95% for hazard targets, effectively enhancing identification efficiency and accuracy. During the training process, precision gradually converged to 0.98 with increasing iterations, the multi-scale mean average precision (mAP50-95) steadily increased to 0.95, and recall progressively approached 1.0. The research demonstrates that the method combining GPR with numerical simulation and deep learning possesses good engineering applicability in dike hazard detection, effectively improving the efficiency and accuracy of geological hazard detection in complex environments. The obtained detection results clearly reveal the distribution characteristics of internal structural defects within the dike, providing crucial data support and a scientific basis for subsequent reinforcement design and safety operation and maintenance. This holds positive significance for ensuring the navigation safety of the Wenzaobang River and the stability of coastal flood control.</p>