Automated interpretation of ground penetrating radar (GPR) images is essential for efficient underground cavity detection and segmentation. However, existing methods often exhibit limited performance in real-world applications due to the high variability in cavity appearances, which poses significant challenges to the generalization and reliability of deep learning models. To address this issue, this study proposes an uncertainty-aware framework using ensemble learning for underground cavity detection and segmentation in real-world 3D GPR images. The proposed framework ensembles multiple segmentation models trained with different random initializations to capture variations in model predictions and improve robustness against diverse cavity appearances. By integrating predictions from these models, the framework enhances segmentation performance while enabling the quantification of predictive uncertainty through inter-model disagreement. An uncertainty-based filtering strategy is further introduced to identify and filter out unreliable cavity detections, thereby reducing false positives and improving detection reliability. The proposed framework is validated using 3D GPR images collected from real road networks, demonstrating its effectiveness in improving cavity detection and segmentation under realistic inspection conditions. The results highlight the benefits of combining ensemble learning with uncertainty quantification to address the challenges posed by highly variable cavity appearances. This study provides an uncertainty-aware approach to automated GPR interpretation, contributing to more robust and reliable underground cavity detection and supporting the practical application of GPR technology in subsurface infrastructure inspection.