Cross-domain feature-enhanced YOLOv8 model for underground defect detection

Ground penetrating radar (GPR) is widely used to detect underground defects and prevent collapses. However, environmental noise and interference often lead to misidentification and missed detection. This study proposes a cross-domain feature-enhanced YOLOv8 model for underground defect detection. A comprehensive dataset is established, consisting of GPR data from highway field tests, laboratory experiments, and numerical simulations. Both the field and experimental data are categorized as measured data, which capture real-world conditions with significant noise. In contrast, the simulated data are noise-free and provide clear target reflections. Based on the characteristics of the measured and simulated data, a CycleGAN-based cross-domain transformation method is introduced to transform the measured data into simulation-style data, aiming to reduce noise and enhance defect-related features such as cavity reflections and boundaries. To enhance global feature extraction and better detect blurred targets, an efficient vision transformer module is integrated into the YOLOv8 backbone. An engineering case in Zhejiang, China, is used for illustration. Results show that the proposed YOLOv8 model outperforms alternative underground target detection methods in both accuracy and efficiency. Transforming measured data into simulation-style data enhances the detection accuracy of the YOLOv8 model, as evidenced by a notable improvement in mAP50 from 0.61 to 0.87. Compared to the model trained on measured data, the CycleGAN-based model achieves higher initial performance, reaching an mAP50 of 0.57 with 300 samples and 0.86 with 1500 samples. These results demonstrate that the proposed model is well-suited for underground defect detection tasks in complex environments with scarce data.