Enhancing Deep Learning-Based GPR Data Inversion With Unsupervised Domain Adaptation: Comparison of Domain Classifiers

Deep neural network (DNN)-based inversion of ground penetrating radar (GPR) data is gaining significant attention, primarily using simulated data for training due to the lack of ground truth in real-world scenarios. However, models trained on simulation data often perform poorly on real-world data due to domain discrepancies. This study applies an unsupervised domain adaptation (UDA) method to GPR inversion, introducing a domain classifier within the main inversion DNN. By training the main DNN adversarially against the classifier to learn domain-invariant features, the DNN’s performance can be improved even without labeled data in the target dataset. Several simulation datasets with varying characteristics were generated to compare the performance of three domain classifiers. The experiments revealed that the proposed simple domain classifier structure, using $3 \times 3$ convolutional local alignment, outperformed both pointwise local alignment and global alignment (GA) classifiers. The proposed method achieved more than 80% improvement in metrics [mean absolute error (MAE), structural similarity (SSIM), and mean intersection over union (mIoU)] over the non-UDA case, highlighting its effectiveness in addressing domain gaps related to frequency and attenuation. Furthermore, the DNN was trained using unlabeled data from real-world roads measured by on-vehicle GPR, and then applied to a bridge deck slab specimen with known design drawings. From both qualitative and quantitative perspectives, the proposed method demonstrates superior performance in permittivity and material estimation. The method enabled clear detection of rebars, and damage such as deterioration and cracks in the concrete slab. This research is expected to provide new insights into the efficient subsurface damage detection of road infrastructures.