A Novel Transformer–CBAM Module Network Toward Practically Viable Applications of Ground Penetrating Radar Inversion

Ground penetrating radar (GPR) is a nondestructive geophysical tool that emits electromagnetic (EM) waves and captures their echoes to image subsurface structures. However, criticism of conventional imaging persists, stemming from quantitative ambiguity or limited practicability, as deterministic or probabilistic inversion typically suffers from high computational costs, accuracy variability, and initial model sensitivity. To address these issues, this article develops a deep learning (DL) inversion alternative based on the RCUNet-Transformer network. Specifically, we exploit a lightweight variant of the RFDB-U-Net version as the backbone, with a special Transformer encoder module designed to underscore global context feature extraction, and a convolutional block attention module (CBAM) leveraged to prioritize the crucial shallow zones while minimizing attention to irrelevant ones. The network performance is evaluated extensively through the use of synthetic records with diverse conditions, laboratory experiments, and on-site complex field datasets. Our evaluation of reconstruction quality involves utilizing three auxiliary metrics to assess global consistency, scale-invariant features, and edge-level detail preservation, as well as two basic indicators to assess pixel-wise fidelity and structural similarity. The results demonstrate the outperformance of the proposed method over other networks, especially for field dataset elaboration, which showcases its generalization capability in terms of practicability and quantitative accuracy for the potential of practical engineering applications.