Depth Coordconv for Enhancing Unsupervised Domain Adaptation in Fully Convolutional Network-Based GPR Inversion

Deep neural network-based GPR inversion is often trained on simulated datasets due to the difficulty of obtaining the real-world permittivity distribution data necessary for training. Models relying solely on simulated data frequently fail to generalize effectively to real-world scenarios. Therefore, unsupervised domain adaptation (UDA) has been adopted as a promising strategy, and our previous work has made it clear that adversarial-type UDA methods may offer a potential means to bridge the gap between source and target domains. However, when both the main network and the discriminator are composed of Fully Convolutional Networks, the translational symmetry prevents them from capturing positional information within the image. As a result, features at different depths in the source and target datasets may be aligned incorrectly, potentially leading to performance degradation. This work offers the Depth CoordConv solution for this problem. Drawing inspiration from CoordConv, which augments feature channels with spatial coordinate information, our approach explicitly incorporates depth information into the input image. This design promotes the alignment of features sharing similar depth characteristics, thereby facilitating more robust and depth-aware feature extraction in convolutional neural networks. The effectiveness of this method was validated through two datasets generated by FDTD simulations: a noiseless source dataset and a noisy target dataset. As a result, the incorporation of Depth CoordConv was found to enhance both the visual quality of the estimations. When compared to a network trained exclusively on target data (Target Only), the improvements were quantified as a 20.8 % reduction in MAE and a 14.2 % improvement in SSIM.