Deep Prior Information-Constrained Full-Waveform Inversion for Ground-Penetrating Radar

This study proposes an improved full-waveform inversion (FWI) method based on deep learning to address the challenges of high accuracy requirements for initial models, high computational complexity, and low efficiency in traditional ground-penetrating radar (GPR) inversion. Traditional FWI methods rely on an accurate initial model and iteratively update the model through multiple forward processes, which are computationally expensive and time-consuming. To overcome these issues, we introduce a deep learning-driven inversion framework that rapidly generates an accurate initial model, which is then used as a plugin for the FWI process. The deep learning model replaces traditional FWI model corrections and forward processes, with the predicted initial model serving as the foundation for FWI updates, accelerating the inversion and enhancing both robustness and interpretability. In model design, we employ a Unet++ network architecture with depthwise separable convolutions to improve computational efficiency. In the design of the loss function, we combine Multi-Scale Structural Similarity (MS-SSIM) loss with Laplacian pyramid loss to ensure that the inversion results maintain structural similarity while recovering high-frequency details of the image. We validate the proposed method through experiments using synthetic data, comparing results with traditional FWI methods. The inversion outcomes are evaluated using metrics like structural similarity (SSIM), peak signal-to-noise ratio (PSNR), MSE, and inversion time. Results show that the proposed method significantly improves both accuracy and efficiency.