SCB-ADAE: An Attention-based Deep Autoencoder for Ground Penetrating Radar Signal Denoising

Ground Penetrating Radar (GPR) is a widely used geophysical tool for subsurface investigation, including applications in civil engineering, environmental studies, and archaeological explorations. However, GPR signals are often contaminated by various types of noise. These noise factors can significantly degrade the quality of the GPR signal. Existing denoising techniques often struggle to remove complex, non-Gaussian noise or site-specific interference effectively. To address this issue, this study proposes a novel denoising model, the Swin-Conv Block with Attention Denoising Autoencoder (SCB-ADAE), which integrates convolutional and self-attention mechanisms to enhance GPR signal denoising performance. The SCB-ADAE model consists of two key components: the Swin-Conv Block (SCB) and the Attention Denoising Autoencoder (ADAE). The SCB captures high-level features of the raw GPR signal, preserving important details while extracting local and global features. The ADAE module, enhanced with self-attention, focuses on the most relevant components of the signal, suppressing noise and preserving the core features that are essential for accurate interpretation. The process begins by passing the raw GPR signal through the SCB for feature extraction. Next, the ADAE module denoises the extracted features by utilizing self-attention mechanisms. Finally, the denoised signal is passed through a second SCB module for refinement and dimension matching with the original input signal. The model was tested on radar signals contaminated by Gaussian noise at varying levels (5 dB, 7.5 dB, and 10 dB), inhomogeneous-material noise, and real-world GPR signals, with performance evaluated using key metrics such as Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). The SCB-ADAE model consistently outperformed existing state-of-the-art models like U-Net and Denoising Autoencoders. For example, at a noise level of 5 dB, SCB-ADAE achieved an SNR of 31.33 dB, PSNR of 38.59 dB, and SSIM of 0.9817, significantly surpassing SCUNet, which achieved lower scores. As the noise level increased, SCB-ADAE maintained superior performance, demonstrating its ability to handle higher levels of noise effectively. In tests involving radar signals with inhomogeneous-material noise, SCB-ADAE demonstrated a 146.74% improvement in SNR and a 16.65% improvement in PSNR compared to SCUNet, highlighting its capacity to address complex, site-specific noise types. In conclusion, the SCB-ADAE model is an effective solution for denoising GPR signals in noisy environments. Future work should focus on expanding training datasets to include more diverse noise types and exploring transfer learning techniques to improve model generalization across different geological environments.