Speaker
Description
In ground-penetrating radar (GPR) data, noise can significantly degrade the signal-to-noise ratio and resolution, thereby negatively impacting the subsequent interpretation of subsurface anomalies. This paper introduces a deep learning denoising approach utilizing the Feedforward Denoising Convolutional Neural Network (DnCNN) and incorporates the Mish activation function. Multiple-step predictions are performed to construct GPR-DnCNN, enhancing denoising performance. The GPR-DnCNN, grounded in both neural network and statistical principles, autonomously extracts features through convolutional neural networks and utilizes a single residual unit to predict noise. Specifically, inputting noisy GPR data into the GPR-DnCNN, the model learns and outputs the predicted noise, which is then subtracted from the input to obtain denoised GPR data. Validation of the denoising effectiveness is conducted using synthetic GPR data. A comparative analysis with mean filtering and F-X predictive filtering methods is performed, demonstrating that the proposed method outperforms the original DnCNN in eliminating random noise from GPR data. In comparison to other denoising methods, GPR-DnCNN proves to be more effective in suppressing random noise. Furthermore, GPR-DnCNN exhibits applicability in real data denoising, achieving notable denoising results while preserving and emphasizing essential signals.
