In electrical resistivity tomography (ERT), the anomaly effects of different electrode arrays vary depending on the geological model. The appropriate combination of different electrode arrays can optimize detection performance and enhance the reliability of interpretation results. However, traditional inversion methods, constrained by single-array data, sparse observations, and ill-posed problem-solving, often yield low-resolution or inaccurate results. To address the resolution challenges in ERT inversion, inspired by the outstanding fusion and nonlinear mapping capabilities of multi-modal deep learning (DL) image methods, we propose the super-resolution ERT fusion network (SRERTF-Net), which utilizes traditional inversion results of multi-array as the initial models, efficiently leveraging and integrating prior physical information to achieve multi-array data joint super-resolution inversion. In SRERTF-Net, different down-sampling paths are employed to process the inversion results of various electrode arrays, while Inception modules are introduced to enhance feature extraction. Additionally, dense connections are implemented both within and across paths to effectively integrate complementary information from different arrays, ensuring robust multi-modal feature fusion. Finally, we designed training samples that include randomly generated typical structural models and comprehensive complex models, in order to enhance the practicality and adaptability of the network. Experiments on synthetic and field measured data indicate that SRERTF-Net outperforms other methods in terms of resistivity accuracy, resolution, and background performance.

Denoising is a crucial step in ground penetrating radar (GPR) data processing. Conventional denoising algorithms for GPR typically require selecting optimal processing parameters, which can be challenging to achieve in practical applications, resulting in unsatisfactory processing outcomes. In recent years, in order to address the issue of low accuracy in conventional GPR denoising algorithms, denoising neural networks have been applied in the field of GPR. Although conventional denoising neural networks have shown improvements in signal-to-noise ratio (SNR) in some cases, their performance is often inadequate when facing real GPR data with complex random noise, due to the training methods of the networks. To address the challenges in denoising of GPR data, a two-stage denoising method based on deep learning (DL) has been proposed. Initially, conventional GPR data processing is conducted, followed by training a denoising network model using both the processed and unprocessed signals. Leveraging the powerful nonlinear fitting capability of convolutional neural networks (CNNs), an end-to-end mapping relationship is established to obtain the final denoising network model, completing the two-stage denoising process. Finally, this letter validates the proposed two-stage denoising method using synthetic and field data. The radar data obtained through this two-stage denoising method not only improve mean squared error (mse) by 0.17 compared to conventional methods but also increase peak SNR (PSNR) by 8.1. Furthermore, there is a significant enhancement in the integrity of the waveform and the recovery of weak signals.

Traditional cross-hole radar tomography (CRT) usually cannot obtain high-resolution imaging results due to the nonlinearity and multisolution of inversion. To cope with these challenges, we propose a scheme to achieve high-resolution CRT for complex slowness models using deep neural networks (DNNs). Given the inherent difficulty in generating complex geophysical models in batches, by series of processing some remote sensing images from the remote sensing scene classification dataset, we create a real slowness model dataset. Then, we utilize 2-D U-Net to directly construct the mapping relationship between the low-resolution slowness model from the traditional method and the real slowness model. The superiority of our scheme is verified by both synthetic data and measured data. Our scheme can significantly suppress the false anomaly of traditional CRT results and accurately reconstruct the underground target’s geometry, position, and slowness value, and it has excellent accuracy and robustness. In addition, the response data of the slowness model reconstructed by our scheme are closer to the field data.

Noise attenuation is crucial in ground-penetrating radar (GPR) data processing. In recent years, deep learning (DL) methods have shown excellent performance in GPR denoising tasks, but they typically focus only on recovering the target signal, which can lead to over-denoising. To enhance the generalizability and the practicality of denoising networks, we propose a strategy to generate random dielectric models from natural image datasets, which can quickly construct model datasets with low redundancy and reasonable distribution. To enhance the fidelity of GPR denoising, we leverage the powerful nonlinear fitting capabilities of convolutional neural networks (CNNs) and introduce a closed-loop denoising network framework for GPR. The framework consists of a denoising sub-network and a noise extraction sub-network, effectively achieving signal-noise separation in noised GPR data. Specifically, the denoising sub-network is used to recover weak reflection signals and initially remove noise, while the noise extraction sub-network is used to restore the true noise, mitigating the problem of over-denoising. A key innovation of our approach is the integration of bandpass filtering, which enhances the robustness of network training and supports effective weak signal recovery. This network framework forms a closed loop through the residual loss between the signal-noise separation results and the noised GPR data, the closed-loop structure is capable of further refining the signal and noise prediction results of the two subnetworks, thereby enhancing the numerical accuracy of the signal-to-noise separation results. Finally, the effectiveness of the GPR closed-loop denoising network is verified from multiple perspectives using both synthetic and field measured data. The results indicate that our proposed method is more competitive in GPR denoising tasks.

The traditional electrical resistivity tomography (ERT) inversion methods typically produce low-resolution imaging results due to the nonlinear and bulk effect of two-dimensional inversion. In this paper, we propose to directly establish the mapping from the geoelectric model of traditional inversion results (input) to the actual geoelectric models (output) through the fully convolutional networks (FCNs), inspired by the robust nonlinear mapping capabilities of deep learning methods. We designed an ERT resolution enhancement network (ERTReNet) based on the prevailing U-Net architecture, which can conduct end-to-end training and enhance the resolution of traditional inversion imaging results. This methodology has been tested on both synthetic and field measured data. Resolution has been improved, and the resistivity value of both target and geological background are closer to the synthetic model comparing to the tradition method. This work aids in improving the accuracy of subsurface target identification in ERT and serves as a guide for more precise ERT inversion in the future.