Details of the Abstract
| Title of paper | A Deep Learning Framework for Magnetotelluric Impedance Tensor Reconstruction |
| List of authors | Peijie Wang, Peng Han, Xiaobin Chen |
| Affiliation(s) |
Department of Earth and Space Sciences, Southern University of Science and Technology, China, wangpj@sustech.edu.cn, $Department of Earth and Space Sciences, Southern University of Science and Technology, China, hanp@sustech.edu.cn, National Institute of Natural Hazards, Ministry of Emergency Management of China, cxb@pku.edu.cn, |
| Summary | Magnetotellurics (MT) is a geophysical method that infers subsurface conductivity distributions by measuring natural variations in the Earth's electromagnetic field. Therefore, obtaining high-quality impedance tensors is crucial for accurately resolving subsurface structures. This paper introduces a method for reconstructing impedance tensors based on deep residual neural networks (ResNet). The network leverages multi-layer convolutions and residual connections to harness the continuity of the impedance tensor and the interrelationships between different tensor elements, extracting valuable information from noisy impedance tensors to recover the original impedance tensor. Noise-free samples are generated via 3D forward modeling of both actual models and manually designed models. The training set is then augmented by rotating the impedance tensors and finally adding noise. After testing various approaches, logarithmic apparent resistivity and unwrapped phase were chosen as the optimal data combination for training. Additionally, a specialized ResNet model is trained for phase unwrapping. The test results indicate that this method can accurately reconstruct noisy impedance tensors. Furthermore, we applied the reconstructed impedance tensors to enhance M-estimation data processing. In each iteration of the M-estimation process, the reconstructed impedance tensor replaces the impedance tensor from the previous iteration, reducing the impact of outliers introduced by strong noise. The experimental results show that the ResNet-based reconstruction method significantly improves data processing performance, especially when dealing with complex geological structures and high-noise environments. This study highlights the profound potential and broad application prospects of deep learning in magnetotelluric data processing. |
| Session Keyword | 1.0 Instrumentation, data acquisition and processing |
| File upload |
1.0_a_deep_learning_framework_wang_03.pdf
|