AUTL: An Attention U-Net Transfer Learning Inversion Framework for Magnetotelluric Data

Given the limited number of labeled magnetotelluric (MT) filed data samples, current neural network (NN) inversions for MT are primarily rely on synthetic data, which may not fully capture the complexity of true underground resistivity structures. This letter introduces a novel inversion scheme that combines attention U-Net with transfer learning (AUTL) to bridge this gap. The proposed method improves inversion accuracy by integrating field data into the training process through transfer learning (TL), using 3-D models derived from real field measurements as the target dataset. This enhances the authenticity and reliability of the inversion results. Additionally, the incorporation of attention gates (AGs) significantly improves feature extraction by focusing on relevant features. We validate the effectiveness of the AUTL approach using both synthetic and measured data, demonstrating its superior performance in reconstructing underground resistivity structures with high accuracy and noise resistance. This method offers a promising solution for training inversion networks with limited field MT data and lays the groundwork for expanding datasets with more diverse 3-D models in the future.