A Deep Learning Inversion Method for 3-D Electrical Resistivity Tomography Based on Neighborhood Feature Extraction
Qian Guo, Benchao Liu, Yaxu Wang, Dongdong He
Abstract
Electrical resistivity tomography (ERT) is one of the most popular methods in geological exploration. When reconstructing the 3-D resistivity model directly from the apparent resistivity data, it has two main challenges: first, the apparent resistivity data obtained from the survey lines are limited, which is much less than the true model parameters. Second, the sensitivity of the data to the model has uneven spatial distribution. In this article, a novel deep learning algorithm is proposed to reconstruct a 3-D resistivity model directly from apparent resistivity data. The new resistivity inversion deep neural network (DNN) is based on neighborhood feature extraction. By using the limited observational apparent resistivity data profiles, the neighborhood features are extracted through a fully connected network to provide the augmented data so that the spatial correspondence between the input apparent resistivity data and the output resistivity model can be enhanced. A 3-D U-Net convolutional neural network is used to learn the attribute information feature relationship spatially aligned with the resistivity model from these augmented data. After that, the 3-D resistivity model is reconstructed. It is worth to point out that, a depth distance weighting constraint is added into the loss function to balance the sensitivity distribution of the different apparent resistivity data profiles and to improve the imaging effect between apparent resistivity data profiles and areas that far away from these data profiles. Finally, the effectiveness and reliability of the newly proposed DNN are verified through numerical simulations and field tests.