Hybrid Vision Transformer With Convolutional Blocks Approach for Subsurface Electrical Resistivity Tomography Inversion

Electrical Resistivity Tomography (ERT) is widely used for subsurface imaging. Despite recent advancements in machine learning based ERT inversion using conventional deep learning models such as CNN and UNet, accurately delineating resistivity profiles of natural and artificial subsurface anomalies remains challenging. These challenges stem from multi‐scale complex geological heterogeneity, spatial resolution limitations, and the ill‐posed nature of ERT inversion. This study proposes a new hybrid image‐based ERT inversion method leveraging a Vision Transformer (ViT) model incorporating convolutional blocks to map observed pseudo sections of subsurface apparent resistivity to true resistivity distributions. This novel approach is compared to others including CNN AutoEncoder, UNet, and Latent Diffusion. A synthetic data set generated using forward modeling with varying conductive/resistive heterogeneity zone, facilitates model training using a curriculum strategy to improve generalization across complexity levels. To address the absence of true resistivity data for field observations, a residual CNN model is employed to calibrate resistivity colormap ranges of ERT imaging to ensure consistency in resistivity values represented in both input and output images of the ViT model, maximizing inversion accuracy and result interpretation in applications. The proposed method demonstrated superior performance over frequently used comparative models, as the adopted ViT model excels in inversion accuracy, computational efficiency, adaptability to heterogeneity, and robustness to measurement noise. Inversion results showed significant advantages over traditional Gauss‐Newton numerical inversion, which achieved high fidelity in rapidly resolving subsurface anomalies with sharp boundaries and accurately mapped resistivity distributions with high resolution, presenting a promising data‐driven approach for large‐scale ERT imaging applications.