Deep learning inversion of electrical resistivity tomography using an enhanced U-Net for mapping discontinuous permafrost

Electrical resistivity tomography (ERT) is widely used for permafrost investigations. However, conventional inversion algorithms often exhibit limited capacity to resolve the precise geometry and boundaries of permafrost, particularly in the presence of high-contrast stratigraphic interfaces or discontinuous permafrost bodies. These limitations reduce their effectiveness in engineering and environmental applications. In this study, we develop an enhanced U-Net convolutional neural network (CNN) inversion framework designed to improve the accuracy, sharpness, and reliability of permafrost boundary delineation. The framework was developed using finite-element forward simulations of both layered and discontinuous permafrost. Evaluating against conventional smoothness-constrained least-squares methods across three representative synthetic models, the enhanced U-Net demonstrated superior performance in boundary reconstruction, structural continuity, and anomaly definition. The model's practical utility was further validated using two field datasets from the Qinghai-Tibet Plateau (QTP). Benchmarked against borehole observations, the enhanced U-Net recovered the main frozen-ground structures with improved geometric clarity relative to the conventional inversion results, highlighting its potential for geotechnical and environmental applications in cold regions.