2-D Magnetotelluric Inversion Based on Improved DenseUNet

To improve the imaging accuracy and inversion stability of magnetotelluric data in complex geological settings, an enhanced DenseUNet-based architecture is developed in this study. The approach integrates Deformable Convolution v4 to enable adaptive extraction of multi-scale spatial features. An improved Atrous Spatial Pyramid Pooling module is inserted between the encoder and decoder to reinforce cross-level feature correlation and contextual information fusion. A channel attention mechanism is further incorporated to strengthen the representation of key features while mitigating the influence of noise.To comprehensively evaluate the performance of the proposed architecture, a dataset of 200,000 samples was generated through two-dimensional finite-element forward modeling. Comparative experiments were conducted against U-Net, DenseUNet and Res-FromerNet under varying noise levels and input configurations. The results demonstrate that the improved architecture achieves superior convergence speed, inversion accuracy, detail reconstruction capability, and overall stability. It maintains consistent imaging quality under Gaussian noise levels from 0% to 10%, and the joint input configuration produces the lowest inversion errors.Field data collected in Huaining, Anhui Province, were further used to validate practical applicability. The reconstructed resistivity structures show strong agreement with known geological features, exhibiting clear interfaces, continuous layering, and significantly improved spatial resolution. Overall, the enhanced architecture markedly improves the reliability and adaptability of magnetotelluric inversion and shows promising application prospects for complex geological environments.