To address the challenge of detecting the complex geoelectric structures of karstic bauxite deposits, this paper proposes a deep learning inversion method for the opposing coils transient electromagnetic method (OCTEM) integrated with geological model constraints, aiming to improve the inversion accuracy of subsurface resistivity structures. First, based on the karst metallogenic characteristics of these deposits, a geoelectric model with a low- high- low resistivity structure is constructed. A training dataset of resistivity models and their corresponding OCTEM responses is generated using variation coefficients of fixed horizon thicknesses. Second, the UResNet architecture is built by integrating a U- Net as the backbone network with a residual network (ResNet). This architecture enables multi- scale feature extraction through residual modules and establishes a nonlinear mapping relationship between the electromagnetic response data and subsurface electrical structures. Third, the proposed method is verified using synthetic data (both noise- free and noisy) and field- measured data from a karstic bauxite deposit in Yangcheng County, Shanxi Province. The research results show that the method exhibits minimal error in anomaly identification and high inversion accuracy, with the field- measured results showing a strong correlation with known geological data. This deep learning inversion method based on OCTEM can break through the dependence of traditional linear methods on initial models, providing an effective solution for detecting subsurface structures under complex geological conditions.