<p>Groundwater contamination characterization often requires joint estimation of source parameters and heterogeneous hydraulic conductivity fields, a task commonly addressed using data assimilation techniques. However, conventional data assimilation methods typically involve iterative computations, leading to substantial computational costs, especially for high-dimensional, nonlinear problems. To overcome these limitations, this study proposes a hybrid data assimilation method that integrates a novel convolutional neural network (CNN)-based surrogate model, ARNet, with the iterative local updating ensemble smoother (ILUES) algorithm. The ARNet surrogate model was used as an alternative to computationally expensive MODFLOW-MT3DMS numerical simulations, significantly accelerating forward model evaluations. Karhunen–Loève (KL) expansion was employed to reduce the dimensionality of the hydraulic conductivity field. On this basis, the KL-ARNet-ILUES hybrid inversion framework is developed for efficient joint estimation. Comparative experiments against the KL-ILUES approach demonstrate a remarkable improvement in computational efficiency, achieving approximately 98.32% reduction in computational time. While introducing minor approximation errors, the KL-ARNet-ILUES framework maintains satisfactory inversion accuracy for both source parameters and hydraulic conductivity fields. Furthermore, this study systematically evaluates the impact of varying monitoring well configurations, representing different levels of data availability, on inversion performance. Results indicate that monitoring well density significantly influences the inversion results; denser configurations effectively reduce parameter uncertainty and enhance convergence. Notably, the influence of surrogate model approximation errors is more pronounced for specific metrics (e.g., root mean square error (RMSE) of source parameters) and under sparse observation conditions.</p>