Urban rivers often experience significant water quality degradation due to the pollution from wet-weather discharges. Accurate pollution source identification (PSI) is essential for effective river management. However, traditional PSI methods face challenges, including high computational demands and difficulties in addressing equifinality. To overcome these issues, this study introduces an innovative approach that integrates deep learning (DL) with data assimilation (DA). Three DL models-simple Convolutional Neural Networks, ResNet, and UNet-were evaluated as surrogate models for the computationally expensive river water quality model (RWQM). Additionally, three advanced DA methods - DREAM_(ZS), ESMDA, and ILUES - were applied to estimate high-dimensional RWQM parameters. In a numerical case study of a river segment involving 50 unknown parameters across five pollution sources, we assessed the performance of eight approaches for PSI and examined the impacts of monitoring schemes and observation errors. Results showed that UNet provided the highest accuracy in surrogate modeling, while ILUES delivered the best DA performance. The combined UNet-ILUES approach demonstrated a remarkable improvement in computational efficiency, achieving a 406-fold gain compared to the RWQM-ILUES approach. Validated through a real-world water quality degradation event in the Outer Qinhuai River, the UNet-ILUES approach demonstrates strong potential as an efficient solution for characterizing the dynamics of pollution from WWDs in urban rivers, leveraging the combined strengths of DL and DA.