This paper introduces a novel data assimilation (DA) method called DA$_{(DL)}$ that leverages deep learning (DL) to address the limitations of traditional DA methods, which often suffer from low computational efficiency or are constrained by the Gaussian assumption. DA$_{(DL)}$ is designed to handle non-linear, high-dimensional, and non-Gaussian problems in hydrological systems. The method generates a large volume of training data from the prior ensemble and trains a DL model to update system knowledge using multiple predictors, including model parameters, model outputs, and the innovation vector. The iterative form of DA$_{(DL)}$ is proposed for highly non-linear models, and data augmentation and local updating strategies are introduced to enhance its performance in small ensemble size and equifinality issues. Two case studies, one involving non-Gaussian distributions and the other Gaussian distributions, demonstrate the effectiveness of DA$_{(DL)}$ compared to traditional DA methods like ES$_{(K)}$ and ES$_{(DL)}$. The results show that DA$_{(DL)}$ provides more accurate and reliable estimates, particularly in non-Gaussian scenarios. The paper concludes by discussing the potential improvements and future directions for DA$_{(DL)}$.This paper introduces a novel data assimilation (DA) method called DA$_{(DL)}$ that leverages deep learning (DL) to address the limitations of traditional DA methods, which often suffer from low computational efficiency or are constrained by the Gaussian assumption. DA$_{(DL)}$ is designed to handle non-linear, high-dimensional, and non-Gaussian problems in hydrological systems. The method generates a large volume of training data from the prior ensemble and trains a DL model to update system knowledge using multiple predictors, including model parameters, model outputs, and the innovation vector. The iterative form of DA$_{(DL)}$ is proposed for highly non-linear models, and data augmentation and local updating strategies are introduced to enhance its performance in small ensemble size and equifinality issues. Two case studies, one involving non-Gaussian distributions and the other Gaussian distributions, demonstrate the effectiveness of DA$_{(DL)}$ compared to traditional DA methods like ES$_{(K)}$ and ES$_{(DL)}$. The results show that DA$_{(DL)}$ provides more accurate and reliable estimates, particularly in non-Gaussian scenarios. The paper concludes by discussing the potential improvements and future directions for DA$_{(DL)}$.