A new method called DL-EnKF combines deep learning with the Ensemble Kalman Filter to improve accuracy in data assimilation for complex systems. By embedding deep neural networks in the EnKF, the DL-EnKF outperforms the traditional EnKF in accurately predicting nonlinear behaviors with just a small ensemble. Experiments show that the DL-EnKF is more accurate in strongly nonlinear situations and provides better results than deep learning alone due to positive feedback during assimilation cycles. The DL-EnKF can even improve outcomes when training on EnKF analyses with large ensembles or imperfect models.