In order to improve the accuracy of flood forecasting, the paper proposes a retroactive correction method for the whole intermediate state variables of a sub-basin based on the ensemble Kalman filter. According to the different geomorphic feature and confluence time of each sub-basin, its corresponding retroactive time is found, and then, the Xinanjiang model is combined with to correct the state of each sub-basin before a certain time, gradually reducing the accumulation of errors. The results of the ideal model show that the intermediate state of sub-basin is effectively corrected, relative error of peak and volume decreases, and the deterministic coefficient increases. 12 historical floods of the Dapoling Basin, selected as an example, are effectively modified by the proposed method. As a result, the method can effectively improve the accuracy of flood forecasting and can be widely used in actual flood forecasting.