The performance of real-time flood forecasting can be improved by updating with the real-time observations. The performance of filtering is determined by the state variables and therefore the criteria for choosing state variables is proposed. With this criteria, a real-time updating method of XAJ model is proposed by using the Unscented Kalman Filter and the conceptual XAJ model. The effectiveness of the new method is supported by a real case study where the filter is applied to flood forecasting in Shaowu Basin, Min River. The results shows that the method using UKF can remarkably update the state variables and improve the accuracy of flood forecasting. It is practical and can be applied to real flood forecasting tasks.