Abstract: Initial error is one of the key factors that have great effects on the accuracy of numerical forecast. To study the characteristics of initial error and its influence on the numerical prediction, an analysis procedure of the sensitive initial error of numerical forecast is developed based on WRF adjoint model and is used in the investigation of typhoon Kammuri (0809). The validity of the linear assumption on the study of typhoon case is firstly assessed prior to discussing any adjoint analysis results. It is done by evaluating the evolution differences of the perturbation between linear and nonlinear development, showing that the nonlinear perturbation evolution is well represented by the linear assumption during 24-h forecast. The sensitive initial error is then constructed using the information derived from adjoint sensitivity analysis, finding that the reference coefficient from 0.01 to 0.08 is proper to build the sensitive initial error. The result of 0.08 is the best in this case study. The numerical forecast error could be reduced and the prediction bias of typhoon trace could be improved greatly by removing the sensitive initial error from the initial field. This effect of the sensitive initial error derived from 24-h numerical forecast error affects the numerical forecast even within 60 hours. In addition, the analysis reveals that the sensitive initial error of regional short-term numerical forecast concentrates mainly around the weather system. It goes with typhoon circle and the pattern is almost consistent for all physical variables. The sensitive initial error in the middle-upper troposphere has slightly more contribution to the forecast than that in lower troposphere. Comparing the contribution of different physical variable, it is found that wind is the main contributor with pressure and humidity following.