A method of crack identification based on a Kriging surrogate model is presented in this paper. The initial samples are used to construct the initial Kriging model which characterizes the relationship between the crack parameters and their corresponding structural dynamic responses, instead of the dynamical input and output relation for avoiding re-meshing process at every iterative step of optimization and the time-consuming FE calculation. To improve the accuracy of the surrogate model, an optimal point-adding process is carried out to reduce computational cost. For identifying crack parameters based on the constructed Kriging model, a robust stochastic particle swarm optimization (SPSO) algorithm is adopted for enhancing the global searching ability. Numerical studies of cantilever beam and plate with a crack are performed. The effectiveness and noise immunity of this method are demonstrated by the identified results. In addition, the effects of initial sampling size on the identification efficiency and the precision of the identified results are also investigated.