Abstract: Taking the observation data of 27 GPS reference stations in Yunnan State Land Network at different time spans from 2011 to 2017 as examples, the Achievement and Bayesian Information Estimation Criterion (AIC／BIC) criterion is used to analyze the solution results. And so as to determine its optimal colored noise model. The results show that the optimal noise model for each coordinate component of the reference station at different time spans is mainly combination of white+flicker noise (WN+FN) and Power-law (PL) noise models. The optimal noise models of some reference stations in the north and vertical directions will change with the accumulation of observation time, and the proportion of PL models will decrease. While the proportion of WN+FN model will increase. When the time span is greater than 5a, the stability of the reference station noise model is significantly improved. As the accumulation time of the coordinate time series increases, the probability of the occurrence of unknown noise components increases accordingly.