The autonomous navigation system based on starlight atmosphere refraction has simple structure and low cost. Using the system, the spacecraft can get high-precision navigation data. But the measurement model has some imprecise atmospheric parameter, which reduces the navigation precision and reliability seriously. The above effects can be avoided by nervous network fitting starlight atmosphere refraction model on line, using satellite navigation data, when the spacecraft is at low orbit. This method can improve the navigation precision based on starlight refraction. Through simulation, the characteristics of the starlight refraction model and the fitting method were researched, and the precision analysis of the fitting models were also made.