The traditional fixed optical axis observation approach reduces positioning accuracy and reliability in the starlight refraction indirect horizon-sensing positioning method by producing a small number of refracted stars with an uneven distribution. Therefore, a high-precision starlight refraction celestial navigation observation method is proposed. Based on the geometric relationship between the stratospheric refraction band and the star sensor’s field of view (FOV), a relation model is established between the FOV coverage area of the refraction band and the star sensor’s elevation angle. By maximizing the FOV coverage area and combining the flight parameters, the optimal elevation angle is determined to increase the observable refracted stars. Additionally, a quantitative relationship model is developed between positioning precision and the spatial arrangement of refracted stars. Based on the minimum position dilution of precision (PDOP) criterion, the optimal spatial geometric configuration of refracted stars is derived, thereby obtaining the optimal azimuth angle of the optical axis. This enables continuous and uniform observation for refracted starlight, so as to improve the positioning accuracy of celestial navigation. Simulation results demonstrate that compared with the conventional method, the proposed method achieves 75.7% and 66.7% improvements in position and velocity estimation accuracy, respectively.