Abstract: The precise observation model for stellar refraction navigation using the least squares support vector machine（LS-SVM）is introduced.The process involves high fidelity simulations of stellar refraction via ray tracing，considering changes of atmospheric parameters with altitude，to determine the nonlinear mapping between true altitude and refraction angle.The LS-SVM algorithm is applied to fit this relationship accurately，creating a precise model for refracted apparent height.Simulations show the model achieves high precision，with an average absolute error of 0.986m.When integrated into a stellar navigation system，it significantly enhances navigation accuracy，reducing the average positioning error to 130.7m and velocity error to 0.1479m/s，thereby validating the effectiveness of the modeling approach.This approach is significant for the high-precision study of autonomous navigation using stellar refraction.

Key words: stellar refraction navigation, observation model, ray tracing, LS-SVM, machine learning