IMPROVING THE ACCURACY OF THE LOCAL GEOID MODEL OF THE ASTANA CITY USING MACHINE LEARNING METHODS

Abstract

The development of a high-precision regional geoid model is an important step in the modernization of Kazakhstan’s vertical reference framework and in improving the efficiency of height determination from GNSS data. This paper proposes a hybrid approach to refining the local geoid model for the Astana area by combining the LSMSA gravimetric base solution with statistical modeling of residual corrections. The base geoid surface is computed using the Least Squares Modification of Stokes’ Formula with additive corrections and is validated against GNSS/levelling control points. To reduce discrepancies with observations, a correction surface is formed from residuals using a 7-parameter Helmert transformation and machine learning regression models (GPR, SVR, LSBoost). On the independent test subset, SVR reduced RMSE from 0.062 to 0.047 m, corresponding to a 24% decrease in error. The results show that the proposed scheme reduces the spread of residual errors and improves the agreement of the geoid model with GNSS/levelling data.