The information of height provided by the GNSS (Global Navigation Satellite System) is purely geometrical, and in most engineering papers, the height must be referenced to the geoid. Provided we have a sufficient number of Bench Marks (BMs) with known horizontal and vertical coordinates, it is nearly always possible to adjust mathematical expressions that allow for the interpolation of geoidal heights. The aim of this paper is to evaluate the efficiency of Artificial Neural Network (ANN) in the process of predicting geoidal heights, having the State of São Paulo as the area of study. The information used is based on a set of 157 BMs, evenly distributed all across the State. The horizontal coordinates (latitude and longitude) and the vertical coordinates (geometrical, orthometrical and geoidal heights) of these BMs are known. From the 157 BMs, 115 were used for the training of RNA and 42 in the process of simulation to assess the efficiency of the model proposed. Efficiency is based in determining the discrepancies (error) between known geoidal heights and those which were obtained by the neural model. As a contribution to this research, we have compared the values simulated with the Earth Gravitational Model 2008 (EGM2008) and with the MAPGEO2004 as well. In terms of results, the RNA produced a mean absolute error of 0.19 m ± 0.14 m and a strong correlation (R² = 0.9871) with the values taken as true. Statistically, the tests showed that there was no difference between known geoidal heights and those which were provided by the neural model for a level of significance of 5%. When we compare these results with the EGM2008 and MAPGEO2004, the RNA has an error reduction of 0.07 and 0.44 m, respectively.
GPS; MAPGEO2004; EGM2008; artificial neural networks; geoidal height
