Limitations of GPT2

Processing of GNSS data collected on an airplane requires special handling of the troposphere. Due to large variations in altitude, the a priori hydrostatic and wet delays must be regularly updated since the tropospheric zenith delay (TZD) process noise would not allow proper tracking of those variations. However, the temperature and pressure reduction schemes employed in GPT2 do not seem well suited for this purpose, as demonstrated herein using data from the NGS kinematic challenge.

For temperature reduction, the GPT2 model [Lagler et al., 2013] replaced the constant lapse rate of -6.5⁰C/km by lapse rates dependent on location. Figure 1 below shows the interpolated temperature lapse rate as a function of the airplane latitude, for altitudes between 10 and 11.5 km. The rapid airplane displacements clearly emphasize that nearby grid points were assigned very different lapse rate values. While this is not an issue at sea level for a static receiver, when the airplane is flying at an altitude of 11 km, the differences in temperatures become very significant and impact wet zenith delay computations.

Figure 1: Temperature lapse rate provided by GPT2 as a function of latitude, for altitudes between 10 km and 11.5 km.


As opposed to the GPT model, GPT2 also provides values for the predicted water vapor pressure, although this measure only seems valid at the altitude of the grid points. For high altitudes, the water vapor pressure given exceeds the saturation vapor pressure, leading to a relative humidity greater than 100%. The pressure reduction process is based on a virtual temperature, also dependent on the water vapor pressure, and provides values significantly different from GPT at high altitudes.

The combined effect of incorrect temperature, pressure and water vapor on the total zenith delay is depicted in Figure 2, showing differences with the GPT model. Large differences reaching more than 20 cm exist between models, and the consequences on the PPP solution are obviously disastrous…

Figure 2: Differences in total zenith delay between GPT2 and GPT.

Perhaps that the GPT2 model may never have been intended to be used in such applications, and perhaps that if I had understood the model before blindly using it, I wouldn’t have been surprised by the results. My suggestion is still to put a note in the GPT2 source code on those shortcomings, which would save (de)buggers like me a few days of work.


Lagler, K., M. Schindelegger, J. Böhm, H. Krásná, and T. Nilsson (2013), GPT2: Empirical slant delay model for radio space geodetic techniques, Geophys. Res. Lett., 40, 1069–1073, doi:10.1002/grl.50288.

Write a comment

Comments: 0