Slant tropospheric models based on machine learning


METADATA ONLY
Loading...

Date

2025

Publication Type

Other Conference Item

ETH Bibliography

yes

Citations

Altmetric
METADATA ONLY

Data

Rights / License

Abstract

Tropospheric errors pose a major challenge for high-precision applications of space geodetic techniques such as GNSS, InSAR, and VLBI. To estimate tropospheric delays and support real-time geodetic applications, empirical tropospheric models are needed. For example, GPT3, a state-of-the-art model developed by TU Wien, provides meteorological parameters, mapping functions and tropospheric gradients at any time and location and has been widely used by the community. However, traditional models rely on simplistic representations of tropospheric parameters, such as mean values combined with annual and semi-annual amplitudes. They also assume symmetry and use gradient compensation to account for asymmetries, which may limit their accuracy. Given the powerful modeling capabilities of machine learning, we explore its potential to enhance empirical models by directly predicting mapping functions in any direction. In this study, we processed over five years of ray-tracing results from ERA5 to create a training dataset and developed machine-learning-based models for mapping functions. We demonstrate that these models can predict mapping factors in a single step for arbitrary elevation and azimuth angles. Our new model is significantly faster in computation and more compact in size than the traditional GPT3 model, while maintaining comparable accuracy regarding global performance.

Publication status

published

External links

Editor

Book title

IAG Scientific Assembly 2025: Abstract Book

Journal / series

Volume

Pages / Article No.

100 - 100

Publisher

International Association of Geodesy

Event

IAG Scientific Assembly 2025

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Machine learning; Mapping function; Space geodesy; Ray-tracing; ERA5

Organisational unit

09707 - Soja, Benedikt / Soja, Benedikt

Notes

Conference lecture on September 5, 2025

Funding

Related publications and datasets