EGU25-19045, updated on 15 Mar 2025
https://doi.org/10.5194/egusphere-egu25-19045
EGU General Assembly 2025
© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.
AI for GNSS Reflectometry: Setting a New Benchmark for Earth Surface and Atmospheric Monitoring
Milad Asgarimehr1, Daixin Zhao1,2, Tianqi Xiao1,3, Hamed Izadgoshasb1,4, and Jens Wickert1,3
Milad Asgarimehr et al.
  • 1Section 1.1 Space Geodetic Techniques, German Research Centre for Geosciences GFZ, Potsdam, Germany
  • 2Data Science in Earth Observation, Technical University of Munich, Munich, Germany
  • 3Institute of Geodesy and Geoinformation Science, Technische Universitaet Berlin, Berlin, Germany
  • 4Department of Information Engineering, Electronics, and Telecommunications Sapienza University of Rome, Italy

GNSS Reflectometry (GNSS-R) satellite constellations offer unprecedentedly large datasets. This creates a unique opportunity to harness the power of AI for Earth system monitoring using GNSS-R. By using these vast datasets, AI models can “learn” effectively and adaptively. The AI for GNSS-R (AI4GNSS-R) project unlocks the potential of deep learning to produce high-quality and innovative data products.

Previously, we introduced CyGNSSnet, a deep learning model based on a CNN-LSTM architecture, which achieved an RMSE of 1.36 m/s—representing a substantial 28% improvement over the officially operational retrieval algorithm. Building on this success, we now present an advanced version of the model that incorporates external precipitation data through data fusion. This enhanced approach achieves an RMSE of 1.57 m/s for rain-affected data, significantly improving wind speed predictions under extreme weather conditions. For land monitoring, we demonstrate the retrieval of Vegetation Water Content (VWC) from GNSS-R data. Using architectures such as LeNet, our models achieve RMSEs below 0.6 kg/m² compared to SMAP VWC data, validating GNSS-R's capability for  global vegetation moisture monitoring. A highlight of our research is the development of a GNSS-R general foundation model using self-supervised learning. This model facilitates the fusion of multimodal data and enables scalable and accurate retrieval of variety of parameters such as wind speed, soil moisture, and VWC with limited labeled data. Our findings emphasize the scalability, adaptability, and applicability of next-generation AI models for GNSS-R. These advancements establish a new benchmark for the achievable quality and extends application spectrum of spaceborne GNSS-R.

How to cite: Asgarimehr, M., Zhao, D., Xiao, T., Izadgoshasb, H., and Wickert, J.: AI for GNSS Reflectometry: Setting a New Benchmark for Earth Surface and Atmospheric Monitoring, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19045, https://doi.org/10.5194/egusphere-egu25-19045, 2025.